Deep Learning Based Apple Surface Defect Detection Research

Apples are fruits with high vitamin and trace element content in our daily life,and are currently loved by people all over the world,and are grown quite widely,even becoming a major regional economic product.In the process of development,the apple industry is not only able to meet the daily needs of the general public,but also to drive local fruit farmers to increase their income and become rich.During the growth period,most apples may have different types of surface defects due to unfavorable natural conditions,resulting in poor quality and reduced yield,and farmers will suffer huge losses.The accurate identification of apple surface defects is a key part of the revitalization of rural industries.With the accelerated development of information technology,various new theories and methods have emerged,and it is a general trend to cross-fertilize the theoretical methods in artificial intelligence with the agricultural field.Taking apples in Aksu region as the research object,this paper conducts experimental research and analysis of apple surface defect detection by using theoretical knowledge and technical methods related to deep learning model for the problems of low accuracy and long time spent on apple surface defect detection recognition.It mainly includes the following aspects.（1）Establishment of apple surface defect image dataset.The original dataset was expanded by using data enhancement operations,and the image size was adjusted to 600 pixels × 600 pixels,and the image format was saved as JPG.All images are saved in PASCAL VOC format.（2）Construct a defect detection model on the apple surface.In order to more accurately identify the types of defects on the apple surface,this paper constructs an apple surface defect detection model using Faster R-CNN as the original model,which uses VGG16 as the feature extraction network.in order to avoid the arbitrary initialization of parameters and reduce the time required for the model during the training process,Under the migration learning method,the model and weights after the pre-training were obtained by conducting pre-training inside the Plant Village public dataset,and then transferred to the neural network model used in this paper.The experimental results show that the accuracy of the apple surface defect model constructed in this paper reaches 90.43% and the m AP is 89.64%.（3）Improvement based on the Faster R-CNN-based apple surface defect detection model.For some of the apple surface defects exist small defect area,containing the characteristic information is not obvious,resulting in relatively low detection accuracy.Firstly,the traditional convolution is replaced by the null convolution to expand the convolution kernel field of perception and capture the contextual information in the feature map in a better multi-scale way,while ensuring the same size of the output feature map;then the global average pooling layer is used to replace the fully connected layer to reduce the parameters in the network;then the batch normalization layer is added to normalize the data distribution and enhance the generalization ability of the network;finally,the Re LU activation function is replaced with Swish activation function to better improve the nonlinear expression in the network.The experimental results show that the improved Faster R-CNN＿VGG16-AL model achieves an average accuracy of 93.25% on average,and the average detection time is shortened to 203 s,which can better detect the defects on the apple surface.The m AP of the improved model on the original dataset and the expanded dataset increased by 5.29% and 4.02%,respectively,compared with the original model,which further verified the effectiveness of the improved model and could better achieve apple surface defect detection.