Research On Apple Surface Defect Detection And Recognition Based On Deep Learning

China’s apple production ranks first in the world,yet high production volume and low quality has been the biggest obstacle facing the marketing and export of our fruit.It is difficult to avoid the problem of apples being crushed and damaged during a series of processes such as growing,picking,transporting,and storing.And these apples not only affect their own value,making the loss of nutrients,but also infect other normal flesh around them,bringing more serious economic losses to the farmers.Therefore,defect detection on the surface of apples has a strong practicality to realize automatic fruit sorting.In order to replace the manual detection with low accuracy,high labor intensity and low efficiency,this paper takes red Fuji apples as the research object and uses the techniques and methods related to deep learning models to study and analyze the defects on the apple surface.The research includes the following aspects:(1)Establishment of apple surface defects dataset.Images of apples with three types of defects: diseases and insect pests,mechanical injury,and rot are collected using both web crawler techniques and field photography.The self-built dataset is expanded to meet the requirements of model training using data augmentation methods.Manual annotation of all defect images is completed using annotation tools and made into MS COCO dataset format.(2)The DETR-based algorithm for apple surface defect detection is proposed.The Transformer network is applied to the apple surface defect detection task to address the problem that the feature extraction network ResNet50 of DETR has insufficient ability to extract certain important features as well as low efficiency.The channel attention mechanism SE Layer with deformable convolution(DCN)is used to improve CSP-Darknet53,and a feature extraction network with stronger feature characterization ability is obtained.For the problem that DETR is less effective in detecting small targets,deformable attention encoder is used instead of attention encoder,which allows the model to effectively utilize the multiscale feature maps containing information of many small targets to improve the detection ability of small targets.To further improve the effect of apple surface defect detection,a joint regression loss combining CIOU Loss and L1 Loss is proposed to solve the degradation problem caused by GIOU Loss and make the model have better convergence.The experimental results show that the improved DETR algorithm not only improves the detection accuracy of the model,but also reduces the number of parameters and the complexity of the model,and the mAP value is improved by 3.4% compared with the previous one.(3)Design of apple surface defect detection and identification software.Using PySide2 framework and Python language to design and implement apple surface defect detection and identification software,the software is mainly divided into user login module,defect detection module and real-time information display module.