Action recognition based on attention mechanism and depthwise separable residual module

Aiming at the deficiencies of the lightweight action recognition network YOWO, a dual attention mechanism is proposed to improve the performance of the network. It is further proposed to use the depthwise separable convolution to replace part of the ordinary convolution after the 2D and 3D fusion and use residuals to merge the feature maps of different convolutional layers of the network, which improves the performance and speed of the network. First, to more effectively obtain salient features from the space and channel dimensions, add the CBAM space and channel attention module to the network; then, to make the parameters of the network more lightweight, it is proposed to use depthwise separable convolution to replace part of the ordinary convolution in the YOWO network. From experiments on the UCF101-24 and J-HMDB-21 datasets, compared with YOWO network, the improved method has significantly improved the accuracy and speed of action recognition.