Search

One of the pandemics that have caused many deaths is the Coronavirus disease 2019 (COVID-19). It first appeared in late 2019, and many deaths are increasing day by day until now. Therefore, the early diagnosis of COVID-19 has become a salient issue. Additionally, the current diagnosis methods have several demerits, and a new investigation is required to enhance the diagnosis performance. In this paper, a set of phases are performed, such as collecting data, filtering and augmenting images, extracting features, and classifying ECG images. The data were obtained from two publicly available ECG image datasets, and one of them contained COVID ECG reports. A set of preprocessing methods are applied to the ECG images, and data augmentation is performed to balance the ECG images based on the classes. A deep learning approach based on a convolutional neural network (CNN) is performed for feature extraction. Four different pre-trained models are applied, such as Vgg16, Vgg19, ResNet-101, and Xception. Moreover, an ensemble of Xception and the temporary convolutional network (TCN), which is named ECGConvnet, is proposed. Finally, the results obtained from the former models are fed to four main classifiers. These classifiers are softmax, random forest (RF), multilayer perception (MLP), and support vector machine (SVM). The former classifiers are used to evaluate the diagnosis ability of the proposed methods. The classification scenario is based on fivefold cross-validation. Seven experiments are presented to evaluate the performance of the ECGConvnet. Three of them are multi-class, and the remaining are binary class diagnosing. Six out of seven experiments diagnose COVID-19 patients. The aforementioned experimental results indicated that ECGConvnet has the highest performance over other pre-trained models, and the SVM classifier showed higher accuracy in comparison with the other classifiers. The resulting accuracies from ECGConvnet based on SVM are (99.74%, 98.6%, 99.1% on the multi-class diagnosis tasks) and (99.8% on one of the binary-class diagnoses, while the remaining achieved 100%). It is possible to develop an automatic diagnosis system for COVID based on deep learning using ECG data. [ABSTRACT FROM AUTHOR]

This special issue of the circuits, systems and signal processing journal focuses on recent advances and improvements in leading-edge integrated machine learning algorithms and signal processing systems. Although signal processing has been studied over several decades, the computer industry is only beginning to understand how signal processing techniques can be well integrated for the development of human-machine interfaces with the advancement of machine learning algorithms. As advances in signal processing tools and machine learning algorithms are becoming more powerful in terms of functionality and communicative capabilities, their contribution to the journal of circuits, system and signal processing is becoming more significant. [Extracted from the article]

Highlights from the article: Deep learning employs deep neural networks (DNNs), which are neural networks with more than one hidden layer, with recently developed initialization and training strategies using massive amounts of diverse data as examples. The paper provides an overview of recent approaches to deep learning as applied to a range of speech processing tasks, primarily for automatic speech recognition (ASR), but also text-to-speech and speaker, language, and emotion recognition. The authors propose integrating deep neural network (DNN)-HMM technique with the HiLAM method where the state alignment information from the HiLAM is used to discriminatively train a DNN to further improve the system performance.

Multimedia data have increased dramatically today, making the distinction between desirable information and other types of information extremely important. Speech/music discrimination is a field of audio analytics that aims to detect and classify speech and music segments in an audio file. This paper proposes a novel feature extraction method called Long-Term Multi-band Frequency-Domain Mean-Crossing Rate (FDMCR). The proposed feature computes the average frequency-domain mean-crossing rate along the frequency axis for each of the perceptual Mel-scaled frequency bands of the signal power spectrum. In this paper, the class-separation capability of this feature is first measured by well-known divergence criteria such as Maximum Fisher Discriminant Ratio (MFDR), Bhattacharyya divergence, and Jeffreys/Symmetric Kullback–Leibler (SKL) divergence. The proposed feature is then applied to the speech/music discrimination (SMD) process on two well-known speech-music datasets—GTZAN and S &S (Scheirer and Slaney). The results obtained on the two datasets using conventional classifiers, including k-NN, GMM, and SVM, as well as deep learning-based classification methods, including CNN, LSTM, and BiLSTM, show that the proposed feature outperforms other features in speech/music discrimination. [ABSTRACT FROM AUTHOR]

Environment Sound Classification (ESC) has been a challenging task in the audio field due to the different types of ambient sounds involved. In this paper, we propose a method for the ESC tasks based on the CAR-Transformer neural network model, which includes stages of sound sample pre-processing, deep learning-based feature extraction, and classifier classification. We convert the one-dimensional audio signal into two-dimensional Mel Frequency Cepstral Coefficients (MFCC) and use them as the feature map of the audio. The CAR-Transformer model was used for feature extraction, and after dimensionality reduction of the extracted feature map, we use the fully connected layer as a classifier of the feature map to obtain the final results. The method achieves a classification accuracy of 96.91% on the UrbanSound8K dataset, while the number of parameters in the model is only 0.16 M. The results of this paper were compared with other state-of-art research. [ABSTRACT FROM AUTHOR]

With the rapid development of information transfer technology and the influence of complex electromagnetic environment, the signal components in communication systems are becoming more and more complex, and the spectrum congestion is further aggravated, which seriously affects the performance of communication systems. Single-channel blind source separation technology is an effective solution, but the existing algorithms are deficient in terms of estimation accuracy, time delay and computational complexity. Based on the powerful feature extraction capability and timing signal processing capability of deep learning, this paper proposes an end-to-end time-domain blind separation convolutional network structure, including three parts: encoder, separation module and decoder. The attention mechanism module is also introduced into the encoding layer of the network to enhance the deep feature extraction ability of the network, highlight the influence of important timing features of communication signals, further improve the separation performance and enhance the practicality of the network. The experimental results show that the algorithm has superior performance and can effectively separate the mixed communication signal of various modulation types, and it can also extract own-side communication signals from high-power masked signals. Compared with the existing algorithms, the network in this paper has high separation accuracy and low computational complexity, which breaks the bottleneck of traditional single-channel blind separation algorithms in complex communication electromagnetic environment. [ABSTRACT FROM AUTHOR]

The cry of a premature infant is an attempt to connect with its mother or others. The newborns are communicated in different ways depending on the reason for their screams. In recent days, the preprocessing, feature extraction, and classification of audio signals require expert attention and a lot of effort. In this paper, a novel deep convolutional recurrent neural network (DCR net) has been proposed to classify the premature infant cry signal into different categories. The acquisition of the cry signal generally requires a lengthy observation period and several activity processes to obtain all the signals of the premature infant. The relevant multi-class frequency features are extracted by using the MFCC (Mel-frequency cepstral coefficient), BFCC (Bark-frequency cepstral coefficient), and LPCC (linear prediction cepstral coefficient) features, which are combined to create a fused feature matrix that is helpful in the classification of pathological crying. Based on these features, the DCR net is used to classify sounds in the premature infant cry. The sound of the target cry signal is classified into five categories: "neh" means hunger, "heh" means discomfort, "eh" means burping, "eair" means cramps, and "owh" means fatigue. The efficiency of the DCR net was estimated with some metrics such as specificity, precision, accuracy, recall, and F1 score. The experimental fallouts disclose that the proposed DCR net attains a better classification accuracy of 97.27% for identifying infant cry signals. The DCR net increases the overall performance range by 8.61%, 11.58%, 0.54%, and 17.03% better than SVM-RBF, MFCC-SVM, optimized deep learning model, and hidden Markov model, respectively. [ABSTRACT FROM AUTHOR]

Speech enhancement is an important method for improving speech quality and intelligibility in noisy environments. An effective speech enhancement model depends on precise modelling of the long-range dependencies of noisy speech. Several recent studies have examined ways to enhance speech by capturing the long-term contextual information. For speech enhancement, the time–frequency (T–F) distribution of speech spectral components is also important, but is usually ignored in these studies. The multi-stage learning method is an effective way to integrate various deep learning modules at the same time. The benefit of multi-stage training is that the optimization target can be iteratively updated stage by stage. In this paper, speech enhancement is investigated by multi-stage learning using a multi-stage structure in which time–frequency attention (TFA) blocks are followed by stacks of squeezed temporal convolutional networks (S-TCN) with exponentially increasing dilation rates. To reinject original information into later stages, a feature fusion block (FB) is inserted at the input of later stages to reduce the possibility of speech information being lost in the early stages. The S-TCN blocks are responsible for temporal sequence modelling tasks. The time–frequency attention (TFA) is a simple but effective network module that explicitly exploits position information to generate a 2D attention map to characterize the salient T–F distribution of speech by using two branches, time-frame attention and frequency attention in parallel. Extensive experiments have demonstrated that the proposed model consistently improves the performance over existing baselines across two widely used objective metrics such as PESQ and STOI. A significant improvement in system robustness to noise is also shown by our evaluation results using the TFA module. [ABSTRACT FROM AUTHOR]

This paper considers the deep-learning-aided compressed video sensing problem. To this end, a deep generative adversarial network has been proposed to provide an approximation of the non-reference frame using its corresponding reference frame. The tests confirm the superiority of this scheme over the conventional methods used earlier. Furthermore, two scenarios have been suggested for deep compressed video sensing and recovery. In the first scenario, the difference between the non-reference frame and its approximation obtained from the pre-trained network is compressively sampled and transmitted to the receiver where the proposed residual reconstruction network is adopted to reconstruct the signal. The second scenario utilizes a pre-trained network followed by an augmented layer to approximate the non-reference frames. In the transmitter, the parameters of the augmented layer are trained for the current non-reference block. Instead of transmitting the samples of the block, the parameters of its trained augmented layer are sent to the receiver where the reconstruction is done using the same pre-trained network. The performances of the proposed scenarios demonstrate their objective and subjective superiority over the state-of-the-art algorithms in both the reconstruction quality and run time. [ABSTRACT FROM AUTHOR]

Convolutional neural network (CNN) is one of the most dominant deep learning networks with good generalization ability. Its high performance in solving large and complex learning problems has enabled usability in IoT devices. However, CNN involves a substantial amount of convolution operations, which demand a large number of power-consuming multipliers. This hinders the deployment of deep CNNs on mobile and IoT edge devices owing to restricted power–area constraints. In this paper, we propose a low-complex methodology named 'minimal modified distributed arithmetic' (M2DA) for convolutional neural network (CNN) by exploiting the data symmetry and consequently storing only the unique kernel coefficient's combinations and the size of required memory and multiplication operations can be reduced, leading to power–area efficient design. For validation, a low-complex CNN architecture for activity recognition application is designed and synthesized in Synopsys using the UMC 65 nm technology wherein average 36.89% and 51.63% improvement is achieved in power and area, respectively, compared to conventional MDA methodology. To demonstrate the significance of the proposed M2DA methodology, we have also implemented the Alexnet which is the most widely and publicly available CNN model for the image classification problem. [ABSTRACT FROM AUTHOR]

The digital revolution places great emphasis on digital media watermarking due to the increased vulnerability of multimedia content to unauthorized alterations. Recently, in the digital boom in the technology of hiding data, research has been tending to perform watermarking with numerous architectures of deep learning, which has explored a variety of problems since its inception. Several watermarking approaches based on deep learning have been proposed, and they have proven their efficiency compared to traditional methods. This paper summarizes recent developments in conventional and deep learning image and video watermarking techniques. It shows that although there are many conventional techniques focused on video watermarking, there are yet to be any deep learning models focusing on this area; however, for image watermarking, different deep learning-based techniques where efficiency in invisibility and robustness depends on the used network architecture are observed. This study has been concluded by discussing possible research directions in deep learning-based video watermarking. [ABSTRACT FROM AUTHOR]

With the development of deep learning techniques, the field of direction of arrival (DOA) estimation has also made significant progress. However, the accuracy of DOA estimation using end-to-end neural networks (NNs) heavily relies on the classification step of the networks, which necessitates the use of large and representative datasets. Additionally, conventional speech presence probability (SPP) estimation methods based on the ideal ratio mask (IRM) may misclassify time-frequency (T-F) bins dominated by non-speech and noise, which hinders the accurate extraction of directional information. To improve the robustness of existing DOA estimation algorithms, this paper proposes a DOA estimation method with T-F bin selection. In terms of output, instead of using IRM-based SPP, our proposed approach focuses on the a posteriori SPP, a deliberate choice aimed at circumventing potential confusion. For input optimization, we construct features that encompass spatial, temporal, and directional information concurrently, and these are coupled with a frequency bin-wise recurrent neural network (RNN) model to attain precise multi-channel SPP estimation. Subsequently, these SPP estimates are utilized to extract local information for DOA estimation. Moreover, the cascaded structure ensures that the model has the ability to complete out-of-label tasks, effectively reducing the dataset requirements by training only a subset of direction information to achieve omnidirectional DOA estimation. Besides, this contributes to the algorithm's ability to eliminate its reliance on the step size, setting it apart from other end-to-end methods. Simulation results validate that the proposed method achieves higher accuracy and lower error compared to both NN-based end-to-end approaches and traditional full-band approaches under various conditions of reverberation and signal-to-noise ratio. [ABSTRACT FROM AUTHOR]

The active use of the Internet and multimedia content has recently escalated copyright violations. Digital content, especially images and videos are subject to vulnerable attacks. It is also possible that an attacker might remove the watermark from the original image. Therefore, the copyright of digital images must be secured to prevent them from being inappropriately misused. This paper proposes an Enhanced Chimp Optimization algorithm based on Deep Fusion Convolutional Neural Network (ECO-DFCNN) for robust watermarking. The proposed framework consists of an embedding and extraction network to embed and extract the watermark. The octave convolutional model introduced in the embedding network captures various features and decreases spatial redundancy. In addition, the ECO algorithm is introduced to overcome the trade-off between robustness and imperceptibility by determining the optimal strength factor. The pyramid feature extraction module in the extraction network extracts the local features and the dilated convolutions minimize the model parameters. The proposed ECO-DFCNN method is tested against various attacks such as histogram equalization, compression, cropping, scaling, blurring, and median filtering. The proposed ECO-DFCNN method is evaluated and the performance is determined by comparing the obtained results with the existing watermarking techniques. The results show that the proposed ECO-DFCNN watermarking method is robust against various attacks while maintaining excellent imperceptibility with a high Peak Signal-to-Noise Ratio of 54.64 dB, Normalized Correlation of 0.98 and Structural Similarity Index Measure of 0.97, and low Bit Error Rate of 0.038. [ABSTRACT FROM AUTHOR]

Super-resolution (SR) technology based on deep learning has achieved excellent results. However, too many convolution layers and parameters consume a very high computational cost and storage space when training the model, which dramatically limits the practical application. To solve this problem, this paper proposes a lightweight feature asymmetric distillation SR network (FADLSR). FADLSR constructs the feature extractor module through the stacked feature asymmetric distillation block (FADB). It extracts the low-resolution image features hierarchically and integrates them to obtain more representative features to improve the SR quality. In addition, we design a new focus block and add it to FADB to improve the quality of feature acquisition. We also introduce asymmetric convolution to strengthen the key features of the skeleton region. Detailed experiments show that our FADLSR has achieved excellent results in objective evaluation criteria and subjective visual effect on the test sets of Set5, Set14, B100, Urban100, and Manga109. The parameters of our model are roughly the same as those of the current state-of-the-art models. Moreover, in terms of model performance, FADLSR is 10–15% higher than the comparison algorithms. [ABSTRACT FROM AUTHOR]

Deep learning (DL)-based compressed sensing (CS) has been applied for better performance of image reconstruction than traditional CS methods. However, most existing DL methods utilize the block-by-block measurement and each measurement block is restored separately, which introduces harmful blocking effects for reconstruction. Furthermore, the neuronal receptive fields of those methods are designed to be the same size in each layer, which can only collect single-scale spatial information and has a negative impact on the reconstruction process. This paper proposes a novel framework named multi-scale dilated convolution neural network for CS measurement and reconstruction. During the measurement period, we directly obtain all measurements from a trained measurement network, which employs fully convolutional structures and is jointly trained with the reconstruction network from the input image. It need not be cut into blocks, which effectively avoids the block effect. During the reconstruction period, we propose the multi-scale feature extraction (MFE) architecture to imitate the human visual system to capture multi-scale features from the same feature map, which enhances the image feature extraction ability of the framework and improves the performance of image reconstruction. In the MFE, there are multiple parallel convolution channels to obtain multi-scale feature information. Then, the multi-scale features information is fused and the original image is reconstructed with high quality. Our experimental results show that the proposed method performs favorably against the state-of-the-art methods in terms of PSNR and SSIM. [ABSTRACT FROM AUTHOR]

Over the last few years, deep learning on irregular 3D data given its wide range of applications has become one of the active topics in the field. While field programmable gate array (FPGA)-based acceleration of deep learning models has been proved to produce power-efficient designs in comparison with other platforms such as CPUs and GPUs, only a few studies have been conducted regarding the models that consume point clouds as their input. Although tailoring the hardware designs to specific networks could lead to better optimization opportunities, it is also important to keep the reusability of the design in mind, especially for a new and evolving topic like learning on point clouds. In this work, we have aimed to achieve reusability by keeping the hardware isolated from the computational graph. Considering the numerous types of layers used in dynamic graph convolutional neural network (DGCNN) and its popularity, our proposed design aims for the thorough acceleration of DGCNN. The challenges including 18 types of tensor operations, achieving burst transfers, dealing with kernel complexities, external memory banks, in-order and out-of-order execution modes, and approaches with multiple processing elements have been explained in details throughout the paper. Our experiments on a single FPGA with a single bitstream, DDR4 memory subsystem, and Float32 data type demonstrated speedups of 2.73 × to 8.4 × compared to a sequential single-threaded implantation on an Intel Core i7 6700HQ. [ABSTRACT FROM AUTHOR]

Speech emotion recognition (SER) has recently increased because of vast innovations in human–computer interaction and affective computing. In recent years, numerous deep learning-based schemes presented for SER have shown significant improvement over the traditional machine learning approaches. Most deep learning-based faced SER systems face challenges due to data imbalance problem that occurs due to unequal samples in the database. The input to two-dimensional CNN uses traditional MFCC for SER. It degrades the quality of deep attributes because of the higher variance, frequency resolution problem and spectral leakage problem of traditional MFCC. This paper proposed the novel Multi-taper Mel Frequency Logarithmic Spectrogram to enrich the Deep Convolutional Neural Network effectiveness for SER. Further, Generative Adversarial Network is used for speech emotion data augmentation during training to deal with data scarcity problems in SER. The performance of the proposed SER scheme is validated using the Berlin EmoDB and RAVDESS datasets. The proposed method provides SER accuracy of 96.65% and 97.12% for the EmoDB and RAVDESS dataset, respectively, and significantly improves over the recent techniques. [ABSTRACT FROM AUTHOR]

Several works report on overcoming the challenge of complex general-purpose signal processing algorithms that result in quick utilization of chip area resources (like FPGA in implementing system on chip or network on chip). However, with the advent of machine learning and deep learning algorithms, reducing this complexity and immense use of resources is a challenging task that can improve the system's performance in real time. One of the approaches is algorithm optimization, which consumes fewer resources and produces an equivalent performance. This approach may likely be accepted in less-sensitive applications, like voice or video processing. Recently, an optimized algorithm of correlation-less Wiener–Hopf-based adaptive channel equalizer is reported, in which the overall complexity of the system is reduced by employing a more compact way of weight optimization. This paper is the continuation of that work that reports the modification in the Steepest-Descent-based adaptive channel equalizer algorithm by keeping the autocorrelation matrix out of the algorithm. The proposed design is simulated and tested at equivalent spectral performance and produces comparable performance with fewer chip resources when translated on FPGA. The proposed design supports that algorithm-level optimization may be accepted extensively for the optimal hardware-based design of DSP systems. [ABSTRACT FROM AUTHOR]

Multi-channel speech enhancement techniques are mainly based on optimal multi-channel speech estimators that comprise a minimum variance distortionless response (MVDR) beamformer followed by a single-channel Wiener post-filter. There are two problems in the application of this theoretically optimal solution. The first is the high sensitivity of the MVDR beamformer to errors in the estimated acoustic transfer function (ATF). The second is the accuracy of the time-varying post-filter coefficients estimated from non-stationary speech and noise. Mask-based beamforming developed in the last decade considerably improves the performance of the MVDR beamformer. In addition, the estimated time–frequency mask can be successfully used in post-filter design. In this paper, we propose several improvements to this approach. First, we propose an end-fire microphone array with a better directivity index than the corresponding broadside array. The proposed microphone array is composed of unidirectional microphone capsules that increase the directivity of the microphone array. Second, we propose preprocessing using a delay-and-sum beamformer before estimating the ideal ratio mask (IRM). Next, we propose a simplified generalized sidelobe canceller (S-GSC), which does not need to estimate ATF. We also improved the design of its blocking matrix by scaling the null space eigenvectors of the speech covariance matrix. The proposed computationally efficient multiple iteration method improves the adaptation of the S-GSC parameters. Finally, we improved the previous IRM-based post-filter, considering the SNR improvement at the output of the S-GSC beamformer. The integral speech enhancement procedure was tested on real room recordings using PESK, STOI, and SDR measures. [ABSTRACT FROM AUTHOR]

The interpretation of human voices holds importance across various applications. This study ventures into predicting age, gender, and emotion from vocal cues, a field with vast applications. Voice analysis tech advancements span domains, from improving customer interactions to enhancing healthcare and retail experiences. Discerning emotions aids mental health, while age and gender detection are vital in various contexts. Exploring deep learning models for these predictions involves comparing single, multi-output, and sequential models highlighted in this paper. Sourcing suitable data posed challenges, resulting in the amalgamation of the CREMA-D and EMO-DB datasets. Prior work showed promise in individual predictions, but limited research considered all three variables simultaneously. This paper identifies flaws in an individual model approach and advocates for the novel multi-output learning architecture, the Speech-based Emotion, Gender and Age Analysis (SEGAA) model. The experiments suggest that Multi-output models perform comparably to individual models, efficiently capturing the intricate relationships between variables and speech inputs, all while achieving improved runtime. [ABSTRACT FROM AUTHOR]

An image caption generator produces syntactically and semantically correct sentences to narrate the scene of a natural image. A neural image caption (NIC) generator is a popular deep learning model for automatically generating image captions in plain English. The NIC generator combines a convolutional neural network (CNN) encoder and a long short-term memory (LSTM) decoder. This paper investigates the performance of different CNN encoders and recurrent neural network decoders for finding the best NIC generator model for image captioning. Besides, we test the image caption generators with four image inject models and with decoding strategies such as greedy search and beam search. We conducted experiments on the Flickr8k dataset and analyzed the results qualitatively and quantitatively. Our results validate the automated image caption generator with ResNet-101 encoder, and the LSTM/gated recurrent units decoder outperforms the popular neural image caption NIC generator in the presence of par-inject concatenate conditioning and beam search. For quantitative assessment, we used R O U G E L , C I D E r D , and B L E U n scores to compare the different models. [ABSTRACT FROM AUTHOR]

Deep neural network (DNN), being an important member of machine learning family, has been employed to serve a wide range of applications in the area of signal and image processing like pattern recognition, speech recognition, language processing, image segmentation, etc. To this aim, this paper concentrates on the design of a narrow transition-band finite impulse response (FIR) filter with the aid of back-propagation-based deep learning approach. The proposed deep learning-based approach offers a unified design framework for a variety of FIR filters. Convergence behaviour of the proposed algorithm has been proved analytically in situations when weights between adjacent layers are updated continuously. Simulation results have shown the frequency response characteristics of several FIR filters with narrow transition-band, designed with the help of proposed approach. Advantage of our design strategy has also been established in terms of magnitude response over a number of state-of-the-art techniques of recent interest. Simulation results have shown noticeable improvement in terms of transition bandwidth when compared with few existing works. Designed filter is subsequently implemented on Altera's Cyclone IV field programmable gate array (FPGA) chip, and hardware efficiency of the suggested design has strongly been established by correlating its hardware cost with many of the state-of-the-art FIR filters. [ABSTRACT FROM AUTHOR]

In recent years, with the rapid development of artificial intelligence, human activity recognition has become a research focus. The complex, dynamic and variable features of human activities lead to the relatively low accuracy of the traditional recognition algorithms. In order to solve the problem, this paper will propose a novel structure named hierarchical deep LSTM (H-LSTM) based on long short-term memory. Firstly, the original sensor data are preprocessed by smoothing and denoising; then, the feature will be selected and extracted by time–frequency-domain method. Secondly, H-LSTM is applied to the classification of these activities. Three public UCI datasets are used to conduct simulation with the realization of the automatic extraction of feature vectors and classification of outputting recognition results. Finally, the simulation results testify to the outperformance of the H-LSTM network over other deep learning algorithms. The accuracy of H-LSTM network in human activity recognition is proved to be 99.15%. [ABSTRACT FROM AUTHOR]

Identifying a scene based on the environment in which the related audio is recorded is known as acoustic scene classification (ASC). In this paper, a bi-level light-weight Convolutional Neural Network (CNN)-based model is presented to perform ASC. The proposed approach performs classification in two levels. The scenes are classified into three broad categories in the first level as indoor, outdoor, and transportation scenes. The three classes are further categorized into individual scenes in the second level. The proposed approach is implemented using three features: log Mel band energies, harmonic spectrograms and percussive spectrograms. To perform the classification, three CNN classifiers, namely, MobileNetV2, Squeeze-and-Excitation Net (SENet), and a combination of these two architectures, known as SE-MobileNet are used. The proposed combined model encashes the advantages of both MobileNetV2 and SENet architectures. Extensive experiments are conducted on DCASE 2020 (IEEE AASP Challenge on Detection and Classification of Acoustic Scenes and Events) Task 1B development and DCASE 2016 ASC datasets. The proposed SE-MobileNet model resulted in a classification accuracy of 96.9% and 86.6% for the first and second levels, respectively, on DCASE 2020 dataset, and 97.6% and 88.4%, respectively, on DCASE 2016 dataset. The proposed model is reported to be better in terms of both complexity and accuracy as compared to the state-of-the-art low-complexity ASC systems. [ABSTRACT FROM AUTHOR]

Spoofing attacks posed by generating artificial speech can severely degrade the performance of a speaker verification system. Recently, many anti-spoofing countermeasures have been proposed for detecting varying types of attacks from synthetic speech to replay presentations. While there are numerous effective defenses reported on standalone anti-spoofing solutions, the integration for speaker verification and spoofing detection systems has obvious benefits. In this paper, we propose a spoofing-robust automatic speaker verification system for diverse attacks based on a multi-task learning architecture. This deep learning-based model is jointly trained with time-frequency representations from utterances to provide recognition decisions for both tasks simultaneously. Compared with other state-of-the-art systems on the ASVspoof 2017 and 2019 corpora, a substantial improvement of the combined system under different spoofing conditions can be obtained. [ABSTRACT FROM AUTHOR]

People with sensory difficulties like dumbness, or with a disease like laryngeal cancer are the major causes of loss of production of human voice signal. This sensory difficulty leads to the use of sign language for their communication with a normal person. A normal person requires a special skill to decode sign language. This paper presents an efficient methodology to recognize the uttered word with the facial expression of speaker using deep learning framework. The proposed model generates an artificial acoustic signal along with an appropriate expression for the words uttered by mute people. It employs two deep learning architectures to recognize the uttered word and facial expressions, respectively. The recognized facial expression is then appended to the artificial acoustic signal. In this work, the uttered word is recognized by a combination of HOG + SVM classifier and a fine-tuned VGG-16 ConvNet with LSTM network using transfer learning. Furthermore, the facial expression of speaker is recognized using a combination of Haar-Cascade classifier and a fine-tuned MobileNet with LSTM network. A detailed evaluation on the proposed model shows that the accuracy of the model has improved by 40% and 10% for uttered word recognition and facial expression recognition, respectively. [ABSTRACT FROM AUTHOR]

At present, the deep learning super-resolution (SR) method has achieved excellent results, but it also faces problems such as large models, high computational cost, a large amounts of training data, and poor interpretability. However, traditional machine learning-based methods still have room for improvement in feature extraction and model structure. This paper constructs a gradient embedding cascade forest structure on the basis of random forest and proposes a limit gradient embedding cascaded forest SR (LGECFSR) model. In feature construction, we not only adopt the first-order gradient, the second-order gradient, and other features of the image but also fuse the information of the original LR image. In addition, image blocks of different sizes are used for training, which increases the model's generalization ability. Compared with the state-of-the-art machine learning-based methods, our method achieves the best performance and the second-best computational speed. In addition, compared with some deep learning-based methods, our model has a similar reconstruction effect and the best computational speed. In detail, for some reconstruction tasks, the Multi-Adds of LGECFSR is one-tenth to one-4000th of that of some current models. However, the SR performance of LGECFSR is the same or slightly better than that of some current classical algorithms. [ABSTRACT FROM AUTHOR]

This paper presents a novel end-to-end unsupervised deep learning approach for video anomaly detection. We propose to utilize the Perception Generative Adversarial Net (Perception-GAN), which is trained using the initial portion of the video. The generator of the perceptual-GAN learns how to generate events similar to the normal events, while the discriminator of the perceptual-GAN learns how to distinguish the abnormal events from the normal events. At testing time, only the discriminator is used to solve our discriminative task (abnormality detection). Through combining the generative adversarial loss and the proposed perceptual adversarial loss, these two networks can be trained alternately to solve the anomaly detection task. A two-stream networks framework and an update strategy is employed to improve the detection result. We test our approach on three popular benchmarks and the experimental results verify the superiority of our method compared to the state of the arts. [ABSTRACT FROM AUTHOR]

This paper provides a comprehensive analysis of the effect of speaking rate on frame classification accuracy. Different speaking rates may affect the performance of the automatic speech recognition system yielding poor recognition accuracy. A model trained on a normal speaking rate is better able to recognize speech at a normal pace but fails to achieve similar performance when tested on slow or fast speaking rates. Our recent study has shown that a drop of almost ten percentage points in the classification accuracy is observed when a deep feed-forward network is trained on the normal speaking rate and evaluated on slow and fast speaking rates. In this paper, we extend our work to convolutional neural networks (CNN) to see if this model can reduce the accuracy gap between different speaking rates. Filter bank energies (FBE) and Mel frequency cepstral coefficients are evaluated on multiple configurations of the CNN where the networks are trained on normal speaking rate and evaluated on slow and fast speaking rates. The results are compared to those obtained by a deep neural network. A breakdown of phoneme-level classification results and the confusion between vowels and consonants is also presented. The experiments show that the CNN architecture when used with FBE features performs better on both slow and fast speaking rates. An improvement of nearly 2% in case of fast and 3% in case of slow speaking rates is observed. [ABSTRACT FROM AUTHOR]

Electroencephalogram (EEG) signals, generated during the neuron firing, are an effective way of predicting such seizure and it is used widely in recent days for classifying and predicting seizure activity. But EEG signals generated during an epileptic seizure are highly nonstationary and dynamic in nature and contain very crucial information about the state of the brain. Due to this randomness, the accuracy of analysis of EEG data by conventional and visual methods is reduced drastically. This paper aims at enhancing epilepsy seizure detection using deep learning models with an FPGA implementation of the short-time Fourier transform block. Detection of seizure has been achieved in the following stages: (1) time–frequency analysis of EEG segments using STFT; (2) extraction of frequency bands and features of interest; and (3) seizure detection using convolutional neural network (CNN) and bidirectional long short-term memory (Bi-LSTM). For this work, the Bonn EEG dataset has been used. The maximum error of ~ 0.13% was encountered while the comparison of STFT output generated via proposed hardware architecture vs the output generated via simulation was done. The average classification accuracy of 93.9% and 97.2% was achieved by CNN and Bi-LSTM models, respectively, considering all frequency bands for epileptic and non-epileptic patients. [ABSTRACT FROM AUTHOR]

In this paper, we consider the problem of sparse signal recovery using a learned dictionary in multiple measurement vectors (MMVs) case. Employing deep neural networks, we provide two new greedy algorithms to solve sparse MMV problems. In the first algorithm, we create a stacked vector of measurement matrix columns and a new measurement matrix, which can be assumed as the Kronecker product of the primary compressive sampling matrix and a unitary matrix. In order to reconstruct sparse vector corresponding to this new set of equations, a four-layer feed-forward neural network is applied. In the second algorithm, joint sparse structure of the sparse vectors is considered. Recurrent neural networks are employed to extract the joint sparsity structure. In addition, we utilize an over-complete dictionary obtained from an unsupervised learning procedure. Simulation results illustrate the benefit of using the proposed methods. Finally, the proposed algorithms are applied for pilot-based channel estimation in massive multiple-input multiple-output systems to improve the channel state information recovery performance. [ABSTRACT FROM AUTHOR]

Image restoration and recognition are important computer vision tasks representing an inherent part of autonomous systems. These two tasks are often implemented in a sequential manner, in which the restoration process is followed by a recognition. In contrast, this paper proposes a joint framework that simultaneously performs both tasks within a shared deep neural network architecture. This joint framework integrates the restoration and recognition tasks by incorporating: (i) common layers, (ii) restoration layers and (iii) classification layers. The total loss function combines the restoration and classification losses. The proposed joint framework, based on capsules, provides an efficient solution that can cope with challenges due to noise, image rotations and occlusions. The developed framework has been validated and evaluated on a public vehicle logo dataset under various degradation conditions, including Gaussian noise, rotation and occlusion. The results show that the joint framework improves the accuracy compared with the single task networks. [ABSTRACT FROM AUTHOR]

In this paper, we combine superpixel and deep learning models to propose a new unsupervised image segmentation based on region-combined color images. Compared to other region merging algorithms, our algorithm can automatically segment color images without human interaction. The algorithm has three phases. In the first phase, we use the mean shift algorithm to obtain non-overlapping over-segmented regions. Firstly, the image is initially segmented by the superpixel segmentation algorithm, then the saliency map is obtained by the superpixel similarity, and the semi-supervised region is merged into an unsupervised algorithm by the saliency map. Finally, the resulting picture is sent to the deep learning model for training to get the final segmentation picture. A large number of experiments have been carried out, and the results show that the scheme can reliably extract the contour of the object from the complex background. [ABSTRACT FROM AUTHOR]

Image depth estimation is an important technology for obtaining scene depth for 3D images, and it has developed rapidly in the field of computer vision. In this paper, convolutional neural networks and conditional random fields are unified into a deep learning framework to build a computer vision model. First, based on the architecture of multiscale CNN and CRF, information of the scene image is obtained from computer video using depth learning, and then, a new frame model is built to predict the depth of images in a computer video. Second, three international standard datasets are used for training, and the results show that the CNN–CRF model can use a small number of samples to complete high-precision training and has a better estimation effect on images outside the dataset. The experimental results show that the CNN–CRF model can recover the depth and speculate the 3D structure of the scene, which has predictable application and development value for improving production efficiency. [ABSTRACT FROM AUTHOR]

Facial expression offers an important way of detecting the affective state of a human being. It plays a major role in various fields such as the estimation of students' attention level in online education, intelligent transportation systems and interactive games. This paper proposes a facial expression recognition system in which two channels of featured images are used to represent a 3D facial scan. Features are extracted from the local binary pattern and local directional pattern using a fine-tuned pre-trained AlexNet and a shallow convolutional neural network. The feature sets are then fused together using canonical correlation analysis. The fused feature set is fed into a multi-support vector machine (mSVM) classifier to classify the expressions into seven basic categories: anger, disgust, fear, happiness, neutral, sadness and surprise. Experiments were carried out on the Bosphorus database using tenfold cross-validation with mutually exclusive training and testing samples. The results show an average accuracy of 87.69% using an mSVM classifier with a polynomial kernel and demonstrate that the system performs better by characterizing the peculiarities in facial expressions than alternative state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Agriculture plays a significant role in the growth and development of any nation's economy. But, the emergence of several crop-related diseases affects the productivity in the agriculture sector. To cope up this issue and to make aware the farmers to prevent the expansion of diseases in crops and to implement effective management, crop disease diagnosis plays its significant role. Researchers had already used many techniques for this purpose, but some vision-related techniques are yet to be explored. Commonly used techniques are support vector machine, k-means clustering, radial basis functions, genetic algorithm, image processing techniques like filtering and segmentation, deep structured learning techniques like convolutional neural network. We have designed a hybrid approach for detection of crop leaf diseases using the combination of convolutional neural networks and autoencoders. This research paper provides a novel technique to detect crop diseases with the help of convolutional encoder networks using crop leaf images. We have obtained our result over a 900-image dataset, out of which 600 constitute the training set and 300 test set. We have considered 3 crops and 5 kinds of crop diseases. The proposed network was trained in such a way that it can distinguish the crop disease using the leaf images. Different convolution filters like 2 × 2 and 3 × 3 are used in proposed work. It was observed that the proposed architecture achieved variation in accuracy for the different number of epochs and for different convolution filter size. We reached 97.50% accuracy for 2 × 2 convolution filter size in 100 epochs, while 100% accuracy for 3 × 3 filter size which is better than other conventional methods. [ABSTRACT FROM AUTHOR]

Recent advances in digital audio source recognition, particularly within judicial forensics and intellectual property rights domains, have been significantly propelled by deep learning technologies. As these methods evolve, they introduce novel models and enhance processing capabilities crucial for audio source recognition research. Despite these advancements, the limited availability of high-quality labeled samples and the labor-intensive nature of data labeling remain substantial challenges. This paper addresses these challenges by exploring the efficacy of self-attention mechanisms, specifically through a novel neural network that integrates the Squeeze-and-Excitation (SE) self-attention mechanism for identifying recording devices. Our study not only demonstrates a relative improvement of approximately 1.5% in all four evaluation metrics over traditional convolutional neural networks but also compares the performance across two public datasets. Furthermore, we delve into the self-attention mechanism’s adaptability across different network architectures by embedding the Squeeze-and-Excitation mechanism within both residual and conventional convolutional network frameworks. Through ablation studies and comparative analyses, we reveal that the impact of self-attention mechanisms varies significantly with the underlying network architecture. Additionally, employing a transfer learning strategy has allowed us to leverage data from a baseline network with extensive samples, applying it to a smaller dataset to successfully identify 141 devices. This approach resulted in performance enhancements ranging from 4% to 7% across various metrics, highlighting the transfer learning method’s role in advancing digital audio source identification research. These findings not only validate the Squeeze-and-Excitation self-attention mechanism’s effectiveness in audio source recognition but also illustrate the broader applicability and benefits of incorporating advanced learning strategies in overcoming data scarcity and enhancing model adaptability. [ABSTRACT FROM AUTHOR]

This paper proposes a deep hybrid model for stereophonic acoustic echo cancellation (SAEC). A two-stage model is considered, i.e., a deep-learning-based Kalman filter (DeepKalman) and a gated convolutional recurrent network (GCRN)-based postfilter, which are jointly trained in an end-to-end manner. The difference between the proposed DeepKalman filter and the conventional one is twofold. First, the input signal of the DeepKalman filter is a combination of the original two far-end signals and the nonlinear reference signal estimated from the microphone signal directly. Second, a low-complexity recurrent neural network is utilized to estimate the covariance of the process noise, which can enhance the tracking capability of the DeepKalman filter. In the second stage, we adopt GCRN to suppress residual echo and noise by estimating complex masks applied to the output signal of the first stage. Computer simulations confirm the performance advantage of the proposed method over existing SAEC algorithms. [ABSTRACT FROM AUTHOR]

Recently, neural network-based deep learning methods have been popularly applied to computer vision, speech signal processing and other pattern recognition areas. Remarkable success has been demonstrated by using the deep learning approaches. The purpose of this article is to provide a comprehensive survey for the neural network-based deep learning approaches on acoustic event detection. Different deep learning-based acoustic event detection approaches are investigated with an emphasis on both strongly labeled and weakly labeled acoustic event detection systems. This paper also discusses how deep learning methods benefit the acoustic event detection task and the potential issues that need to be addressed for prospective real-world scenarios. [ABSTRACT FROM AUTHOR]

Deep learning-based speech enhancement approaches like deep neural networks (DNN) and Long Short-Term Memory (LSTM) have already demonstrated superior results to classical methods. However, these methods do not take full advantage of temporal context information. While DNN and LSTM consider temporal context in the noisy source speech, it does not do so for the estimated clean speech. Both DNN and LSTM also have a tendency to over-smooth spectra, which causes the enhanced speech to sound muffled. This paper proposes a novel architecture to address both issues, which we term a conditional generative model (CGM). By adopting an adversarial training scheme applied to a generator of deep dilated convolutional layers, CGM is designed to model the joint and symmetric conditions of both noisy and estimated clean spectra. We evaluate CGM against both DNN and LSTM in terms of Perceptual Evaluation of Speech Quality (PESQ) and Short-Time Objective Intelligibility (STOI) on TIMIT sentences corrupted by ITU-T P.501 and NOISEX-92 noise in a range of matched and mismatched noise conditions. Results show that both the CGM architecture and the adversarial training mechanism lead to better PESQ and STOI in all tested noise conditions. In addition to yielding significant improvements in PESQ and STOI, CGM and adversarial training both mitigate against over-smoothing. [ABSTRACT FROM AUTHOR]

Many classic chirp signal processing algorithms may show significant performance degradation when the signal-to-noise ratio (SNR) is low. To address this problem, this paper proposes a pre-filtering method in time-domain based on deep learning. Different from traditional signal filtering methods, the proposed denoising convolutional neural network (DCNN) is trained to recover the pure signal from the noisy signal as much as possible. Following denoising, we use two classic chirp signal parameter estimation algorithms to estimate the parameters of the DCNN output. The simulation results show that, compared with no DCNN processing, the parameter estimation accuracy is significantly improved. This is mainly due to the powerful pure signal extraction ability of DCNN, which can significantly improve the SNR and the accuracy of signal parameter estimation. [ABSTRACT FROM AUTHOR]

In this paper, a novel Hybrid Deep Ensemble (HDE) is proposed for automatic speech disfluency classification on a sparse speech dataset. Categorizations of speech disfluencies for diagnosis of speech disorders have so long relied on sophisticated deep learning models. Such a task can be accomplished by a straightforward approach with high accuracy by the proposed model which is an optimal combination of diverse machine learning and deep learning algorithms in a hierarchical arrangement which includes a deep autoencoder that yields the compressed latent features. The proposed model has shown considerable improvement in downgrading processing time overcoming the issues of cumbersome hyper-parameter tuning and huge data demand of the deep learning algorithms with high classification accuracy. Experimental results show that the proposed Hybrid Deep Ensemble has superior performance compared to the individual base learners, and the deep neural network as well. The proposed model and the baseline models were evaluated in terms of Cohen's kappa coefficient, Hamming loss, Jaccard score, F-score and classification accuracy. [ABSTRACT FROM AUTHOR]

COVID-19 virus has become a very critical human health hazard. Many variants are reported, and still, the virus is mutating. Thus, we get new strains now and then. COVID-19 detection at an early stage is an important issue that will help in the efficient management of the disease. This work studies COVID-19 audio signals originating from breathing, coughing, and vowel sounds. In the literature, most of the works on this topic use MFCC-based features. In this work, various methods are proposed for COVID-19 detection. The proposed methods use accumulated bispectrum features that capture the distinctive properties of COVID-19 in the above signals. Three new methods are proposed for COVID-19 detection. The performance of the presented methods is analyzed in detail, and comparison with the state-of-the-art methods is given. For various signals, considerable performance improvement is seen in the proposed methods. The CNN and ResNet-50 network models are used in this study. [ABSTRACT FROM AUTHOR]

Plant diseases significantly impact agricultural productivity. Early disease identification and diagnosis are critical for plant protection. Recent deep learning approaches have substantially aided detection of plant diseases. However, existing plant leaf disease identification techniques do not provide sufficient disease detection accuracies when the input image resolution is low. Images of grape leaves taken in the field may be low-resolution (LR) in nature due to limited lighting and varying weather conditions. Such LR images may result in incorrect real-time disease diagnosis. Hence super-resolution is a way to solve this problem. This research work proposes a novel Residual Skip Network-based Super-Resolution for Leaf Disease Detection (RSNSR-LDD) in the Grape plant. The input LR image is separated into two subcomponents using the guided filtering technique. Features from the two subcomponents are extracted using single convolutional layer and four layers of the two-channel residual skip network. Concatenation is used to combine the features of two channels. Following this, a decoding block and a convolutional layer are utilized to generate the super-resolution (SR) image. A new collaborative loss function is proposed for training. The obtained SR image is given to the Disease Detection Network (DDN) for grape leaf disease detection. This approach, based on an LR image, assists the farmer in spotting grape plant diseases very early. Proposed model was extensively trained and tested on PlantVillage, Grape 400, Grape Leaf Disease datasets with multiple super-resolution scaling factors for various grape leaves images. For different super-resolution scaling factors such as Χ2, Χ4, Χ6, the proposed model RSNSR-LDD achieved accuracies of 97.19%, 99.37%, and 99.06% for the PlantVillage dataset, 96.88%, 97.12%, and 95.43% for the Grape400 dataset, and 100% for the Grape Leaf Disease dataset. [ABSTRACT FROM AUTHOR]

Low-light enhancement plays an important role in overcoming image degradation and degraded performance of high-level computer vision tasks at night due to poorly illuminated environments. Existing low-light enhancement methods are often difficult to guarantee the accuracy of color and image details in complex low-light conditions. To solve the above problem, we design a low-light enhancement generative adversarial network with a dual-path design (DPGAN). We adopt a super-resolution subnetwork to provide high-gradient detail information of low-light images in one path, while a feature-forward network is designed in the other path. The feature-forward network consists of a feature extraction subnetwork and a feature refinement subnetwork; the former is designed for initial feature extraction, while the latter is then applied for deep refinement. Extensive experiments demonstrate the excellent performance of DPGAN in low-light image enhancement tasks. [ABSTRACT FROM AUTHOR]

The atmospheric conditions like rain degrade visibility, creating problems for computer vision applications. In single-image de-raining, the lack of temporal information creates challenges. Rain removal requires a high-frequency layer extraction, as rain affects an image's high-frequency layer (detail layer). This article has proposed using a combination of image decomposition and Inception ResNet v2 (IR v2) network for single-image rain removal. A guided filter decomposes the rainy image into the base and detail layers. The use of the Inception ResNet v2 (IR v2) network is proposed to create the residual map of the rain streaks from the high-frequency layer of the input image. The de-rained image obtained by subtracting the residual from the original rainy image provides comparable results. SSIM is added to MSE to train the IR v2 network, which improves the network's performance. The complexity of the network (IR v2) is high. A systematic reduction in the network is done first at the module level and later at convolution filters. The proposed refinement decreases network complexity and reduces execution time without degrading the performance. An extensive test using natural and artificial rainy images reveals that the proposed refined Inception ResNet v2 (rIR v2) favorably competes against the recent single-image de-raining techniques. [ABSTRACT FROM AUTHOR]

This study presents a computer-aided diagnosis (CAD) system that uses deep learning to diagnose migraine from electroencephalogram (EEG) signals. The proposed method converts EEG signals into scalogram images using the continuous wavelet transform method. Afterward, these scalogram images were given as input to the proposed convolutional neural network (CNN) deep learning network. With the proposed CNN deep learning model, 100% classification success has been achieved with very high accuracy compared to other related literature studies. Comparative analyses of the classification success of the proposed method and the classification performances obtained using state-of-the-art CNN models (VGG16, DenseNet121, ResNET101, and Xception) are presented. It has been observed that this study is superior to state-of-the-art CNN deep learning models compared with low computational complexity, simple network structure (less depth), and high classification performance. In addition, interpretable Gradient-weighted Class Activation Mapping (GradCAM) images were obtained to support the specialist in the diagnosis of migraine. In the study, it is revealed that there is a relationship between EEG recordings and migraine disease in terms of frequency components. When interpretable GradCAM images are examined, it is seen that mid- and high-frequency components are essential in distinguishing between migraine and healthy individuals. Considering the mentioned aspects, it is seen that this study has the potential to be used as a CAD system that can contribute to the expert opinion in distinguishing migraine patients and healthy controls. [ABSTRACT FROM AUTHOR]

Acoustic Scene Classification (ASC) aims to categorize real-world audio into one of the predetermined classes that identifies the recording environment of the audio. State-of-the-art ASC algorithms have excellent performance in terms of accuracy due to the emergence of deep learning algorithms. In particular, Convolutional Neural Networks (CNN) have set a new benchmark in ASC due to their promising performance. Despite the emergence of new frameworks, the interest in ASC is growing progressively with a shift of focus from enhancing accuracy to reducing model complexity. In this work, we introduce the AtResNet, a residual atrous CNN for low complexity acoustic scene classification. The AtResNet utilizes dilated convolutions and residual connections to reduce the number of model parameters. To further enhance the performance of AtResNet, we introduce a multi-scale feature representation method called multi-scale mel spectrogram (ms2). To compute the ms2, we evaluate the mel spectrogram on the wavelet subbands of the signal. We assessed AtResNet with ms2 on three benchmark datasets in ASC. The results suggest that our method significantly outperformed the CNN-based techniques in addition to a baseline system based on log mel spectrum for signal representation. AtResNet offers a 28.73% reduction in the model parameters against a baseline CNN. Furthermore, the AtResNet has a model size of 81 KB with post-training quantization of network weights. It makes AtResNet suitable for deployment in context-aware devices. [ABSTRACT FROM AUTHOR]

We propose in this work, a new approach for feature extraction based on deep Self-Organizing Map (SOM) network, named Generalized Unsupervised Deep SOM (G-UDSOM). This work presents an enhancement of the classic unsupervised deep SOM (UDSOM) algorithm in two ways. First, we modify the UDSOM Sub-sampling module in such a way that the image reconstruction phase is applied only in the feature construction phase. Second, the modified sub-sampling module learn feature of different sizes and resolutions through different map and patch sizes, which allows building parallel learning architecture. We add compact reconstruction constraint through the SOM learning based on convolutional auto-encoder. Thus. G-UDSOM allows extracting latent representation through its internal layer, by minimizing cost function. This function is composed of two terms: the cost SOM function mean square error of reconstruction. The second objective of our extended version is to achieve reduced computation time while improving accuracy and generalization capability. Different from the earlier proposed UDSOM architectures, G-UDSOM uses parallel SOMs without any reconstruction in the hidden layers. We evaluate our proposed approach on MNIST and STL-10 datasets. Experiment results show that our proposed approach outperforms deep SOMs models in terms of classification accuracy, training time and computational complexity. Our method is also can be generalized across different dataset without any pre-trained models. [ABSTRACT FROM AUTHOR]

The existing speech enhancement algorithm has shown poor performance under low Signal Noise Ratios (SNRs). To resolve this problem, a speech enhancement algorithm based on binaural sound source localization and cosh measure filtering is proposed. Firstly, the algorithm uses a sound source localization algorithm based on head correlation functions and two-level deep learning to extract the spatial information of the binaural sound source and determine the spatial position of the sound source. The beamforming method is then used to remove the noises in different directions from the speech. Finally, the Wiener filtering of cosh measure based on logarithmic relation is used to remove the noise in the same direction as the speech to achieve speech enhancement. Experiments show that the proposed algorithm has better robustness and denoising ability than the contrast algorithms. [ABSTRACT FROM AUTHOR]