DarkActionNet: An efficient neural network for low-light animal behavior recognition

Yuming Chen, Tianzhe Jiao, Chaopeng Guo, Zhe Zhang, Jie Song

Abstract: Recognition of animal behavior in low-light environments plays a crucial role in medical experiments, smart agriculture, biological monitoring, and ecological research. However, the degradation of image quality under low-light conditions poses significant challenges to the accuracy and efficiency of behavior recognition. To address these issues, we propose an end-to-end network model named DarkActionNet, which integrates image enhancement and behavior recognition networks into a unified framework, significantly improving the accuracy and computational efficiency of animal behavior recognition in low-light conditions. We evaluated the model comprehensively on the Animal Kingdom and Kinetics-400 datasets and analyzed the performance improvements contributed by different key modules on the low-light mouse behavior dataset. Experimental results demonstrated that DarkActionNet achieved mAP/Top-1 accuracy of 34.1%, 80.8%, and 90.2% on the Animal Kingdom, Kinetics-400, and low-light mouse behavior datasets, respectively, while requiring only 50.2 GFLOPs and 30.8 MB of parameters, underscoring its potential for low-light animal behavior recognition tasks.