A Research On Recognition Method Of Bobbin Based On Computer Vision

In traditional textile industry,the sorting at the end of bobbin production is highly dependent on manual sorting,which can not meet the production demand of high efficiency and intelligence.Therefore,it is of great significance to study the automatic and accurate identification method of bobbin.This thesis takes the bobbin image recognition technology that relies on the core sticker labeling as a research topic.Aiming at the problems of low accuracy and small application area of the existing bobbin recognition system,a new automatic detection method is designed based on computer vision technology.This thesis mainly studies the shape area detection and pattern shape classification of the bobbin image under the industrial background.The specific contents are as follows:(1)This thesis studies the precise positioning method of shape region in bobbin image,which is divided into the detection of sticker region and shape region.On the premise that the classical target detection algorithm can not locate the sticker area,this thesis starts from the shape characteristics of the bobbin,finds the upper edge of the bobbin through edge detection and contour search,then finds all the ellipse areas by ellipse fitting,and then finds the ellipse area corresponding to the sticker on the upper edge of the bobbin by using the self-designed screening rule,and designs a flattening algorithm to convert it into a rule.In this thesis,a double threshold segmentation algorithm based on HSV color space histogram analysis is proposed.Compared with the classical threshold segmentation algorithm,the algorithm in this thesis has better performance.(2)The color cast detection and correction method of bobbin image is studied in this thesis.In order to solve the possible color cast phenomenon of bobbin image due to the influence of imaging devices,the classic brightness algorithm is used to detect and correct the color cast of bobbin image.The experimental results show that for the bobbin image with few components and total reflection area,it is more suitable to use Gray World method for color cast detection and White Patch method for color cast correction..(3)This thesis studies the classification algorithm of bobbin shape pattern based on transfer learning.In order to make the model focus more on texture and structure in image feature extraction,binary image is used as training sample,and a small number of training images are expanded to thousands of training sets by using rotation,scaling and clipping methods,and then it is used for the pre trained images on large data sets perception network is trained to complete migration learning.Compared with the self built shallow convolutional neural network,the test results show that the migration learning can obtain a higher classification accuracy,which reaches 92.73%.In order to solve the problem that the model has poor classification effect on the defective binary test images,this thesis uses morphological processing to enhance these images,and improves the accuracy of the test to be closer to the training results.