Research On Fish Lesion Image Segmentation Method Based On Deep Learning

As a large aquaculture country,aquaculture industry in China has an important position in the national economy,which is of great significance in increasing national economic income,raising employment rate,improving poverty,promoting social development and protecting ecological environment.In aquaculture,fish farming is an important part of it,but in fish farming,as the scale of farming gradually becomes larger and the density of farming increases,it is difficult to regulate the water quality of fish in the growth and development stage,coupled with the influence of the external environment,it is very easy to induce all kinds of infectious diseases,thus causing serious economic losses to aquaculture.China as a large country of culture and research of Pelteobagrus fulvidraco,but in the actual culture process of fish disease prevention and monitoring are mostly artificial,coupled with the fish living environment underwater,it is not easy to detect the occurrence of disease in the first time,thus missing the best treatment period,resulting in increased losses.Realising the smart development of the yellow catfish farming industry will effectively reduce the risk and cost of farming producers.In this paper,we study the segmentation of yellow catfish body surface disease spot images based on deep learning techniques,analyse and compare the performance of different semantic segmentation algorithms on fish disease spot image segmentation,and design an ESA-Attention U-Net network for yellow catfish disease spot segmentation,which improves on the basis of the Attention U-Net by incorporating the pyramid on the original attention mechanism pooling and self-attentiveness mechanisms to fuse features at different scales,while further improving segmentation performance by enhancing the attention mechanism.The experimental results show that the model can segment the diseased areas in the images more accurately,as follows:(1)Data set acquisition and annotation: data were obtained mainly by photographing farmed Pelteobagrus fulvidraco,including both diseased and healthy fish,and then the data were sorted and filtered,and the sorted data were annotated;(2)Comparing the performance of three semantic segmentation algorithms,U-Net,Attention U-Net and Swin U-Net,on the data set of sick spots of Pelteobagrus fulvidraco,through the analysis of different evaluation indexes,and selecting the network with the best overall performance as the target of subsequent improvement and optimisation;(3)Due to the better performance of the Attention U-Net network on the segmentation of yellow catfish spots,this study improved and optimized the Attention U-Net model,and compared the improved algorithm with the original model;(4)Based on the segmentation results of the improved model,the percentage of diseased areas in the image is obtained,which provides a reference for monitoring and early warning of fish diseases in aquaculture.