Peach Appearance Grading Method Based On Deep Learning

The grade processing of peaches is the main link in the market circulation.The appearance grade of peaches directly determines its economic value.At present,the appearance grading of peaches in China is still mainly based on manual grading and mechanical grading.The manual grading will waste a lot of manpower,and the mechanical grading is easy to cause the secondary damage of peaches,and it is difficult to guarantee the grading quality,which leads to the lack of competitiveness in the international market.Existing classification technologies mainly use image processing and traditional machine learning algorithms,and most of them use hyperspectral and other invisible light for detection.The application of CNN algorithms such as residual network has not been reported.Therefore,a higher accuracy grading method for peach appearance is of great significance.In this paper,seven kinds of common peach varieties in the market were taken as objects to study the grading method of peach appearance.The main research contents are as follows:(1)Construction of peach appearance grade data set.In order to test the peach appearance grade,a total of 540 samples of 7 types of peach varieties commonly found in the fruit wholesale market were selected in this paper,including 438 normal fruits and 102 defective fruits.A total of 2143 peach images were taken from different angles by image acquisition equipment.The peach images were classified according to the domestic peach grade classification standard.The peach appearance grade data set was established by clipping,rotation,filtering and segmentation,etc.(2)A grading method for peach appearance based on machine learning.The image processing algorithm was used to segment the peach foreground and extract the peach color,fruit shape,texture and other features.The machine learning peach appearance grade grading model and deep learning peach appearance grade grading model were established,and the validity of the model was tested.(3)Peaches appearance grading method based on deep learning.The Resnet34 model was studied and combined with the characteristics of peach appearance data set,and an improved model with multi-scale convolution was proposed.The experiment showed that the accuracy of the improved Resnet34 was 1.54 to 2.84 percentage points higher than that of the previous model.The VGG16 model was studied.Aiming at the problems of large number of model parameters and slow speed of training and reasoning,an improved scheme was proposed to replace the last pool layer with the convolutional layer and delete the redundant full connection layer.The number of model parameters decreased by 85.38% compared with the original VGG16,and the experiment showed that the improved VGG16 could still achieve the accuracy of the original VGG16.(4)A multi-model fusion grading method for peach appearance grade based on prior knowledge was proposed.The experimental results showed that the accuracy of this method for peach verification set and test set reached 95.32% and 95.57%.