Recognition And Counting Of Residual Bait Images Of Penaeus Vannamei Based On Neural Network

Penaeus vannamei is rich in nutrients and delicious in taste.It is loved by the majority of people in China and has high economic value.According to incomplete statistics,Penaeus vannamei accounts for65% of all marine shrimp farms,including freshwater cultured Penaeus vannamei,whose production has accounted for 80% of the total production of all cultured shrimp in China.In order to increase production,high-density aquaculture is often adopted.As the aquaculture density increases,feed residues and animal manure will directly affect the growth of prawns,causing serious water pollution and disrupting ecological balance.Therefore,the development of a shrimp residual bait identification system and the statistical work of residual bait quantity after the accurate identification of the bait will help to further understand the feeding law of white shrimp in aquaculture,reduce the cost of feeding and reduce water pollution.First,establish a full convolutional neural network(FCN)model for identifying residual bait;second,establish a sporadic residual bait counting model based on the fast labeling method of connected domains;third,establish a variety of applications based on deep learning technology The neural network model for counting the sticky bait targets,and the optimal model is selected.The main research contents and conclusions are as follows:(1)The original image collected by the digital camera is cropped into a small image of 320 pixels ×320 pixels,and the image is manually labeled using the online labeling software Labelme,and the image target is set as two kinds of labels for remnant bait and white shrimp.Input the original image and the labeled image(Maskmap)into FCN-8s for training to achieve automatic segmentation of the two types of targets.Experiments show that the FCN-8s model has achieved good recognition results,the training recognition accuracy rate reaches 99%,and the verification recognition accuracy rate reaches 98%.(2)Use a variety of counting models for the remnant bait count of the remnant bait image.The methods tried in this article include: Connected Component Quick Marking(CCL),Vgg-16 model neural network model counting and Nasnet-A model neural network image classification and counting method.According to the adhesion degree of different baits,the counting model is used for counting processing.When there is only a small amount of residual bait,the CCL method can be used to quickly count.The results show that,for the sporadic residual bait,the CCL model has the best recognition effect,with a counting accuracy of 97.5%.(3)For a large number of heavily adhered residual baits,they cannot be counted manually due to their dense distribution.The 320-pixel*320-pixel remnant image is further divided into 32-pixel*32-pixel small images to ensure that the number of remnants in each small image is between 0-5.The image data training uses shallow Densen neural network,Vgg-16,NASNet-A and other neural network models for modeling and analysis.The classification model results show that NASNet-A is the best counting method for residual bait.The accuracy of the prediction set counted by the NASNet-A model can reach 88.29%.