Abnormal Behavior Recognition of Underwater Fish Body Based on C3D Model

The behavior of fish is the direct embodiment of fish life. It is of great significance for the management of mariculture to recognize the abnormal behavior of fish. Traditional abnormal behavior monitoring uses manual monitoring, which will cost a lot of manpower and material resources. With the development of science and technology and the progress of the times, deep learning methods have made great progress in the field of video behavior recognition, which provides an opportunity for abnormal behavior recognition and detection. Firstly, in this paper, a two-category video dataset of fish behavior is constructed. Dataset consisting of 200 videos, including 100 videos of abnormal fish behavior and 100 videos of normal fish behavior. Among them, 150 videos are used for training the network model, and 50 videos are used for testing the model performance. Secondly, C3D deep network model is used to conduct an abnormal behavior recognition experiment on the self-made fish behavior dataset in a complex environment. Meanwhile, the softmax cross-entropy loss and the weight attenuation L2 regularization are used to calculate the total loss. For the experimental results of video classification, this paper makes a more intuitive visualization. ROC curve and t-SNE dimensionality reduction algorithm are used to analyze and evaluate the experimental results. The experiment results show that the C3D model we used has a better recognition effect on fish abnormal behavior and can be used to recognize fish abnormal behavior in actual aquaculture video surveillance.