Study On Fruit And Vegetable Images Recognition Based On Deep Learning And Its Application On Intelligent Refrigerator

Fruit and Vegetable images collected in intelligent refrigerator contain abundant information such as shape,quality and freshness of them.Recognizing these images intelligently can grasp the consumption of fruit and vegetable of users and analyze their preferences,merchants then can make fruit and vegetable products or personalized recipes recommendation by terminal device.At present,the refrigerator food identification mainly includes manual inputting type,bar code scanning,radio frequency identification on the method of manual operation,they are all cumbersome and of heavy workload.The traditional image recognition methods are required to manually design the feature extractor,which is of high requirement to image acquisition and noise,and is of one-sidedness and low rubostness of the extracted feature.As the representative of Deep Learning,convolutional neural network takes the raw image as input directly,it can autonomously select features from low-level to high-level and then to do the recognition function.It does not need to design the feature exacting algorithm and has a good resistance to the distortion or deformation of specific form of the image,so it's a good algorithm for processing image recognition of fruit and vegetable.This paper presents an algorithm for fruit and vegetable images recognition based on improved convolutional neural network and convolutional-recursive neural network,convolutional layer can extract feature autonomously by convolutional operation to reduces the manual intervention,overcome the triviality and one-sided problem;the sampling layer and recursive layer may greatly reduce the dimensionality of the image data to avoid the data dimension disaster.In this paper,the main work and innovations are as follows:(1)Systematically expounds the status of domestic and foreign research on fruitand vegetable identification technology,summarizes the development course of Deep Learning,research status at home and abroad and its achievements in the field of image recognition,focusing on the relevant algorithm principle of convolutional neural network and classical artificial neural network.(2)According to the characteristics of image recognition tasks,tricks of ReLU and Dropout are used to improve the conventional convolutional neural network.The network uses ReLU as activation function to prevent the divergence of gradient and avoid saturation phenomenon,accelerate the convergence of network as well.Introducing an appropriate proportion of Dropout to thin the hidden layer can improve the overfitting problem.The experiment is based on the Supermarket Produce Dataset,the recognition rate of the network is 5.1%higher than that of the same method,which proves the effectiveness of the tricks of the network to improve the recognition rate of fruit and vegetable images.(3)A joint convolutional-neural network based on ROI(Regions of Interest)image blocks for training convolutional kernels is proposed.The network selects target ROI from original image and trains feature detector(convolutional kernels)of convolutional layer based on the blocks of ROI,then it integrates Gray and RGB characteristics of convolution and recursive network to train the classifier so that enhancing the effectiveness of characteristics express.The experimental result shows that the convolutional kernels trained based on ROI can extract strong distinguishable feature and effectively improve the recognition rate by integrating multi-channel information of fruit and vegetable.(4)Elaborates the development history of intelligent refrigerator and the application of related technology,discusses the research significance of the recognition of fruit and vegetable in intelligent refrigerator,puts forward the module designments of intelligent refrigerator subsystem and the corresponding platform access scheme,studies the food management and fruit and vegetable recommended application model of the program.