Video Anomaly Detection Of Fish Based On Self-supervised Deep Learning

Aquaculture is the main source of aquatic products in China.With the increasing demand for aquatic products and the continuous expansion of aquaculture scale,higher requirements are put forward for the automation and intelligentization of aquaculture in China.In the process of fish farming,anomalies in the external environment,such as abnormal water temperature,abnormal dissolved oxygen concentration,abnormal light condition and abnormal chemicals,will cause different abnormal behavior reactions,such as avoidance behavior,abnormal swimming,death,etc.If these abnormal behaviors of fish can not be found in time and handled properly,it will cause a lot of economic losses.Therefore,as one of the important links to strengthen the monitoring and early warning,risk assessment and emergency disposal of aquatic animal diseases,fish anomaly detection has high research value and practical value.With the promotion and gradual popularization of modern aquaculture modes such as deep sea cage aquaculture and industrialized circulating water aquaculture in China,monitoring and anomaly detection of fish manually is limited by the monitoring scope and monitoring time,which can not meet the needs of anomaly detection of fish stocks in the context of large-scale fish farms.Relying on the rapid development of computer vision technology and machine learning technology,in recent years,more and more researchers have carried out automatic fish stock anomaly detection through fish images and videos.However,due to the wide variety and different forms of fish stock abnormal behavior,it is usually difficult to cover all fish stock abnormal behavior with artificially designed features.These fish anomaly detection methods are difficult to detect fish anomaly comprehensively and effectively.In addition,some classification based fish video anomaly detection methods require a large number of normal fish video data and abnormal fish video data at the same time,so that the actual detection effect is limited by the completeness of abnormal fish video data,and the performance of the detection of the anomalies that does not exist in abnormal data is poor.Inspired by the research on surveillance video anomaly detection,this paper proposes two fish video anomaly detection methods based on self-supervised deep learning.Aiming at the problem of insufficient utilization of video temporal information and global spatial information in current prediction based video anomaly detection methods,this paper combines U-Net and Video Vision Transformer to propose video anomaly detection model Trans Anomaly to obtain richer temporal information and global context features in video clips.In order to make Video Vision Transformer suitable for image generation task,this paper improves the temporal and spatial Transformer Encoder in Video Vision Transformer and uses it for video future frame prediction.Through experiments on several video anomaly detection benchmark data sets,the effectiveness of the Transformer encoding module proposed in this paper is proved.Also,the sliding window size and stride applicable to each data set when calculating the anomaly score are determined.On the fish video anomaly detection data with red crucian carp as the experimental fish,Trans Anomaly proposed in this paper can effectively detect the motion anomalies of fish,and the AUC of the anomaly detection result is 0.919.When detecting the static anomalies of fish,such as dead fish,due to the characteristics of the video anomaly detection method based on prediction,the model failed to detect the static anomalies,and the AUC of abnormal detection result is 0.343.In view of the problem that the prediction based video anomaly detection model fails to detect the static anomalies of fish,this paper further proposes a video anomaly detection model FP-CFLOW based on feature prediction and conditional normalizing flow.Conditional normalizing flow has a good anomaly detection effect in the anomaly detection of industrial products,and similar to the anomalies of industrial products,the static anomalies of fish is usually manifested as the appearance change of fish.Therefore,this paper attempts to use conditional normalizing flow to improve the detection effect of static anomalies of fish.At the same time,the model predicts the image features,in order to maintain the detection effect of the model on fish motion anomalies.Experiments are carried out on the fish video anomaly detection dataset.The results show that FP-CFLOW significantly improves the anomaly detection performance of the model by adding the feature prediction module.Compared with the video anomaly detection model proposed in Chapter 3,the AUC of FP-CFLOW for static anomaly detection of fish is increased from 0.343 to 0.967.Although the AUC of motion anomaly detection is 0.848,which is lower than that of Trans Anomaly,FP-CFLOW has better detection performance on the whole test dataset,which is increased from 0.607 to 0.913.The research of this paper show that self-supervised deep learning models can effectively detect fish video anomalies.At the end of this paper,the advantages and shortcomings of the proposed models are summarized,and the future development direction of fish video anomaly detection method based on self supervised deep learning is prospected.