Research On Fish Recognition Algorithm Based On Machine Vision

Fish identification is an important foundation for the development and utilization of marine fishery resources.It has a wide range of applications in aquaculture,fish monitoring,and fish sorting lines.The study of fish identification will also maintain the balance of marine ecological resources.In order to realize the automatic recognition of fish images,many problems are currently faced,such as impaired underwater image quality,fish data set is small,and the recognition accuracy is difficult to improve.All these bring difficulties to the research of fish identification methods.Therefore,through the research and analysis of domestic and foreign techniques and results,this article has launched an in-depth study on fish identification methods in view of the difficulties faced.The main work of this paper is as follows:(1)In the preprocessing stage of underwater images,aiming at the problem of image quality damage,this paper studies and improves on the basis of dark channel prior algorithm to realize underwater image enhancement.This article estimates the background light by finding the smallest difference point.Then estimate the transmittance of the three color channels according to the attenuation coefficient ratio,and use the median filter to optimize the transmittance.The method in this paper improves the contrast of underwater images,and enhances the texture,color and other feature information.Comparing this method with the classical enhancement algorithm,the subjective and objective quality evaluation shows that this method can obviously improve the quality of underwater damaged images,and make good preparations for fish recognition.(2)In the recognition method based on traditional machine learning,this paper makes full use of the edge,texture and shape features of fish images,and proposes a fish recognition algorithm based on feature fusion of ULBP and Gabor.First,segment the enhanced image,and the fish target contour is obtained by edge detection and boundary tracking.The mask map is generated by contour,thus the foreground fish image is segmented.Then,features are extracted by fusion of ULBP and Gabor.Compared with different feature fusion methods,this feature fusion method improves the classification and recognition performance.Finally,fish recognition is completed by SVM classifier.Experimental results show that,compared with traditional machine learning recognition methods,this algorithm has better performance in recognition accuracy and time.(3)In the recognition method based on convolutional neural network,aiming at the problem that the number of samples is small and the recognition accuracy is difficult to improve.This paper adopts the transfer learning method to design a network model suitable for the fish recognition task,avoiding the complex process of manual feature extraction.First,compare the recognition performance of different network models through experiments on the fish data set,select the Alex Net network as the pre-training model.Then,freeze the low-level convolutional layer,fine-tune the high-level,and adjust and optimize the model structure in terms of activation function,batch normalization layer and parameters.According to the experimental comparison,the performance of the model is improved by the fine-tuning method.Finally,the automatically extracted features are used to train the classifier to judge the fish species.The fish images are recognized online,and the accuracy and real-time performance of the recognition algorithm are improved in the case of small samples.(4)The two proposed recognition algorithms are compared and analyzed in terms of technology and performance.On the platform of Matlab and Jupyter Notebook,the research algorithms are verified by experiments,and an application program for fish recognition is designed,which has certain application value.The paper has 48 figures,15 tables,and 84 references.