Research On The Feeding Behavior Of The Pearl Gentian Grouper Based On Machine Vision

Our country is a traditional nation to raise aquaculture industry.Today,studying new technology and method for aquaculture,improving productive efficiency,reducing artificial cost,is of great significance to our country’s aquaculture industry.At present,the feeding method of fish is mainly artificial observation.This method determines the feeding amount mainly based on the subjective experience and consciousness of breeders,which is easy to cause under-feeding or over-feeding,underfeeding,slowing down the growth rate of fish,lengthening the breeding cycle,and causing certain economic losses.Excessive feeding will produce food residue,resulting in water pollution.With the change of breeding methods,the experience-based manual observation feeding method is not suitable for modern fine breeding,and modern breeding urgently needs a new efficient feeding method.To study the feeding rules of fish,improve feeding methods,increase the utilization rate of bait,formulate scientific feeding strategies,and improve the economic benefits of aquaculture are the focus of researchers and research hotspots in recent years.With the rapid development of science and technology,using computer vision technology to analyze the feeding behavior of fish has become an efficient and reasonable method and effective means.Factory farming has become an inevitable trend in the development of modern fisheries.The premise of realizing intelligent feeding is accurate assessment of fish feeding intensity.From the perspective of exploring precise feeding methods in circulating aquaculture,this paper proposes a method to evaluate feeding behaviors of fish by collecting feeding images,which provides a theoretical basis for precise feeding.The main work and conclusions of this paper are as follows:(1)The image segmentation method of fish swarm is proposed.In view of the high density of fish in the actual factory recirculating aquaculture environment,underwater images have drawbacks such as color bias,unbalanced contrast and blurring,and it is difficult to extract features.When traditional semantic segmentation network is used to segment fish images in complex environments,there are problems such as easy loss of deep information and not obvious detail extraction.Therefore,the Unet network is improved: The first is to add BN layer of batch standardization before activating function Relu to accelerate network convergence.The second is to add residual module in the process of network decoding to solve the problem of network degradation and gradient dispersion.The experimental results show that the accuracy of the model is 98.34%,that of m Io U is 94.6% and that of m PA is 97.07%.This method has the characteristics of low memory requirement,high recognition accuracy and fast recognition speed,and can accurately segment the image of feeding fish,which has important application value in accurate animal management and welfare breeding.(2)The evaluation model of feeding intensity of fish shoals was established.The improved Mobile Net V3 model is applied to the image recognition task of fish feeding,which solves the problem of poor robustness in traditional image processing methods.Based on the Mobile Net V3 model,the following improvements are made: 1)ECA attention module is used to replace the SE attention module of the original network,which effectively enhances the feature refinement ability of the model and improves the classification accuracy.2)The optimization of classifier removes the redundant structure in the original Mobile Net V3 model,reduces the number of parameters,and improves the accuracy of the model.Experimental results show that the accuracy of the improved model reaches 97.1 %.