Research On Apple Surface Defect Detection Technology Based On Multi-band Image

China has always been an important country for production of apples,with its planting area and output consistently leading the world,but its export ratio has been relatively low.This problem is mainly due to the outdated detection methods of pre-sales apples and the substandard quality of finished products.Therefore,there is important scientific significance and practical value to detect defects of harvested apples and realize apple grading.It can improve China’s apple export ratio and international competitiveness.Based on the multi-band images of apples,this paper detects surface defects and then grades the apples.The main research contents are as follows:(1)An image acquisition system is designed to obtain the images of apples.The respective advantages of visible light machine vision system and hyperspectral imaging system are considered,a multi-bands image acquisition system is designed.It combines a near-infrared CCD camera and filters to acquire images of apples.This system can provide five single-band images for each apple in the visible and near-infrared band ranges.(2)The traditional image algorithm is used to extract surface defects on apples.Firstly,the apple image is preprocessed by background removal and median filtering to avoid the influence of background and noise.Secondly,in order to solve the problem of uneven brightness distribution in the apple image,a ring-by-ring brightness correction algorithm is used.This algorithm can compensate for the edge grayscale of the apple and expand the contrast between the defect and non-defect areas.Then,this algorithm is optimized to reflect the real situation of defect regions and avoid mis-segmentation of stem.Finally,an optimized local threshold segmentation algorithm is used to extract apple surface defects.(3)An improved U-Net network is proposed to extract surface defects on apples.Firstly,data label and data augmentation are performed on the training set.Secondly,the trained U-Net is used to extract surface defects on apples.Then,in order to solve the shortcomings of U-Net during defect extraction,the network structure is modified by using dilated convolutions instead of ordinary convolutions in the down-sampling path and adding attention mechanism in the up-sampling path.Finally,a composite loss function is used to train the improved U-Net.Compared with the original U-Net,the improved U-Net focuses more on the segmentation of defect areas and defect edges,and improves the accuracy of extracting slight defects and irregular defect edges.(4)The defective apples are graded.Firstly,according to the national standard GB/T10651-2008,apples are divided into intact fruit,substandard fruit,and defective fruit.Then,a single pixel area calibration experiment is carried out to statistically analyze the defect area extracted by traditional algorithms,U-Net,Deep Lab V3+,and the improved U-Net.Finally,defective apples are graded.