Search

Caries may be halted or reversed in their progression by early detection, better hygiene habits, and coadministered drugs. The major clinical procedures for identifying dental caries are visual-tactile examination and dental radiography. However, due to their location, approximate caries exceedingly difficult to detect and affect the clinical assessment. Incorrect interpretations may also hinder the diagnostic procedure. Computational approaches and technology can be used to help dentists assess caries. Teledentistry has the ability to improve dental health care by providing access to dental care services from a remote location. Teledentistry helps identifying various stages of caries lesions using neural network and devices connected to the internet. This research develops an image classification for teledentistry systems using depthwise separable convolutional neural network. The trainable parameters reduction of depthwise separable convolution (DSC) successfully reduces the computational cost of conventional convolutional neural networks (CNN). As a result, the DSC model is reduced by 91.49% when compared to the traditional CNN model. Several DSC models improve conventional CNN accuracies in the training, validation, evaluation, and testing stages.

The semantic segmentation of apples from images plays an important role in the automation of the apple industry. However, existing semantic segmentation methods such as FCN and UNet have the disadvantages of a low speed and accuracy for the segmentation of apple images with complex backgrounds or rotten parts. In view of these problems, a network segmentation model based on deep learning, DeepMDSCBA, is proposed in this paper. The model is based on the DeepLabV3+ structure, and a lightweight MobileNet module is used in the encoder for the extraction of features, which can reduce the amount of parameter calculations and the memory requirements. Instead of ordinary convolution, depthwise separable convolution is used in DeepMDSCBA to reduce the number of parameters to improve the calculation speed. In the feature extraction module and the cavity space pyramid pooling module of DeepMDSCBA, a Convolutional Block Attention module is added to filter background information in order to reduce the loss of the edge detail information of apples in images, improve the accuracy of feature extraction, and effectively reduce the loss of feature details and deep information. This paper also explored the effects of rot degree, rot position, apple variety, and background complexity on the semantic segmentation performance of apple images, and then it verified the robustness of the method. The experimental results showed that the PA of this model could reach 95.3% and the MIoU could reach 87.1%, which were improved by 3.4% and 3.1% compared with DeepLabV3+, respectively, and superior to those of other semantic segmentation networks such as UNet and PSPNet. In addition, the DeepMDSCBA model proposed in this paper was shown to have a better performance than the other considered methods under different factors such as the degree or position of rotten parts, apple varieties, and complex backgrounds.

Insulator devices are important for transmission lines, and defects such as insulator bursting and string loss affect the safety of transmission lines. In this study, we aim to investigate the problems of slow detection speed and low efficiency of traditional insulator defect detection algorithms, and to improve the accuracy of insulator fault identification and the convenience of daily work; therefore, we propose an insulator defect detection algorithm based on an improved MobilenetV1-YOLOv4. First, the backbone feature extraction network of YOLOv4 ‘Backbone’ is replaced with the lightweight module Mobilenet-V1. Second, the scSE attention mechanism is introduced in stages of preliminary feature extraction and enhanced feature extraction, sequentially. Finally, the depthwise separable convolution substitutes the 3 × 3 convolution of the enhanced feature extraction network to reduce the overall number of network parameters. The experimental results show that the weight of the improved algorithm is 57.9 MB, which is 62.6% less than that obtained by the MobilenetV1-YOLOv4 model; the average accuracy of insulator defect detection is improved by 0.26% and reaches 98.81%; and the detection speed reaches 190 frames per second with an increase of 37 frames per second.

Vision is an important way for unmanned mobile platforms to understand surrounding environmental information. For an unmanned mobile platform, quickly and accurately obtaining environmental information is a basic requirement for its subsequent visual tasks. Based on this, a unique convolution module called Multi-Scale Depthwise Separable Convolution module is proposed for real-time semantic segmentation. This module mainly consists of concatenation pointwise convolution and multi-scale depthwise convolution. Not only does the concatenation pointwise convolution change the number of channels, but it also combines the spatial features from the multi-scale depthwise convolution operations to produce additional features. The Multi-Scale Depthwise Separable Convolution module can strengthen the non-linear relationship between input and output. Specifically, the multi-scale depthwise convolution module extracts multi-scale spatial features while remaining lightweight. This fully uses multi-scale information to describe objects despite their different sizes. Here, Mean Intersection over Union (MIoU), parameters, and inference speed were used to describe the performance of the proposed network. On the Camvid, KITTI, and Cityscapes datasets, the proposed algorithm compromised between accuracy and memory in comparison to widely used and cutting-edge algorithms. In particular, the proposed algorithm acquired 61.02 MIoU with 2.68 M parameters on the Camvid test dataset.

Residential electricity consumption forecasting plays a crucial role in the rational allocation of resources reducing energy waste and enhancing the grid-connected operation of power systems. Probabilistic forecasting can provide more comprehensive information for the decision-making and dispatching process by quantifying the uncertainty of electricity load. In this study, we propose a method based on stationary wavelet transform (SWT), quantile regression (QR), Bidirectional nested long short-term memory (BiNLSTM), and Depthwise separable convolution (DSC) combined with attention mechanism for electricity consumption probability prediction methods. First, the data sequence is decomposed using SWT to reduce the complexity of the sequence; then, the combined neural network model with attention is used to obtain the prediction values under different quantile conditions. Finally, the probability density curve of electricity consumption is obtained by combining kernel density estimation (KDE). The model was tested using historical demand-side data from five UK households to achieve energy consumption predictions 5 min in advance. It is demonstrated that the model can achieve both reliable probabilistic prediction and accurate deterministic prediction.

In this article, a remote sensing image change detection method based on depthwise separable convolution with U-Net is proposed, which omits the tedious steps of generating and analyzing the difference map in the traditional remote sensing image change detection method. First, two images having c -channel each can be specifically stacked into a 2 c -channel image, and the change detection can be converted to an image segmentation problem, an improved full convolution network (FCN) called U-Net is exploited to directly separate the changing regions. Because the capability of the deep convolution network is proportional to the depth of the network and a deeper convolution network means the increase of the training parameters, we then replace the original convolution in FCN by the depthwise separable convolution, making the entire network lighter, while the model performs slightly better than the traditional convolution operation. Besides that, another innovation in our proposed method is to use a preference control loss function to meet the different needs of precision and recall rate. Experimental results validate the effectiveness and robustness of the proposed method.

Depthwise separable convolution (DSC) significantly reduces parameter and floating operations with an acceptable loss of accuracy and has been widely used in various lightweight convolutional neural network (CNN) models. In practical applications, however, DSC accelerators based on graphics processing units (GPUs) cannot fully exploit the performance of DSC and are unsuitable for mobile application scenarios. Moreover, low resource utilization due to idle engines is a common problem in DSC accelerator design. In this paper, a high-performance DSC hardware accelerator based on field-programmable gate arrays (FPGAs) is proposed. A highly reusable and scalable multiplication and accumulation engine is proposed to improve the utilization of computational resources. An efficient convolution algorithm is proposed for depthwise convolution (DWC) and pointwise convolution (PWC), respectively, to reduce the on-chip memory occupancy. Meanwhile, the proposed convolution algorithms achieve partial fusion between PWC and DWC, and improve the off-chip memory access efficiency. To maximise bandwidth utilization and reduce latency when reading feature maps, an address mapping method for off-chip accesses is proposed. The performance of the proposed accelerator is demonstrated by implementing MobileNetV2 on an Intel Arria 10 GX660 FPGA by using Verilog HDL. The experimental results show that the proposed DSC accelerator achieves a performance of 205.1 FPS, 128.8 GFLOPS, and 0.24 GOPS/DSP for input images of size 224×224×3.

Sperm detection performance is particularly critical for sperm motility tracking. However, there are a large number of non-sperm objects, sperm occlusion and poorly detailed texture features in semen images, which directly affect the accuracy of sperm detection. To solve the problem of false detection and missed detection in sperm detection, a multi-sperm target detection model, Yolov5s-SA, with an SA attention mechanism is proposed based on the YOLOv5s algorithm. Firstly, a depthwise, separable convolution structure is used to replace the partial convolution of the backbone network, which can ensure stable precision and reduce the number of model parameters. Secondly, a new multi-scale feature fusion module is designed to enhance the perception of feature information to supplement the positional information and high-resolution of the deep feature map. Finally, the SA attention mechanism is integrated into the neck network before the output of the feature map to enhance the correlation between the feature map channels and improve the fine-grained feature fusion ability of YOLOv5s. Experimental results show that compared with various YOLO algorithms, the proposed algorithm improves the detection accuracy and speed to a certain extent. Compared with the YOLOv3, YOLOv3-spp, YOLOv5s and YOLOv5m models, the average accuracy increases by 18.1%, 15.2%, 6.9% and 1.9%, respectively. It can effectively reduce the missed detection of occluded sperm and achieve lightweight and efficient multi-sperm target detection.

An accurate and efficient automatic brain tumor segmentation algorithm is important for clinical practice. In recent years, there has been much interest in automatic segmentation algorithms that use convolutional neural networks. In this paper, we propose a novel hierarchical multi-scale segmentation network (HMNet), which contains a high-resolution branch and parallel multi-resolution branches. The high-resolution branch can keep track of the brain tumor’s spatial details, and the multi-resolution feature exchange and fusion allow the network’s receptive fields to adapt to brain tumors of different shapes and sizes. In particular, to overcome the large computational overhead caused by expensive 3D convolution, we propose a lightweight conditional channel weighting block to reduce GPU memory and improve the efficiency of HMNet. We also propose a lightweight multi-resolution feature fusion (LMRF) module to further reduce model complexity and reduce the redundancy of the feature maps. We run tests on the BraTS 2020 dataset to determine how well the proposed network would work. The dice similarity coefficients of HMNet for ET, WT, and TC are 0.781, 0.901, and 0.823, respectively. Many comparative experiments on the BraTS 2020 dataset and other two datasets show that our proposed HMNet has achieved satisfactory performance compared with the SOTA approaches.

Research on coal foreign object detection based on deep learning is of great significance to safe, efficient, and green production of coal mines. However, the foreign object image dataset is scarce due to collection conditions, which brings an enormous challenge to coal foreign object detection. To achieve augmentation of foreign object datasets, a high-quality coal foreign object image generation method based on improved StyleGAN is proposed. Firstly, the dual self-attention module is introduced into the generator to strengthen the long-distance dependence of features between spatial and channel, refine the details of the generated images, accurately distinguish the front background information, and improve the quality of the generated images. Secondly, the depthwise separable convolution is introduced into the discriminator to solve the problem of low efficiency caused by the large number of parameters of multi-stage convolutional networks, to realize the lightweight model, and to accelerate the training speed. Experimental results show that the improved model has significant advantages over several classical GANS and original StyleGAN in terms of quality and diversity of the generated images, with an average improvement of 2.52 in IS and a decrease of 5.80 in FID for each category. As for the model complexity, the parameters and training time of the improved model are reduced to 44.6% and 58.8% of the original model without affecting the generated images quality. Finally, the results of applying different data augmentation methods to the foreign object detection task show that our image generation method is more effective than the traditional methods, and that, under the optimal conditions, it improves APbox by 5.8% and APmask by 4.5%.

As a major daily task for the popularization of artificial intelligence technology, more and more attention has been paid to the scientific research of mental state electroencephalogram (EEG) in recent years. To retain the spatial information of EEG signals and fully mine the EEG timing-related information, this paper proposes a novel EEG emotion recognition method. First, to obtain the frequency, spatial, and temporal information of multichannel EEG signals more comprehensively, we choose the multidimensional feature structure as the input of the artificial neural network. Then, a neural network model based on depthwise separable convolution is proposed, extracting the input structure’s frequency and spatial features. The network can effectively reduce the computational parameters. Finally, we modeled using the ordered neuronal long short-term memory (ON-LSTM) network, which can automatically learn hierarchical information to extract deep emotional features hidden in EEG time series. The experimental results show that the proposed model can reasonably learn the correlation and temporal dimension information content between EEG multi-channel and improve emotion classification performance. We performed the experimental validation of this paper in two publicly available EEG emotional datasets. In the experiments on the DEAP dataset (a dataset for emotion analysis using EEG, physiological, and video signals), the mean accuracy of emotion recognition for arousal and valence is 95.02% and 94.61%, respectively. In the experiments on the SEED dataset (a dataset collection for various purposes using EEG signals), the average accuracy of emotion recognition is 95.49%.

Deep learning achieves substantial improvements in face detection. However, the existing methods need to input fixed-size images for image processing and most methods use a single network for feature extraction, which makes the model generalization ability weak. In response to the above problems, our framework leverages a cascaded architecture with three stages of deep convolutional networks to improve detection performance. The network can predict face in a coarse-to-fine manner. We replace the standard convolution with a combination of separable convolution and residual structure in the network. Extensive experiments on the challenging FDDB and WIDER FACE benchmarks demonstrate that our method achieves competitive accuracy to the state-of-the-art techniques while keeps real-time performance.

The overhead contact system (OCS) is an indispensable part of high-speed railway power supply system. The stable operation of the OCS depends on the stability of the contact between the pantograph and catenary. The dropper is an important guarantee for the stable contact. The three faults of the dropper, such as slack, breakage and disappearance can damage the quality of traction power supply and reduce the safety of the railway operation. Therefore, it is necessary to efficiently detect the dropper fault and guide the maintenance in time. Recently, automatic dropper fault detection methods based on monitor-video have been introduced to improve railway operation safety. However, the existing methods were still not stable enough in complex backgrounds. To improve the accuracy and real-time performance of the dropper fault detection, this paper proposed a method combining depthwise separable convolution with object detection network to detect the dropper fault. The proposed method consists of two stages. First, a dropper progressive location network (DPLN) was adopted to obtain the dropper. The DPLN was mainly composed of a pantograph location network (PLN) and a dropper location network (DLN). Then, the dropper fault recognition network (DFRN) was used to recognize the type of dropper fault. The experimental results demonstrated the accuracy and real-time performance of the proposed method.

Sound event detection (SED), which is typically treated as a supervised problem, aims at detecting types of sound events and corresponding temporal information. It requires to estimate onset and offset annotations for sound events at each frame. Many available sound event datasets only contain audio tags without precise temporal information. This type of dataset is therefore classified as weakly labeled dataset. In this paper, we propose a novel source separation-based method trained on weakly labeled data to solve SED problems. We build a dilated depthwise separable convolution block (DDC-block) to estimate time-frequency (T-F) masks of each sound event from a T-F representation of an audio clip. DDC-block is experimentally proven to be more effective and computationally lighter than “VGG-like” block. To fully utilize frequency characteristics of sound events, we then propose a frequency-dependent auto-pooling (FAP) function to obtain the clip-level present probability of each sound event class. A combination of two schemes, named DDC-FAP method, is evaluated on DCASE 2018 Task 2, DCASE 2020 Task4, and DCASE 2017 Task 4 datasets. The results show that DDC-FAP has a better performance than the state-of-the-art source separation-based method in SED task.

Camera/image-based localization is important for many emerging applications such as augmented reality (AR), mixed reality, robotics, and self-driving. Camera localization is the problem of estimating both camera position and orientation with respect to an object. Use cases for camera localization depend on two key factors: accuracy and speed (latency). Therefore, this paper proposes Depth-DensePose, an efficient deep learning model for 6-degrees-of-freedom (6-DoF) camera-based localization. The Depth-DensePose utilizes the advantages of both DenseNets and adapted depthwise separable convolution (DS-Conv) to build a deeper and more efficient network. The proposed model consists of iterative depth-dense blocks. Each depth dense block contains two adapted DS-Conv with two kernel sizes 3 and 5, which are useful to retain both low-level as well as high-level features. We evaluate the proposed Depth-DensePose on the Cambridge Landmarks dataset, which shows that the Depth-DensePose outperforms the performance of related deep learning models for camera based localization. Furthermore, extensive experiments were conducted which proven the adapted DS-Conv is more efficient than the standard convolution. Especially, in terms of memory and processing time which is important to real-time and mobile applications.

Currently, many methods of object pose estimation use images or point clouds alone for pose estimation. This leads to their inability to accurately estimate the object pose in the case of occlusion and poor illumination. Second, these models have large parameters and cannot be deployed on mobile devices. Therefore, we propose a lightweight two-terminal feature fusion network, which can effectively use images and point clouds for accurate object pose estimation. First, Pointno problemNet network is used to extract point cloud features. Then the extracted point cloud features are combined with the images at pixel level and the features are extracted by CNN. Secondly, the extracted image features are combined with the point cloud point by point. Then feature extraction is performed on it using the improved PointNet++ network. Finally, a set of center point features are obtained and pose estimation is performed for each feature. The pose with the highest confidence is selected as the final result. Furthermore, we apply depthwise separable convolutions to reduce the amount of model parameters. Experiments show that the proposed method exhibits better performance on Linemod and Occlusion Linemod datasets. Furthermore, the model parameters are small, and it is robust in occlusion and low-light situations.

The image classification based on cloud computing suffers from difficult deployment as the network depth and data volume increase. Due to the depth of the model and the convolution process of each layer will produce a great amount of calculation, the GPU and storage performance of the device are extremely demanding, and the GPU and storage devices equipped on the embedded and mobile terminals cannot support large models. So it is necessary to compress the model so that the model can be deployed on these devices. Meanwhile, traditional compression based methods often miss many global features during the compression process, resulting in low classification accuracy. To solve the problem, this paper proposes a lightweight neural network model based on dilated convolution and depthwise separable convolution with twenty-nine layers for image classification. The proposed model employs the dilated convolution to expand the receptive field during the convolution process while maintaining the number of convolution parameters, which can extract more high-level global semantic features to improve the classification accuracy. Also, the depthwise separable convolution is applied to reduce the network parameters and computational complexity in convolution operations, which reduces the size of the network. The proposed model introduces three hyperparameters: width multiplier, image resolution, and dilated rate, to compress the network on the premise of ensuring accuracy. The experimental results show that compared with GoogleNet, the network proposed in this paper improves the classification accuracy by nearly 1%, and the number of parameters is reduced by 3.7 million.

In order to improve the accuracy of underwater object classification, according to the characteristics of sonar images, a classification method based on depthwise separable convolution feature fusion is proposed. Firstly, Markov segmentation is used to segment the highlight and shadow regions of the object to avoid the loss of information caused by simultaneous segmentation. Secondly, depthwise separable convolution is used to learn the deep information of images for feature extraction, which produces less network computation. Thirdly, features of highlight and shadow regions are fused by the parallel network structure, and pyramid pooling is added to extract the multi-scale information. Finally, the full connection layers are used to achieve object classification through the Softmax function. Experiments are conducted on simulated and real data. Results show that the method proposed in this paper achieve superior performance compared with other models, and it also has certain flexibility.

WOS:000760108100003 The high dimensionality of hyperspectral remote sensing images (HRSI) affects the classification performance. Therefore, most HRSI classification methods use dimension reduction methods as a solution for high dimensionality. It is aimed to extract useful features with dimension reduction methods. At the end of this process, the data dimension is reduced and the transaction cost is decreased. In this study, LDA, PCA, IPCA, ICA, SPCA, RPCA and SVD dimension reduction methods were applied as a preprocessing step to improve HRSI classification performance. Since HRSI is volumetric data and has a spectral dimension, 2D CNN cannot extract good distinguishing features from spectral dimensions. Because 2D CNN only considers spatial information. With 3D CNN, spectral-spatial features are extracted simultaneously. However, 3D CNN increases the computational cost. Therefore, in this study, Hybrid 3D/2D CNN method is used together with dimension reduction methods. Hybrid CNN method consists of a combination of 3D CNN, 2D CNN and depthwise separable convolution. While 3D CNN extracts common spectral-spatial features, more spatial features are learned with 2D CNN used after 3D CNN. With depthwise separable convolution, it reduces the number of parameters and overfitting is prevented. The applications performed on the frequently used HRSI benchmark datasets show that the classification performance of the proposed method is better than the compared methods. In addition, Indian pines, HyRANK-Loukia, Botswana and Pavia of University datasets are used to examine the effect of dimension reduction methods used together with the hybrid 3D/2D CNN method on classification performance. As a result of the applications, the best classification accuracies were obtained in PCA, LDA and IPCA with Indian pines, PCA with Pavia of university, PCA and IPCA with Salinas, PCA, RPCA and LDA dimension reduction methods with HyRANK-Loukia.

Recent work has shown significant progress in the direction of synthetic data generation using Generative Adversarial Networks (GANs). GANs have been applied in many fields of computer vision including text-to-image conversion, domain transfer, super-resolution, and image-to-video applications. In computer vision, traditional GANs are based on deep convolutional neural networks. However, deep convolutional neural networks can require extensive computational resources because they are based on multiple operations performed by convolutional layers, which can consist of millions of trainable parameters. Training a GAN model can be difficult and it takes a significant amount of time to reach an equilibrium point. In this paper, we investigate the use of depthwise separable convolutions to reduce training time while maintaining data generation performance. Our results show that a DepthwiseGAN architecture can generate realistic images in shorter training periods when compared to a StarGan architecture, but that model capacity still plays a significant role in generative modelling. In addition, we show that depthwise separable convolutions perform best when only applied to the generator. For quality evaluation of generated images, we use the Fr\'echet Inception Distance (FID), which compares the similarity between the generated image distribution and that of the training dataset., Comment: 6 pages, 8 figures, To appear in the Proceedings of Southern African Universities Power EngineeringConference/Robotics and Mechatronics/Pattern Recognition Association of South Africa(SAUPEC/RobMech/PRASA), January 20-30 2019, Bloemfotein, South Africa

The super-resolution generative adversarial network (SRGAN) is a seminal work that is capable of generating realistic textures during single image super-resolution. However, the hallucinated details are often accompanied by unpleasant artifacts. To further enhance the visual quality, we propose a deep learning method for single image super-resolution (SR). Our method directly learns an end-to-end mapping between the low/high-resolution images. The method is based on depthwise separable convolution super-resolution generative adversarial network (DSCSRGAN). A new depthwise separable convolution dense block (DSC Dense Block) was designed for the generator network, which improved the ability to represent and extract image features, while greatly reducing the total amount of parameters. For the discriminator network, the batch normalization (BN) layer was discarded, and the problem of artifacts was reduced. A frequency energy similarity loss function was designed to constrain the generator network to generate better super-resolution images. Experiments on several different datasets showed that the peak signal-to-noise ratio (PSNR) was improved by more than 3 dB, structural similarity index (SSIM) was increased by 16%, and the total parameter was reduced to 42.8% compared with the original model. Combining various objective indicators and subjective visual evaluation, the algorithm was shown to generate richer image details, clearer texture, and lower complexity.