Research On Fish Lateral Line Scale Based On Convolutional Neural Network

Fish lateral line scale is one of the most important sensory organs and phenotypic attributes in fish body.The number of lateral line scales is an important basis for fishery researchers to classify fish.For a long time,the work of lateral line scale counting mainly depends on the researchers' manual work.It is low level of automation and easy to make mistakes,which has greatly hindered the development and progress of fishery breeding and research.As an objective and nondestructive detection method,machine vision technology has the incomparable advantages of traditional manual acquisition methods in obtaining the phenotypic attributes of fish.At present,the research on the acquisition of freshwater fishes based on machine vision focuses on the size,color,shape and other attributes.However,the research on the detection of lateral line scales is rarely reported.The main reason is that the color,texture and structure of the lateral line and non lateral line scales are little difference in the fish's surface,so it is difficult to distinguish them from the traditional method of machine vision.In order to solve the problem and provide a reference for further research on lateral line scale count,a method based on convolutional neural network for lateral line scale recognition is proposed in this paper.This paper researches and discusses from the aspect of sampling method,network structure and sampling strategy,and the main research contents are as follows:1)Using the improved LetNet-5 model to construct the convolutional neural network(CNN)for lateral line scales identification.Putting forward three kinds of different sampling size method according to the distribution of the lateral line scales on fish body surface.In addition,a lateral line scale recognition experiment based on manual feature extraction is designed to comparised with CNN.The results show that when under the same sample set,the CNN has obvious advantages in the task of recognizing the lateral line scales.2)Adjusting the Batch Size and the Kernel Size respectively in order to solve the over-fitting problem in the training process of CNN,and conducting a contrast experiment.The results show that when Batch Size=20,Kernel Size=5×5,accuracy of the model on validation set is 98.26%,compared to 97.39%before adjustment,increased by about 0.9 percentage points,and the losses of cost function are 0.0128 and 0.0624 before and after adjustment,reduced by nearly a half,it solves the over-fitting problem which is easy to occur in network training under small samples.3)Two sample expansion schemes are proposed to solve the problem that the recognition rate of the model is not high on the test set.Firstly,5400 samples of training set and 1035 samples of validation set are generated based on the expansion of the virtual samples by adding noise and geometric transformation.The model trained by this set of samples can recognize the lateral line scales of fish in the test set,and can effectively improve the positive sample recognition rate,but there are a large number of negative samples which have been wrongly identified.Secondly,an active sampling method based on the uncertainty sampling strategy and the deterministic error sampling method is proposed to make better from sampling strategy,improve sample labeling efficiency,rich sample structure,improve the test set lateral line scale recognition effects,and the problem of error identification negative samples has been changed dramatically.4)Finally,the Support Vector Regression(SVR)algorithm is used to fit the identified lateral line scales.In conclusion,the experiment results show that compared with the traditional manual feature extraction method,the CNN has obvious advantages in the task of the lateral line scale recognition.When the training samples are small,the network structure has an important influence on the experimental performance.Besides that,the extension of the virtual sample and the active learning sampling method can effectively improve the recognition rate of fish lateral line scales on the test set.Finally,fitting the lateral line to done the work of lateral line detection,which can serve as a reference for further research on automatic counting of lateral line scales.