Study On Automatic Of Surface Diseases In Cage Cultured Fish Based On Deep Learning

Aquaculture is the fastest growing source of high-protein resources for global food production and is considered the most efficient and sustainable primary method to meet the growing demand for food,contributing to global economic development and social stability.Global aquaculture production reached 114.5 million tonnes,including 82.1 million tonnes of fish(e.g.finfish,shellfish and crustaceans).China is the world’s largest aquaculture producer,accounting for 57.50% of global aquaculture production in 2020,with farmed production reaching 52.24 million tonnes.Diseases in farmed fish are considered to be a major problem affecting the sustainable growth of many aquaculture species;therefore,the timeliness and accuracy of detection and identification of diseased fish are critical to prevent fish disease outbreaks and provide timely treatment,thereby avoiding massive fish mortality and reducing economic losses to fishermen and related aquaculture enterprises.Traditional fish disease monitoring requires manual real-time monitoring by staff with specialized knowledge,which has some limitations.Therefore,using image recognition technology combined with knowledge about sick fish detection,using computer analysis of fish body surface abnormalities in video to assess the health status of fish instead of manual fish disease detection,combined with deep learning methods can effectively achieve the detection of farmed sick fish in complex scenarios underwater,improve the accuracy of detection,and reduce the human and material resources consumed by manual fish disease detection.This paper addresses the problem that fish diseases are prone to occur in net tank aquaculture and that manual real-time monitoring of farmed fish health is extremely costly and difficult to achieve,resulting in high economic losses.In this study,we propose a method to automatically monitor the health status of cultured fish in net pens based on the images and video data of cultured fish from the Penghu No.1pelagic aquaculture net pens(Zhuhai Spider Island aquaculture fishery)and the Yellow Sea Long Whale No.1 pelagic aquaculture net pens(Dachin Island sea area,Changdao County)in 2021,using the YOLO v4 model based on deep learning.The main work and results are as follows.(1)An image acquisition system for fish diseases was designed and a fish disease dataset was constructed.Because the dataset of fish disease recognition based on deep learning is vacant,relying on expert guidance and monograph guidance,a web crawler-based image acquisition system was built to collect Baidu images,bing images and Google images as the web image dataset,combining photos taken at the fish market of Luchao Port(Nanhui New Town,Pudong New Area,Shanghai),photos taken at the pescadores No.1 pelagic aquaculture net tank(Zhaizhou Island,Zhuhai)and Yellow Sea long whale The fish disease dataset was formed by combining the photos taken at the fish market in Luchao Port(Nanhui Town,Pudong New Area,Shanghai),the photos taken at the pescadores(Spider Island,Zhuhai)and the videos taken by the underwater gimbal equipment at the Yellow Sea Long Whale No.1pelagic aquaculture net box(Dachin Island,Changdao County).The dataset includes650 images of normal fish,fish with water mold,fish with hemorrhagic disease,fish with cucumber,fish with ciliopathy and fish with Beneden’s disease,and is labeled with the fish disease dataset.(2)Due to the difficulty of acquiring data of farmed diseased fish and its special characteristics,we proposed fish data enhancement criteria based on the actual situation of underwater Gaussian blurring of fish,illumination changes and fish swimming angle,and enhanced the data of farmed fish disease data.On the basis of increasing the clarity of fish data by underwater defogging operation,and then image enhancement operations such as Gaussian blur,rotation,and random adjustment of brightness,the image data set was enhanced to 6050 images to realize the processing and enhancement of the labeled images.(3)In order to realize the recognition and classification of fish health in net tank culture,three kinds of target detection and recognition network models are commonly used: SSD target detection and recognition network model,Fatser R-CNN target detection and recognition network model and YOLO v4 target detection network,and three types of image datasets are trained for healthy fish,sick fish and dead fish,and the training results are analyzed for the The best detection model for fish disease detection in net tank culture,the attention mechanism is introduced and the impact of the attention mechanism on the sick fish detection task is discussed through evaluation metrics.(4)A YOLO v4 network model based on lightweight design is proposed for lightweight deployment of fish disease detection models on underwater head devices to achieve real-time detection of diseased fish.After comparing different feature extraction networks,the feature extraction module of YOLO v4 is replaced by Moblie Net v3 feature extraction module based on the channel attention mechanism,and the GELU activation function,which is a better performing activation function in GPT 2.0 network,is introduced to improve the recognition accuracy of similar features.The depth-separable convolution is used to replace the large convolutional blocks,which significantly reduces the number of network parameters and the amount of network operations.The experimental results show that the network enhances the intraspecific classification capability on the technology of the original model,and has significant advantages in hardware deployment,network learning capability,detection accuracy and detection speed of the lightweight network.In this study,an improved YOLO v4 network model for detecting fish body surface diseases is proposed by comparing three target detection networks and performing targeted optimization of the optimal network model for fishery use scenarios to solve the problem of difficult detection caused by too fast fish movement and small fish disease area features in real-time detection of farmed diseased fish.The research results of this study can provide scientific methods and techniques for real-time detection of farmed sick fish and promote the development of intelligent monitoring platforms for deep-sea nets.