Research On Surface Defect Recognition Method Of Chemical Fiber Yarn Packages Based On Deep Learning

Chemical fiber filament is an important textile raw material.In order to facilitate storage and transportation,it is usually wound around a paper tube to form a chemical fiber yarn package.Due to some factors in the production process and the shaking of paper tube during the winding process,there will be many defects on the surface of the yarn package.The existence of these defects will affect the level and price of yarn package,so,it is necessary to inspect these defects before the yarn packages sell.Manual detection is by far the most common method,however,this method is not highly reliable and prone to cause the problem of missed detection.In view of this problem,we used deep learning methods to identify the common defects of yarn packages.The main work completed and the results obtained are as follows:(1)According to the actual task requirements,the relevant parameters of the hardware required to collect the photos were calculated and the model was selected,then the camera and the matching lens were chosen,next,the light source and lighting way were determined.At lase,the yarn package photos of clearness and high-contrast were collected,and to some extent they reduced the preprocessing process of the subsequent algorithm recognition part.(2)The photos of tripping filament,bad formation,stained and normal yarn package were collected.Firstly,they were processed into blocks to highlight the small defects.Then data augmentation methods such as shifting and random rotation were used to further increase the number of sample and ensure the balance of different kinds.A filter way was adopted to reduce image noise.At lase,a total of 12465 training samples and 512 test samples of yarn packages were obtained.The training and testing set were made into a file of the format of tfrecord respectively which was suitable for the input of the network model.(3)Considering the task of defect identification of yarn packages and the number of samples,the Alex Net network was chosen as the basic network model and some improvements were made.The lager convolution size of the first layers was replaced with smaller 3×3 to extract more distinguishing features.On top of the feature maps produced by the last convolutional layer,a global maximum pooling way was proposed to replace the traditional full collection,greatly reducing the number of parameters and enhancing the robustness of the network to the image space transformation.The value of some key hyperparameters were analyzed and chosen and the final model was determined.The experimental results show that our improved network achieve a recognition accuracy of 97.1% on the testing set of yarn package and the recognition effect is satisfactory.(4)Aiming at the high cost caused by labeling a large number of samples,from the perspective of how to make full use of fewer labeled samples to quickly improve network performance,an active learning method was proposed.A comprehensive selection criterion which considered the information and diversity of samples was designed.The network model calculated and actively selected unlabeled samples based on the this criterion and then allowed experts to label them accurately.The initial annotation samples for active learning and the conditions to stop were set.Finally,the active learning experiments were performed on the previous improved model.The experimental results show that the method in this paper uses only a few samples of chemical fiber yarn packages to train a network model with the same performance,reducing the labeling cost.