Thanks to the dividends of the rapid development of Internet + and agricultural modernization,the market for tree seeds in agricultural and sideline products continues to grow,and its processing process is gradually becoming automated and intelligent.As the mainstream of it,Xingyue Bodhi is a bead string obtained by Hainan yellow vine through the process of picking,shelling,polishing into beads,grading and selection,and stringing of the same level.Due to the lack of quantitative methods,the traditional manual loose bead grading and selection process often requires repeated comparison by naked eye,which is very cumbersome and timeconsuming and lacks accuracy,and processing enterprises are also facing various difficulties of high labor cost and low production efficiency.The contradiction between the current expanding market demand and the relatively backward processing technology deeply troubles the relevant practitioners.In order to solve the above problems,for the purpose of the research and application of Xingyue Bodhi hand string bead grading detection based on deep learning,the work is carried out in the following aspects:(1)Analyze the object detection and image segmentation in the field of computer vision and deep learning,compare the information and prediction effect that each model can provide in the scattered bead grading,and use the research idea of segmentation and then grading to preliminarily design the network model for the problem that the difference between the scattered beads at all levels is small.(2)Collect loose bead images by self-shooting and merchants,combined with market conditions and expert guidance of cultural play,and formulate a quantifiable loose bead grading method according to the proportion of area of monthly white space in key feature parts;Use Labelme to label the instances of loose beads and moon white space respectively,and add grade labels to the loose bead instances.Data are preprocessed by size normalization and sharpening filtering,and data is enhanced by geometric transformation and color space transformation.(3)Mask R-CNN with target positioning,classification and segmentation functions was selected as the basic model;The pixel statistics branch is added,the model output anchor frame and category information are integrated,and the fractionation mask is calculated to output the scatter bead grading detection results.(4)In order to improve the model segmentation performance to meet the requirements of loose bead grading,Res Net101 and Bi FPN were constructed to improve the basic model backbone network twice.The performance of each model was analyzed by control experiments,and the m Io U of the improved model split the instances of loose beads and moon white space reached 92%,which was 3% higher than that of the basic model,and the m AP of the fractional detection of loose beads reached 96.3%,which was 2.22% higher than that of the basic model and 8.45%higher than that of human vision.(5)Summarize the research results,design and develop a grading detection mini program for Xingyue Bodhi Loose Beads that can be put into application,and provide practical help for relevant practitioners. |