Research And Application On Image Recognition Of Fruits And Vegetables Based On Deep Learning

Nanchang edible agricultural products electronic transformation platform is the concrete manifestation of the “Nanchang Edible Agricultural Products Safety Assurance Level Promotion Project”,which is being piloted in Nanchang's Shenzhen agricultural product market.When market practitioners use the platform,they need to manually distinguish and recognize fruit and vegetable varieties to report incoming goods and trade.However,manual recognition has problems such as judgment errors,input errors,low efficiency,consuming a large amount of labor and so on,which is not conducive to the operation of the platform.Therefore,using image recognition technology to develop fruit and vegetable image recognition systems instead of manual recognition has greater significance and application value.The content of this paper is to build a lightweight fruit and vegetable image recognition model based on deep learning technology,and perform lightweight operations such as pruning and quantization on the model to further improve the recognition speed of the model,and then develop a fruit and vegetable image recognition system based on the model.The main research contents and phased results of this paper are as follows:(1)Research and analysis on convolutional neural networks and lightweight networks.After the background explanation and research status analysis,this paper introduces deep learning models such as Res Net and Mobile Net in detail,and focuses on the residual structure of Res Net,as well as the improvements and advantages of the Mobile Net series of lightweight models,including deep separable convolution,linear activation function,inverted residual structure,h-swish activation function,etc.(2)Construction and data enhancement of the data set.Introduce the fruit and vegetable image data set used in this paper,as well as the enhancement operations of the image data set,including cropping,rotation and flipping,and the division of the data set in this paper.(3)Training and improvement of fruit and vegetable image recognition model.Set up the experimental environment and use the data set to train each model.An improved method of Mobile Net V3 model that uses global maximum pooling instead of global average pooling is proposed.Improve the model and train it.After training,the model is used to predict the test set,and the training results and prediction results are analyzed.Model lightweight operations such as model pruning and model quantization are introduced,the schemes of pruning and quantification are formulated,and then prune and quantify the model.The results of pruning and quantification are analyzed.(4)Design and implementation of fruit and vegetable image recognition system.This paper develops a fruit and vegetable image recognition system based on the improved Mobile Net V3 model.System analysis and system design are carried on based on the method of software engineering.According to the system design,the system is implemented and tested.