Research On Occluded Fish Detection Methods Based On Deep Learning

In recent years,my country’s aquaculture industry has developed rapidly,and has become an important part of my country’s national economy.The use of information technology to effectively manage the aquaculture industry is the need and trend of the development of the aquaculture industry.As the focus of aquaculture,the information management of fish farming is more representative.Fish target detection can assess the number of fish,and then provide corresponding data to support the management of fish farming.However,manual fish detection is inefficient and cannot meet the needs of large-scale aquaculture management.At this time,an efficient and automated fish detection method is needed.At present,the target detection algorithm is mainly based on some public data sets and land acquisition images for research.For underwater fish images,it is difficult to obtain high Average Precision(AP)for fish detection due to problems such as low contrast of underwater images and schools of fish gather and block each other.Therefore,this thesis studies the fish detection in the actual farm environment,the YOLOv4 algorithm with the best balance between detection accuracy and detection speed is used as the research benchmark.The main innovations are as follows:(1)In the part of feature extraction,this thesis proposes a multi-scale memory feature extraction network,which use Feature Memorizer(FM)to refine extraction and memory retention of the multi-level features extracted from the CSPDarknet53 network,enhance the transfer and retention of the contour and texture features of the fish,to make up for the loss of the fish feature information in the process of multi-level feature extraction and transmission from shallow to deep in the CSPDarknet53 network,and to better extract the feature information of the fish in the input fish image.Experiments show that using the feature extraction network proposed in this thesis,the AP value of fish detection is increased from73.50% of the YOLOv4 algorithm to 74.79%.(2)In the part of multi-scale feature fusion,this thesis draws on the semantic embedding branch in the image semantic segmentation network Ex Fuse,and proposes an improved semantic embedding branch,the semantic information of multiple deep features is introduced into the shallow features,which overcomes the problem of unbalanced information fusion caused by only using the semantic information of adjacent layers in the path aggregation network;at the same time,considering that sufficient detail information is required for fish position prediction,the detail information of the four-fold down-sampling feature layer is introduced into the deep features through convolution down-sampling.Experiments show that,the method in this thesis effectively enhances the degree of multi-scale feature fusion,enhances the expression ability of features,can better predict fish targets,and further increases the AP value of fish detection to 76.13%.(3)In the part of post-prediction processing,aimed at the problem that the original DIo U＿NMS algorithm has poor occlusion fish detection results,this thesis proposes an improved DIo U＿NMS algorithm to suppress repeated prediction box,which compensates for the missed detection of hidden fish when the fish is blocked,improves the recall rate of fish detection,and makes the AP value reach 77.99%.