Research On Fish Recognition Based On Deep Learning

According to statistics from the China Fisheries Statistical Yearbook,the total output value of the fishery economy of the whole society has steadily increased in recent years.At the same time as the scale of fisheries is growing rapidly,China's fisheries are gradually developing in the direction of fisheries automation and intelligent fisheries.As an important branch of fisheries automation,fish classification depends on the depth of traditional computer vision to manually set features.The cost is high and the accuracy rate is generally not ideal.In the context of fisheries automation,this paper combines deep learning and transfer learning to apply convolutional neural networks to fish image classification and recognition.Finally,a fish image recognition model based on residual network is proposed to obtain quality.With a higher underwater fish data set and verified fish recognition model,this experiment designed an underwater robot to complete the process from underwater image acquisition to host computer image recognition.The main work and results of this article are as follows:(1)Based on the construction of the fish recognition model,in order to better obtain the recognition model data set,this article designs an underwater shooting robot.By designing the control system of the underwater robot,adjusting the angle and position of the robot under the water is more clear.Photos of underwater fishes,at the same time provide support for the subsequent completion of underwater image shooting,uploading to the host computer,and image recognition model verification.In the design of the underwater control system,the selection of single chip microcomputer and motor drive chip is designed,and the appropriate camera module and data transmission method are selected to realize the control of the underwater robot and the transmission of the underwater image by the host computer.Through the underwater experiment of the robot,the underwater image shooting and transmission are completed,the underwater image is transmitted to the upper computer,and a suitable image is selected to make a data set.After the model training is completed,the images captured by the underwater robot are recognized,and the complete process from underwater fish image collection to classification model classification is completed.(2)This article uses two methods: underwater fish shooting pictures and Python crawler technology to crawl target fish pictures to create data sets,and manual screening to remove pictures that do not meet the standards.Three kinds of fish data sets have been produced.A total of 1080 pictures.In order to improve the training effect and prevent overfitting due to the data set,the data enhancement methods such as geometric transformation and noise addition are used to increase the diversity of training data and optimize the training effect.(3)Design and compare different fish classification models based on convolutional neural networks.This paper first introduces the model that selects the VGG16 network as the basic network,builds the network through transfer learning,and finally obtains the recognition model with the highest recognition accuracy of 94.25% by adjusting different network parameters and training skills.Then select the residual network Res Net50 as the pre-training network,build a residual network model through transfer learning,and obtain the classification model suitable for this data set after parameter training,and achieve 96.82% recognition accuracy.In order to obtain higher recognition accuracy,a method of introducing attention mechanism is proposed.By inserting the non-local operator in the attention mechanism into the residual network in the form of a module,through experiments,the attention-wise residual network is introduced The model recognition performance is the best,and the recognition accuracy reaches 98.16%.