Research On Fish Recognition Based On Deep Learning

With the decline of marine fishery economic resources,coastal countries continue to strengthen the management and maintenance of fishery resources in their territorial waters.As one of the methods of fishery management monitoring,shipboard electronic monitoring can monitor the fishing behavior of fishing vessels for a long time without interruption,but it is necessary to manually check the catch by the massive monitoring video data stored by the fishing vessel and check whether the fishing boat has illegal fishing behavior.The accuracy of the judgment of manual review is easily plagued by internal and external factors such as visual fatigue,self-mood and working environment.Because of the increasing labor costs,shipboard electronic surveillance video review requires a method for automatic detection and classification of fish to be used for supplementary review.The traditional fish identification method is mainly through artificial design features.The design features are specific to specific products and have no universality.Artificial feature design requires expert experience and requires high professionalism for algorithm personnel.The limitations of the artificial design features lead to poor recognition of traditional methods in surveillance video scenes with complex backgrounds and illumination.In response to these problems,the thesis proposes a fish recognition method based on deep learning.The main work and results are as follows:1.This paper collects and produces seven fish species including albacore tuna,bigeye tuna,and yellowfin tuna,with a total of 4,520 electronic surveillance backgrounds,and a total of 5,424 fish targets.Because water fog can block the image in the electronic surveillance image,this paper uses the dark channel prior algorithm to enhance the image,thus improving the image color and visibility,reducing the impact of water fog on the annotation and image recognition of artificial data sets.2.Through the experiment of the target detection network based on the region proposal,the improved Faster-RCNN model is selected as the fish target detection network.Because of the slender characteristics of the fish target and the missed detection of the stacking too much,this paper optimizes the anchor window parameters of the RPN network and improves the fractional reconstruction function of the non-maximum suppression algorithm,and finally achieves betterresults.The fish detection effect,mAP reached 81.9%.3.In order to further improve the positioning and classification of fish targets,this paper proposes to divide the fish identification task into two sub-tasks.The first sub-task is fish target positioning,and the single-class target positioning through the Faster-RCNN network,thus cutting out the region of interest.The second sub-task is a fish classification,which will be fine-grained by the Inception-Resnet convolutional neural network with the cropped image input parameters reduced.This method combines the advantages of the target detection network and the classification network to achieve better results for the entire fish identification task.