Human pose estimation (HPE) is crucial for computer vision (CV). Moreover, it’s a vital step for computers to understand human actions and behaviours. However, the huge number of parameters and calculations in the HPE model have brought big challenges to deploy to resource-constrained mobile devices. Aiming to overcome the challenge, we propose a sparse pruning method (SPM) for the HPE model. First, L1 regularisation is added in the training phase of the original model, and network parameters of the convolution layers (CLs) and batch normalisation layers (BNLs) are sparsely trained to obtain a network structure with sparse weights. We then combine the sparse weights of filters with the scaling parameters of the BNLs to determine their importance. Finally, the structured pruning method is used to prune the sparse filters and corresponding channels. SPM can reduce the number of model parameters and calculations without affecting precision. Promising results indicate that SPM outperforms other advanced pruning methods.

Face alignment is a common technology in face recognition and face verification field. Previous works mostly pay attention to improving the accuracy of prediction and ignored the practicability of the method. In this paper, we aim at providing a two-stage face alignment network for mobile platform. Firstly, the network was trained with residual label which is the difference between ground truth and mean shape. Secondly, the input data in the second stage is composed of the original data and generated heatmap which enriched the data types. Finally, a new loss function is used to enhance the convergence of local region. Experimental results show that our method not only provides high precision but also improve the real-time processing performance on the mobile platforms.

With the development of deep learning, neural networks are widely used in various fields, and the improved model performance also introduces a considerable number of parameters and computations. Model quantisation is a technique that turns floating-point computing into low-specific-point computing, which can effectively reduce model computation strength, parameter size, and memory consumption but often bring a considerable loss of accuracy. This paper mainly addresses the problem where the distribution of parameters is too concentrated during quantisation aware training (QAT). In the QAT process, we use a piecewise function to statistics the parameter distributions and simulate the effect of quantisation noise in each round of training, based on the statistical results. Experimental results show that by quantising the Transformer network, we lose less precision and significantly reduce the storage cost of the model; compared with the full precision LSTM network, our model has higher accuracy under the condition of a similar storage cost. Meanwhile, compared with other quantisation methods on language modelling task, our approach is more accurate. We validated the effectiveness of our policy on the WikiText-103 and PENN Treebank datasets. The experiments show that our method extremely compresses the storage cost and maintains high model performance.

The rapid development of computer vision technology provides a basic guarantee for public security reliance on video surveillance. In current video surveillance based on static cameras, accurate and quick extractions of foreground regions of moving objects enable quicker analysis of the behavior of meaningful objects and thus improve the intelligent analysis level of video surveillance. However, there would always occur false detection in the extraction of foreground regions, because of the shaking of tree branches and leaves in the scene and the “ghosting” area caused by the delayed updating of the background model. To solve this problem, this paper proposes a method for the extraction of foreground regions by using spatio-temporal information. This method can accurately extract foreground regions of moving objects by utilizing the difference and complementarity between spatial domain methods and temporal domain methods and further in combination with image processing technology. Specifically, the foreground regions of moving objects can be extracted by the morphological processing of the combination of the spatial information and the morphologically processed temporal information in the video. The experimental results show that the proposed method for the extraction of foreground regions of moving objects in view of the spatio-temporal information can reduce false detections caused by the shaking of tree branches and leaves, and thus effectively extract foreground regions of moving objects.

Summary: Head pose classification is an important part of the preprocessing process of face recognition, which can independently solve application problems related to multi‐angle. But, due to the impact of the COVID‐19 coronavirus pandemic, more and more people wear masks to protect themselves, which covering most areas of the face. This greatly affects the performance of head pose classification. Therefore, this article proposes a method to classify the head pose with wearing a mask. This method focuses on the information that is helpful for head pose classification. First, the H‐channel image of the HSV color space is extracted through the conversion of the color space. Then use the line portrait to extract the contour lines of the face, and train the convolutional neural networks to extract features in combination with the grayscale image. Finally, stacked generalization technology is used to fuse the output of the three classifiers to obtain the final classification result. The results on the MAFA dataset show that compared with the current advanced algorithm, the accuracy of our method is 94.14% on the front, 86.58% on the more side, and 90.93% on the side, which has better performance.