Clothing Detection Action Recognition Based on Hierarchical Transformer Networks

In the application of industrial production operation action analysis based on computer vision, Transformer networks have the problem of large computational overhead and insufficient real-time performance. To address this issue and enable the network to be applied to clothing detection in industrial production operation recognition, a hierarchical Transformer network was constructed using network pooling and unit attention designs, which improved the accuracy of action recognition and focused on computing local information. A cosine re-weighting method was designed to replace the softmax operator in the self-attention mechanism of the Transformer network to reduce network overhead. The robustness of this network was verified on the HMDB51 dataset, achieving an accuracy of 91.7%, with a FLOPs of 42.8 G and 29 M parameters. Compared with other Transformer models, it significantly reduced network overhead while achieving high accuracy. For the analysis of clothing detection actions on a self-built dataset, an edge detection algorithm was added to the network to improve recognition accuracy. The improved network achieved a FLOPs of 22.9 G and 10 M parameters on the self-built clothing detection action dataset, with an accuracy of $\mathbf{9 7. 4 \%}$.