Apple Appearance Quality Grading Based On Residual Neural Network

The total production and planting area of apples in our country are among the top in the world,but we are at a disadvantage in international market competition due to the low level of commercial processing.Grading and sorting of fruits are important steps in improving commercial processing and enhancing the value chain of the industry.Currently,our country’s main grading methods—manual grading and mechanical grading—suffer from issues such as low grading efficiency,single grading features,and easy damage to fruits.However,deep learning-based grading methods can autonomously extract target feature information,reduce the reliance on feature engineering,and achieve end-to-end grading tasks.Compared to the aforementioned grading methods,deep learning can classify fruits of different grades more efficiently,accurately,and without causing any damage.Based on this,this paper focuses on researching apple appearance quality grading methods based on residual neural networks,with Fuji apples as the research subject.The main research contents of this article are as follows:(1)To address the problem of a lack of public and standardized datasets,and to ensure sufficient and diverse data samples,this paper created an apple appearance dataset.This image collection is built upon the initial dataset of apple appearance images created by Professor He Jinrong’s team.The dataset was divided into different grades of apple appearance quality based on the current "Fresh Apple" grading standards and the grading standards of the initial dataset.Then,some standardized and high-quality images were selected from the initial dataset,and two methods of data augmentation,geometric transformation and DCGAN,were used to create two different apple appearance datasets that differ from the initial dataset.The two datasets were compared and analyzed on four different network models,and the experimental results show that data augmentation based on geometric transformation can best represent its own features.Therefore,the dataset generated using this method was chosen as the final dataset for this study.(2)A lightweight improved residual neural network was proposed for apple appearance quality grading.This network model incorporates an improved residual module,in which a group information fusion module is designed to reduce the number of model parameters.To address the issue of the inability of information exchange between groups caused by grouped convolution stacking,the channel shuffle mechanism is introduced in the fusion module to merge information between subgroups and enrich the level of feature expression.In addition,to further improve the grading accuracy,the model incorporates a lightweight channel attention ECA module to adaptively recalibrate the feature response of the channels and enhance the weight of key channel features.Furthermore,the parallel pooling module is used to replace the maximum pooling layer in the model,in order to maximize the retention of input data representation features.Finally,experiments were conducted on the final dataset to compare this model with other models,and the results show that the improved residual model has good classification performance and the smallest number of model parameters.(3)The appearance features of premium apples in the dataset are relatively abstract and similar to adjacent grade fruit classification standards,resulting in a high number of misclassifications.To fully extract the appearance feature information of premium apples,this paper proposes a multi-scale attention residual neural network.The network model uses a multi-scale feature fusion module to extract features with different receptive fields,achieving a balance between local and global information.Additionally,the model incorporates a channel attention SE module in the residual module,enabling the network to explicitly model the relationships between different channels and assign channel weights to further extract discriminative feature information.While the above model improvements improve classification accuracy,they also lead to a significant increase in model parameters and computational complexity.To address this issue,the model reduces parameters by incorporating depthwise separable convolutions and decomposing the initial large convolution layer into multiple smaller convolutions.Experimental results demonstrate that the multi-scale attention residual model has better discriminability in apple appearance grading and sorting.