PUFFER-DETR: Tiger puffer similar abnormal behavior recognition based on transformer

Fish behavior monitoring is crucial for fish farmers to obtain growth information, improve aquatic product quality, and adjust aquaculture strategies. However, the small size, severe occlusion, and similar behavior of fish pose challenges for identifying abnormal behavior. Therefore, this study proposes an abnormal behavior detection method based on PUFFER-DETR. Using the TripletAttention backbone network, the ability of the model to extract features of fish swarm behavior and individual fish behavior in turbid water has been improved. Furthermore, weight calculation is performed on the similar behavioral characteristics between individual fish and the behavioral characteristics of the fish groups to obtain a relationship feature map of similar behavior. Cross-scale feature fusion is performed using SHS-FPN, and the similarity behavior features of individual fish are adjusted to avoid the loss of similarity behavior features during the feature fusion process. Experimental results indicate that PUFFER-DETR achieved the best fusion accuracy at a speed of 127.9 frames per second, with an average accuracy of 92.8 %. This method can accurately detect abnormal behavior of fish and assist aquaculture personnel in judging the growth status of fish. Source code is available at https://github.com/DLOU-FishBehavior/PUFFER-DETR . • Novel Deep Learning model (PUFFER-DETR) for real-time detection of tiger puffer abnormalities in aquaculture. • The model is tailored for turbid water environments, high obstruction, and poor imaging conditions in aquaculture. • Utilizes the Inner-concept FocalerIoU loss function to enhance focus on minority samples and detection accuracy. • Surpasses mainstream YOLO algorithms in terms of average precision, accuracy, and detection speed. • Offers an efficient, automated solution for early warning of fish health issues in aquaculture.