融合注意力的轻量级行为识别网络研究

针对传统的三维卷积神经网络存在参数量多、信息冗余和时序信息提取不充分3个问题，提出了一种融合注意力的轻量级行为识别网络。首先，为轻量化网络参数和融合短中长时序信息，提出了高效残差块来替代两个级联的3×3×3卷积；其次，对通道注意力进行拓展，提出了时间注意力机制，并将两者嵌入在网络中抑制冗余信息对识别结果的影响；最后，在UCF101数据集上进行实验验证该网络的有效性。结果表明，提出的行为识别网络计算成本为8.9 GFlops,参数量为18.0 M,识别准确率为94.8%,与其他行为识别方法相比，以低成本的计算量实现了较高的识别准确率。

Research on lightweight action recognition network integrating attention

A lightweight action recognition network with fused attention is proposed to deal with the three problems of the traditional 3D convolutional neural network: large number of parameters, information redundancy and insufficient extraction of temporal information. First, in order to lighten the network parameters and fuse short-medium-long temporal information, an efficient residual block is developed to replace two cascaded 3×3×3 convolutions; second, by extending the channel attention mechanism, a temporal attention mechanism is derived, and both of the two mechanisms are integrated into the proposed network to suppress the influence of redundant information on recognition results; finally, experiments are conducted on the UCF101 dataset to verify the effectiveness of the network. The results show that the proposed action recognition network has a computational cost of 8. 9 GFlops, a parameter amount of 18. 0 M, and a recognition accuracy rate of 94. 8%, which reveals a high recognition accuracy with a low cost computation in comparison with other behavior recognition networks.