Detection Of Cotton And Hemp Blend Yarn Ratio Based On Deep Learning Technology

Cotton and hemp blended yarn combines the excellent characteristics of cotton fiber and hemp fiber,which makes the fabric feel comfortable and has good moisture absorption and air permeability,so it is loved by the majority of consumers.The determination of the blending ratio of cotton and hemp is a very important indicator.Because both cotton and hemp fibers belong to cellulose fibers,it is impossible to use chemical dissolution method,combustion method and other conventional methods for quantitative analysis.At present,the efficiency of microscopic observation is low,which is time-consuming and laborious.In recent years,with the rapid development of deep learning,it can automatically extract the features of detected objects to identify and locate,and the accuracy is often higher than the traditional image processing technology,which has been applied to various fields.Therefore,this paper explores the feasibility of its application in the detection of cotton and hemp fiber blending ratio based on the deep learning technology.The research work of this paper is as follows:(1)In this paper,the target detection method based on deep learning technology was proposed to detect the blending ratio of cotton and hemp fiber.Through the sampling of cotton and hemp to obtain images,target labeling,data set division,training and so on,the influence of fiber slice length,fiber overlapping,fiber state,fiber sample size and the presence of small fibers on model detection was discussed,and the optimal cotton and hemp fiber detection model was determined.The results show that the proposed method is effective.(2)The detection model of cotton and hemp fiber based on target detection was constructed.Cotton and hemp data set format as VOC2007,In the preparation of the data set,the acquisition of cotton and hemp fiber pictures and the labeling of fiber targets were unified,and the data set wasdivided into training set and test set according to 8:2.In the establishment and training of the model,by comparing the detection accuracy of SSD network and Faster Rcnn network,this paper constructed a detection model similar to Faster Rcnn with resnet50 as the backbone network.On this basis,the influence of different confidence threshold and training epoch number on the detection model was discussed.Model test showed that the recall rate of cotton fiber is 74.6%,the precision rate of cotton is 79.3%,the recall rate of hemp fiber is 76.7%,the precision rate of hemp is 80.0%,and the overall accuracy of the model is low.(3)The detection model of cotton and hemp fiber based on target detection was optimized.The effects of fiber section length,fiber lap,fiber state,fiber data and the presence of small fibers on the detection model were discussed.Comparing the detection results of 0.24 mm,0.30 mm and0.36 mm fiber slice length,it is found that when the fiber slice length is 0.30 mm,the detection accuracy of the model is the best.The precision of the model for cotton fiber is 85.7%,the recall rate is 86.9%,the precision of hemp fiber is 84.2%,the recall rate is 90.9%,and the mean average precision of the model is 85.9%.By analyzing the influence of fiber overlap on model training and prediction,the reason of low detection accuracy of model based on 0.36 mm fiber slice length is explained.Soft NMS in test code algorithm was used to improve the problem of missed detection due to the high degree of fiber overlap,which makes the recall rate of cotton and hemp test model increase by 3%,to 89.9%.In view of the shortage of fiber samples,the images on the training set were enlarged by means of rotation,flipping and mirror image processing.After data augmentation,The precision of cotton fiber reach 90.7%,the recall rate is 91.0%,the precision rate of hemp fiber is 91.4%,the recall rate is 90.7%,and the mean average precision of the model was 89.2%.In addition,according to the principle of multiscale detection,the feature pyramid structure was introduced to reduce the false inspection of fine fibers.Finally,with the optimized model,the precision of the model for cotton fiber is 91.5%,the recall rate is 94.3%,the precision of hemp fiber is 92.4%,the recall rate is 95.2%,and the mean average precision of the model is91.0%.(4)Test and verification of blending ratio of cotton and hemp blended yarn.In this paper,the optimized model was used to test the blending ratio of 60 / 40 and 50 / 50 cotton and hemp blended yarn.Using cashmere to cover cotton and hemp blended yarn to make slices and obtain pictures of cotton and hemp fibers;Then the optimized model is used to detect the fiber target in the picture,and the number of all kinds of fibers is obtained;Finally,the diameter of two kinds of dimension was measured by fiber fineness meter,and the blending ratio of cotton and hemp fiber was obtained by two numerical conversions.The experimental results show that the number of cotton and hemp fibers measured by the model is basically the same as that calculated by human,and the relative error of the blending ratio converted by the two is 0.02,which proves theeffectiveness of the proposed method.The detection model based on target detection for cotton and hemp fiber proposed in this paper is feasible for the measurement of cotton and hemp blended yarn ratio,which provides technical support for the industrialization of related testing methods,and also provides reference for the research of similar fiber detection.