Concerning the missed detection and low recognition rate in the class attendance system based on face recognition, a method that combines student body detection and face angle filtering was proposed by applying the master and slave dual-camera device. First, the bodies of students were detected from the photograph of master camera by the Mask R-CNN algorithm. Then, the slave camera (PTZ (Pan/Tilt/Zoom) camera) was controlled to acquire high-quality magnified image of each student in turn. Next, the face poses were detected and recognized in the magnified images through MTCNN (Multi-Task Convolutional Neural Network) algorithm and FSA(Fine-grained Structure Aggregation)-Net algorithm in order to filter the frontal face image of every student. Finally, the FaceNet algorithm was used to extract the features of the filtered student frontal face images for training or recognition of Support Vector Machine (SVM) classifiers. Experimental results showed that, compared with the Tiny-face algorithm, when the Intersection over Union (IOU) was 0.75, the body detection algorithm had the Average Precision (AP) value increased by about 36% and the detection time reduced by 57%; compared with the method of establishing a multi-pose face database, the method of face angle filtering improved the recognition rate by 4%; and the accuracy of student recognition in the entire classroom was close to 100% in most cases. The proposed method simplifies the student registration process, improves the face recognition rate, and provides new ideas for solving the problem of face missed detection.

Aspect-level sentiment analysis aims to classify the sentiment of text in different aspects. In the aspect-level sentiment analysis of long text, the existing aspect-level sentiment analysis algorithms do not fully extract the features of aspect related information in the long text due to the redundancy and noise problems, leading to low classification accuracy. On the datasets with coarse and fine aspects, existing solutions do not take advantage of the information in the coarse aspect. In view of the above problems, an algorithm named TFN+BERT-Pair-ATT was proposed based on text filtering and improved Bidirectional Encoder Representation from Transformers (BERT). First, the Text Filter Network (TFN) based on Long Short-Term Memory (LSTM) neural network and attention mechanism was used to directly select part sentences related to the coarse aspect from the long text. Next, the related sentences were associated with others in order, and after combining with fine aspects, the sentences were input into the BERT-Pair-ATT, which is with the attention layer added to the BERT, for feature extraction. Finally, the sentiment classification was performed by using Softmax. Compared with the classical Convolutional Neural Network (CNN) based models such as Gated Convolutional network with Aspect Embedding (GCAE) and LSTM based model Interactive Attention Network (IAN), the proposed algorithm improves the related evaluation index by 3.66% and 4.59% respectively on the validation set, and improves the evaluation index by 0.58% compared with original BERT. Results show that the algorithm based on text filtering and improved BERT has great value in the aspect-level sentiment analysis task of long text.

Concerning the problem that distant historical signals cannot be transmitted to the current time under the network structure of Recurrent Neural Network (RNN), Long Short Term Memory (LSTM) network was proposed as a variant of RNN. On the premise of inheriting RNN's excellent memory ability for time series, LSTM overcomes the long-term dependence problem of time series and has a remarkable performance in natural language processing and speech recognition. For the long-term dependence problem of human behavior data as a time series and the problem of not real-time detection caused by using the traditional sliding window algorithm to collect data, the LSTM was extended and applied to the human posture detection, and then a human posture detection method based on LSTM was proposed. By using the real-time data collected by the accelerometers, gyroscopes, barometers and direction sensors in the smartphones, a human posture dataset with a total of 3336 manual annotation data was produced. The five kinds of daily behavior postures such as walking, running, going upstairs, going downstairs, calmness as well as the four kinds of sudden behavior postures of fallling, standing, sitting, jumping, were predicted and classified. The LSTM network was compared with the commonly used methods such as shallow learning algorithm, deep learning fully connected neural network and convolution neural network. The experimental results show that, by using the end-to-end deep learning method, the proposed method has improved the accuracy by 4.49 percentage points compared to the model of human posture detection algorithm trained on the produced dataset. The generalization ability of the proposed network structure is verified and it is more suitable for posture detection.