Design Of Potato Defect Detection And Classification System Based On Convolution Neural Network

Potato is one of the important crops in China.At present,the classification of potatoes is mainly divided into two methods which include manual sorting and mechanical sorting.However,these two methods will cause certain losses in the sorting process.With the rapid development of machine vision and convolutional neural networks,this dissertation proposed the design of a potato defect detection and grading system based on convolutional neural networks.The Pto＿Net model and Pto＿SEResNe Xt model were designed for training.At the same time,the potato defect detection and grading were also designed in the dissertation.The research was mainly carried out from the following four aspects:(1)Construct a potato image acquisition platform through machine vision technology for clearly collection images containing potato defects was used to establish an original potato data set.However,a large amount of data was needed in the training process of the convolutional neural network.Therefore,data enhancement was performed on the basis of the original potato data set in order to obtain more potato images.Finally,an enhanced potato data set was established which was compared with the original potato data set.In the data set,the number of potato images had increased from 234 to 3744 which provides data support for the training of convolutional neural networks.(2)Construct the Pto＿Net model,which contains 17 convolutional layers,5 pooling layers and fully connected layers was the second aspect.In the convolutional layer,a small-scale convolution kernel of the size of was used and the BN layer and Relu activation function were added after the convolutional layer.At the same time,the Global Average Pooling layer was added before the fully connected layer.Then,it was classified by the Softmax function.After being trained on the potato dataset,the Pto＿Net model obtained 94.15%,80.95%,94.32%,and 84.32% accuracy rates on various potato defect classifications.The overall classification and recognition accuracy rate reached 87.83%.(3)After analyzing the principles of the residual network and the incentive network,the Pto＿SEResNe Xt model was proposed and the network model with deeper layers was obtained through the network fusion method.The network training parameters were less than the Pto＿Net model.After training on the potato dataset,the convergence speed and network performance of the Pto＿SEResNe Xt model had been significantly improved with 96.75%,92.77%,95.74%,and 94.59% accuracy rates in each category.The overall classification and recognition accuracy rate reached 94.92 %.(4)Design a potato defect detection and grading system,classify and recognize input images by calling the trained and saved model was the final part.The main functions of the system were image loading,model choosing,and data storage.The realization of the potato defect detection and classification system based on convolution neural network shows that the application of convolutional neural network to the classification and recognition of agricultural products has a broader prospect for improving the efficiency of agricultural products industry.