Human Action Recognition Network Based on Improved Channel Attention Mechanism

To tackle the problem that the existing channel attention mechanism uses global average pooling to generate channel-wise statistics while ignoring its local spatial information, two improved channel attention modules are proposed for human action recognition, namely the Spatial-Temporal (ST) interaction block of matrix operation and the Depth-wise-Separable (DS) block. The ST block extracts the spatiotemporal weighted information sequence of each channel through convolution and dimension conversion operations, and obtains the attention weight of each channel through convolution. The DS block uses firstly depth-wise separable convolution to obtain local spatial information of each channel, then compresses the channel size to make it have a global receptive field. The attention weight of each channel is obtained via convolution operation, which completes feature re-calibration with the channel attention mechanism. The proposed attention block is inserted into the basic network and experimented over the popular UCF101 and HDBM51 datasets, and the results show that the accuracy is improved.