In order to solve the problem that the micro moving gesture features are not obvious and difficult to be identified, a micro moving gesture recognition method based on multi-scale fusion deep network for millimeter wave radar is proposed in this paper. The method is mainly composed of 2D convolution module, multi-scale fusion module and attention mechanism module. The multi-scale fusion module is composed of three residual blocks of different scales, which can obtain receptive fields of different sizes and obtain multi-scale features. Meanwhile, residual blocks of different scales are fused to increase the diversity of the network and better extract the deep features of the data. The Squeeze-and-congestion (SE) attention mechanism module is added to suppress the channel characteristics with little information. This improves the network identification accuracy and reduces the number of parameters and computation. The experimental results show that this method is simple to implement, doesn't need to do complex data preprocessing. The convergence speed of the network is fast, which can realize the effective recognition of the micro moving gesture.
Face detection and alignment are challenging operations due to variations in image angles, background lighting conditions, and intermediate blocking objects. Recent work has shown that these tasks can be improved by using a multi-task cascaded convolutional neural network (MTCNN) architecture. However, it is difficult to implement such an approach in a low-end edge AI system due to its high computational complexity. In this paper, a low cost, low power real-time face detection system based on FPGA MTCNN is designed. Firstly, the computational requirements of the algorithm are analyzed. Based on this analysis, a compromise solution is adopted. Data accuracy is a 16-bit floating point number. Meanwhile, the dual parallel method in and between convolutional kernels is used to reduce latency, and only the output of each level network is stored in block-on-chip RAM to reduce hardware resource consumption. Experimental results show that the system can significantly improve throughput, reduce time delay and power consumption, and Compared to 8-bit quantization, the detection accuracy is improved. The system takes into account the accuracy of the algorithm and hardware resources to realize real-time processing.
For the electrocardiograph correction and equipment maintenance, Traditional electrocardiogram(ECG) simulation methods have defects in fidelity and flexibility. In order to solve this problem, this paper proposes a comprehensive model of wavelet transform and Gaussian distribution, a variety of ECG with freely pre-set features could be simulated by this model, in addition to the flexibility and fidelity in generating simulated ECG. Secondly, the machine learning classification model was used to train the simulated five types of abnormal ECG. MIT-BIH ECG database was used as the test set to obtain the classification results to evaluate the quality of simulated ECG. Finally, the recognition accuracy of five kinds of ECG by the classifier is above 80%. It proves the feasibility of this study, which plays an essential role in the maintenance of a variety of ECG medical equipment and enhancement of ECG data template library.
With the development of Graph Neural Network (GNN), a lot of GNN-based methods have been proposed in text classification. However, GNNs are difficult to capture the word relationship of the context in the document sequence. In addition, these models are common to use one-hot encoding to initialize the node features, which have high dimensions and sparse features. Therefore, it is difficult to represent the correlation between texts. To effectively learn the semantic features of contexts and the association between words, we propose a text classification method based on the BERT-BiLSTM word embedding model and the Graph Convolutional Network (GCN), which is named BBG. The BERT-BiLSTM can learn word embedding representations in different contexts and the semantic relationship of text contexts. The learned representation vectors are provided into GCN as the features of document nodes. The feature representations of the training data and unlabeled test data are learned by performing the label propagation in GCN. Finally, the classification results are output by utilizing the Softmax layer. Experiments on four benchmark datasets show that our method has better performances than existing text classification methods.
In order to solve the problem of inaccurate results in the prediction of sleep stages by traditional networks, a DeepStageNet network was proposed. The network firstly express that deep characteristics through three-layer parallel CNN. Secondly, the extracted features are respectively transferred to the density at the next layer and further backward to the BiLSTM, which continues to transfer information and mine the dependency relationship between the various stages of sleep in the same way, thus realizing the automatic learning of the staging features of EEG signals and the determination of sleep periods. Finally, the experimental results on Sleep-EDF Database show that, after removing the influence of the data in the half-hour awake period before and after sleep, the Staging Accuracy of DeepStageNet Network on the traditiona130s data set and the new 210s data set after secondary processing are 86.5% and 87.7%. This effectively improves the Staging Accuracy of Sleep Model, and has certain practical value.