MNv3-MFAE: A Lightweight Network for Video Action Recognition

Video action recognition aims to achieve the automatic classification of human behaviors by analyzing the actions in videos, with its core lying in accurately capturing the spatial detail features of images and the temporal dynamic features among video frames. In response to the problems of limited action recognition accuracy in videos containing complex temporal dynamics and large network model parameters, this paper proposes an innovative multi-feature fusion information modeling method. This paper designs a plug-and-play multi-feature action extraction (MFAE) module. The module adopts a multi-branch parallel processing strategy and integrates the functions of modeling and extracting temporal features, spatial features, and motion features to ensure the efficient modeling of the spatio-temporal information, inter-frame differences, and temporal dependencies of video actions. Meanwhile, the network employs a lightweight channel attention module (TiedSE), which reduces the complexity of the network model and decreases the number of network parameters. Finally, the effectiveness of the model is demonstrated on the Jester dataset, SomethingV2 dataset, and UCF101 dataset, achieving accuracies of 94.01%, 66.19%, and 96.74% with only 1.45 M parameters, significantly fewer than existing algorithms. The proposed method balances accuracy and computational efficiency in video action recognition, overcoming the shortcomings of traditional algorithms in temporal modeling and demonstrating its effectiveness in the task of video action recognition.