Application Of Fish Recognition Based On Convolutional Neural Network

With the continuous improvement of material living standards,people are more and more concerned about the source and nutritional value of food.However,with the increasing variety of foods,it has become fun and challenging to identify the category of food bought at the market.Convolutional Neural Network has been developing on image recognition tasks,and the accuracy rate on specific tasks can be comparable to humans.In this paper,a convolutional neural network is applied to the problem of fish image recognition,and based on the actual application requirements,the recognition algorithm is implanted on a mobile device equipped with a camera to design an independent application that does not rely on cloud resources.The main tasks completed are as follows:(1)Through the use of web crawler technology and mobile terminal data acquisition applications,we collected a large amount of candidate pictures.Then,to solve the problem of insufficient samples and imbalanced categories,we preprocessed the image with the data augmentation and oversampling techniques,and finally a class-balanced edible fish image dataset was obtained.(2)Taking the Mobile Net V2 network structure as a benchmark model,it is determined to use NAdam as an optimization algorithm.On this basis,the characteristics of the Cross Entropy loss function and Focal Loss function were compared and analyzed,and experiments were designed to verify the performance difference between the two on the edible fish image dataset.Finally,it is determined to use the focus loss function as the network loss function.The threshold shift method is used to further reduce the effect of class imbalance on algorithm accuracy.Finally complete the training to get the classification model.(3)Use the Tensor Flow Lite library to complete the conversion of the model from the PC to the mobile end.On this basis,further deployment and optimization were completed,and the entire process from calling the camera to take pictures to locally identifying images and feeding back the processed results to the user was completed,and the development practice of fish image recognition applications on Android was realized.