Non-destructive Testing Of Cotton Seed Appearance Quality Based On Machine Vision

The quality of seeds affects the yield of agricultural products.The quality inspection of seeds before planting is one of the effective ways to ensure the yield and quality of agricultural products.China’s cotton production is huge,and southern Xinjiang is an important cotton production base in China.The selection of excellent cotton varieties can increase the yield of cotton.At the same time,cotton production is also a pillar industry in this region.However,the quality sorting problem of lint free cotton seed seriously restricts the development of cotton industry in this region.Therefore,in this paper,according to the actual production requirements,the computer vision technology is applied to the quality detection of lint free cotton seeds.Taking Xinluzao 50 cotton seeds as the research object,an accurate and efficient non-destructive detection method of lint free cotton seeds is developed.The main research contents are as follows:(1)Image processing technology is used to preprocess the image of single lint cotton seed,cut the image,retain the center position of the cotton seed,extract the hog features of the cotton seed,and PCA principal component analysis is carried out.The 100 dimensional matrix obtained by the analysis is modeled,and the damage and integrity of the cotton seed are identified under the SVM linear classifier.The experiment shows that the hog features of the cotton seed can be used to identify the damage and integrity of the cotton seed SVM,LDA and KNN are used to analyze the extracted features respectively.The results show that SVM is the best,the accuracy rate is 93.5%,and the recall rate is94.5%.It shows that extracting hog features of cotton seeds and using SVM classifier modeling can identify single cotton seeds damage..(2)Using deep learning,an improved target detection algorithm based on the classic single shot multibox detector(SSD)algorithm is proposed,which can accurately identify the damaged cottonseed seeds in the group de fluffing cotton seeds.Using resnet50 network instead of VGg network in the classic SSD algorithm,as the basic network of SSD,can quickly extract the characteristics of cotton seed image.The experimental shows that the accuracy of the model is 96.1%,the recall rate is 97.3%,and the missing detection rate is 0,which is higher than 92.5%,96.4%,and 1.4% of the classical SSD network model.(3)Taking advantage of the transparency of acrylic board under strong light and white background,cotton seeds are slid into the groove of transparent acrylic board through the feeding device.With the rotation of the turntable,the two upper and lower horizontal angles are 45 ° apart,and the images of upper and lower sides of cotton seeds are collected by double CCD cameras respectively.Using improved yoov4 vision algorithm to detect damaged cotton seeds,the experimental shows that the detection accuracy rate of damaged cotton seeds and non damaged cotton seeds in group cotton seeds is 95.33%,the recall rate is 94.74%,and the missing detection rate is 0%;the detection effect is better than 91.67%,92.21%,1.67% of the original yoov4 network;the damage identification of double-sided group cotton seeds is realized.The built model has a detection accuracy of 93.5% for single cotton seed,96.1% for single-faced cotton seeds,and 95.33% for double-faced cotton seeds.