Research On Appearance Quality Classification Method Of Apple Based On Octave Convolution And Fusion Attention Mechanism

Apples have high nutritional value,are rich in minerals and vitamins,and occupy a vital position in the world fruit market.However,during the process of planting,picking and transporting,apple is damaged by insects,intrusive external forces or improper storage,which may cause the quality of apple fruits to decline,related to the economic benefits of fruit farmers and the entire apple industry.Therefore,classification according to the quality of apple is of great significance for the subsequent storage and processing of apple and food safety.At present,apple quality is mainly sorted manually in china,which has the disadvantages of low efficiency,high cost and high labor intensity.In addition,the accuracy of classification is easily affected by artificial fatigue.In order to meet the needs of agricultural market and enterprises,finding an efficient and non-destructive automatic classification method to solve the problem of apple appearance quality classification,which became the conquer direction of agricultural product quality classification.A novel method based on octave convolution and fusion attention mechanism convolution neural network(OF-Net)is developed to classify apple appearance quality.A set of apple appearance quality classification system based on machine vision was built and image samples were collected through the system.Based on the design of the octave convolution branch and the fusion attention branch structure,the18-layer residual network is selected as the basic network.The octave convolution module is integrated to replace the original traditional convolution of the residual network,and the feature map low-frequency component(apple texture)and high-frequency component(apple defect)are separated,and then combined with the shape and contour characteristics of the apple defect part to classify,which solves the problem of network expression ability decline caused by inconsistent size and distribution of defects.The integration of the fusion attention module,and each feature channel is weighted to enhance the proportion of useful features,which solves the interference problem of noise features,and makes it focus on the shape,size and texture features of apple defects.At the same time,the PReLU activation layer and the batch normalization layer are used in the network structure to speed up the convergence rate of the network.Finally,the features extracted from the two branches are fused by point-by-bit addition,and the classification results are output through the Softmax layer.Experimental results showed that the prediction accuracy of the model reached97.67%,92.02%,94.62%,98.17% and 95.65% respectively of the date of the intact apple,scratches apple,russeting apple,insect bites apple,and rot apple.The overall classification and recognition accuracy rate reached 95.64%,which shows good performance on the task of apple appearance quality classification.