TWE-C3D: Time Weighted 3D Convolutional Neural Network Based on ε -Pareto Dominance NSGA-III
Bin Cao, Shanshan Zhao, Yiming Hao, W.J. Liu · 2024
3D Convolutional Neural Network (C3D) for action recognition suffers from insufficient extraction of temporal information and limited generalizability. Therefore, we proposed a time weighted 3D convolutional neural network (TWE-C3D) based on evolutionary algorithm tuning parameters. To avoid uncontrollability of the parameter tuning process and enhance the generalization capability and accuracy, we introduced the ε-value and proposed a fast converging multiobjective ε-Pareto dominance and regional-angle enhanced NSGA-III (DR-NSGA-III) algorithm. In addition, based on the idea of moving average weighting of time information, we improved the time weighted shift block to better extract the temporal information of video image frames. Experimental results demonstrate that our algorithm outperforms the state-of-the-art algorithms in terms of search efficiency and prediction accuracy. Moreover, our improved time shift block can extract more temporal information and further enhance the classification accuracy of the C3D model.