Detection And Recognition Of Underwater Fish Targets Based On YOLOv3

The detection and identification of fish have important research significance and value in the fields of aquaculture management,water environment monitoring,and fishery resources research.Traditional fish detection and recognition mainly use the combination of artificial design features and machine learning classifiers,and artificial features have the defects of difficult feature extraction,lack of versatility,and very time-consuming.In recent years,deep learning has made great achievements in the field of image classification,effectively improving the accuracy of image detection and recognition.Because deep learning can autonomously learn,and the accuracy of detection and recognition is high and robust.At present,the YOLOv3 algorithm has shown good performance in the field of target detection,and has been widely concerned by academia and industry.Therefore,this paper proposes a detection and recognition model of underwater fish targets based on YOLOv3 algorithm.The main research work of this paper is as follows:(1)Aiming at the inconsistency between the features of different scales under the multi-scale detection of YOLOv3 algorithm,the study uses an Adaptively Spatial Feature Fusion(ASFF)method to filter conflict information in space to suppress inconsistencies.For the features of a certain layer,first integrate and adjust the features of other layers to the same resolution,and then through training,adaptively learn the fusion space weights of feature maps of each scale to find the best fusion method.In addition,for the problem of unstable K-means clustering results of the priori box generation method of the YOLOv3 algorithm,a more advanced priori box is generated by adopting a more advanced clustering method,and IOU is used instead of the traditional Euclidean distance for distance measurement As a judging criterion;(2)For the problem of low positioning accuracy of the prediction loss function of the prediction box of the YOLOv3 algorithm,the study uses the Complete IOU(CIOU)loss as the prediction loss function of the prediction box,and comprehensively considers the distance,overlap rate,and scale are more in line with the prediction box regression mechanism,making the prediction box regression more stable.In addition,for the problem that the YOLOv3 algorithm has poor detection performance for scale diversity targets,by using the Spatial Pyramid Pooling(SPP)module,it is possible to extract multi-scale deep features with different receptive fields and connect them in series by Feature maps arefused in the channel dimension.The experimental results verify the effectiveness of the detection and recognition model of underwater fish targets based on the improved YOLOv3 algorithm.