The Method Of Fish Abnormal Behavior Detection Based On Deep Learning

China is a major fishing nation and the world’s top supplier of aquatic items.With the progressive improvement in demand for fresh aquatic goods and the continual expansion of aquaculture scale,there is a greater demand for aquaculture information technology in our country.Fish in aquaculture will exhibit a variety of stress responses,including avoidance,aimless swimming,and even mortality,in response to external environmental changes in the water temperature,hydrogen ion redox potential concentration,dissolved oxygen concentration,and chemical composition.Due to the high scientific significance and practical utility,it is crucial to make appropriate judgments on fish anomalous behaviors in time to detect varied water quality circumstances.First,using this feature extraction method to capture video abnormalities has problems like insufficient feature learning,difficult feature selection,poor generalization,etc.,which is no longer applicable to large-scale aquaculture.Second,the traditional manual extraction of abstract features is typically based on a priori knowledge extraction.Second,more researchers are using computer vision techniques in conjunction with fisheries research to study abnormal fish behavior,although the term “abnormal behavior” is perplexing because the data are distributed unevenly between normal and abnormal events.Ultimately,the definition of the same behavior differs in various contexts due to the diversity of anomalous samples,their categories,and their richness.Therefore,applying deep learning network models to fish abnormal behavior detection brings new opportunities and challenges to the development of fisheries,and using deep learning network models to automatically extract the motion correlation and appearance characteristics of fish and objectively reflect the fish motion status has become an urgent problem in realizing fish abnormality detection.Inspired by the research on anomaly detection in surveillance videos,this paper proposes two unsupervised deep learning models for fish video anomaly detection methods.To begin,in order to address the issue of insufficient utilization of temporal and spatial information in continuous video,this paper proposes an anomaly detection method based on temporal displacement and attention mechanism,employing Wider Resnet as the base network as the encoder for video frame feature extraction,residual temporal displacement on the temporal network to learn richer temporal information,and residual channel in the decoding stage attenuation.The experimental results show that the method’s frame-level AUCs on the benchmark datasets Ped1 and Shanghai Tech for video anomaly detection are 0.864 and 0.734,respectively,while the method’s frame-level AUCs on the fish dataset are 0.906 and 0.894,demonstrating the method’s effectiveness in detecting fish anomalous behaviors.Second,to solve the prior research technique’s concerns of low anomaly detection accuracy and poor generalization for huge scene data,this work offers an anomaly detection approach combining multilayer memory improvement and residual space-time transformer.We employ the U-Net network’s encoder and decoder to encode and decode video frames,then discover anomalies based on the difference between anticipated and real frames.The residual temporal transformer module and residual spatial transformer module are proposed to improve the network’s ability to model temporal and spatial information by strengthening the connectivity of spatio-temporal information features between consecutive video frames.Because the convolutional neural network is capable of generalization,the memory improvement module is employed instead of the jump link in the U-Net network to ease the encoder’s ability to represent aberrant frames.Furthermore,to address anomalies and increase the network’s detection accuracy,Generative Adversarial Networks(GAN)are employed to build more realistic prediction frames.The experimental results show that the frame-level AUCs of this method are 0.87 and 0.75 on the Ped1 and Shanghai Tech benchmark datasets,respectively,and 0.916 and0.921 on the zebrafish and red carp datasets,respectively,which effectively improve detection accuracy compared to other existing methods and have a more significant effect on the Shanghai Tech dataset.The findings of this research show that the deep learning model can detect anomalies in fish videos more effectively than previous methods,making it more appropriate for large-scale aquaculture.Meanwhile,this work highlights the model’s merits and shortcomings and discusses the future development of deep learning methods in fish video anomaly detection.