Search

This paper presents a novel sound event detection (SED) system for rare events occurring in an open environment. Wavelet multiresolution analysis (MRA) is used to decompose the input audio clip of 30 seconds into five levels. Wavelet denoising is then applied on the third and fifth levels of MRA to filter out the background. Significant transitions, which may represent the onset of a rare event, are then estimated in these two levels by combining the peak-finding algorithm with the K-medoids clustering algorithm. The small portions of one-second duration, called 'chunks' are cropped from the input audio signal corresponding to the estimated locations of the significant transitions. Features from these chunks are extracted by the wavelet scattering network (WSN) and are given as input to a support vector machine (SVM) classifier, which classifies them. The proposed SED framework produces an error rate comparable to the SED systems based on convolutional neural network (CNN) architecture. Also, the proposed algorithm is computationally efficient and lightweight as compared to deep learning models, as it has no learnable parameter. It requires only a single epoch of training, which is 5, 10, 200, and 600 times lesser than the models based on CNNs and deep neural networks (DNNs), CNN with long short-term memory (LSTM) network, convolutional recurrent neural network (CRNN), and CNN respectively. The proposed model neither requires concatenation with previous frames for anomaly detection nor any additional training data creation needed for other comparative deep learning models. It needs to check almost 360 times fewer chunks for the presence of rare events than the other baseline systems used for comparison in this paper. All these characteristics make the proposed system suitable for real-time applications on resource-limited devices. [ABSTRACT FROM AUTHOR]

The recognition of food images is of great significance for nutrition monitoring, food retrieval and food recommendation. However, the accuracy of recognition had not been high enough due to the complex background of food images and the characteristics of small inter-class differences and large intra-class differences. To solve these problems, this paper proposed a food image recognition method based on transfer learning and ensemble learning. Firstly, generic image features were extracted by using the convolutional neural network models (VGG19, ResNet50, MobileNet V2, AlexNet) pre-trained on the ImageNet dataset. Secondly, the 4 pre-trained models were transferred to the food image dataset for model fine-tuning. Finally, different basic learner combination strategies were adopted to establish the ensemble model and classify feature information. In this paper, several kinds of experiments were performed to compare the results of food image recognition between single models and ensemble models on food-11 dataset. The experimental results demonstrated that the accuracy of the ensemble model was the highest, reaching 96.88%, which was superior to any base learner. Therefore, the convolutional neural network model based on transfer learning and ensemble learning has strong learning ability and generalization ability, and it is feasible and practical to apply the method to food image recognition. [ABSTRACT FROM AUTHOR]

Long short-term memory (LSTM) has been effectively used to represent sequential data in recent years. However, LSTM still struggles with capturing the long-term temporal dependencies. In this paper, we propose an hourglass-shaped LSTM that is able to capture long-term temporal correlations by reducing the feature resolutions without data loss. We have used skip connections in non-adjacent layers to avoid gradient decay. In addition, an attention process is incorporated into skip connections to emphasize the essential spectral features and spectral regions. The proposed LSTM model is applied to speech enhancement and recognition applications. The proposed LSTM model uses no future information, resulting in a causal system suitable for real-time processing. The combined spectral feature sets are used to train the LSTM model for improved performance. Using the proposed model, the ideal ratio mask (IRM) is estimated as a training objective. The experimental evaluations using short-time objective intelligibility (STOI) and perceptual evaluation of speech quality (PESQ) have demonstrated that the proposed model with robust feature representation obtained higher speech intelligibility and perceptual quality. With the TIMIT, LibriSpeech, and VoiceBank datasets, the proposed model improved STOI by 16.21%, 16.41%, and 18.33% over noisy speech, whereas PESQ is improved by 31.1%, 32.9%, and 32%. In seen and unseen noisy situations, the proposed model outperformed existing deep neural networks (DNNs), including baseline LSTM, feedforward neural network (FDNN), convolutional neural network (CNN), and generative adversarial network (GAN). With the Kaldi toolkit for automated speech recognition (ASR), the proposed model significantly reduced the word error rates (WERs) and reached an average WER of 15.13% in noisy backgrounds. [ABSTRACT FROM AUTHOR]