Research And Application Of Image Processing Algorithms In Intelligent Fabric Production Line

With the fast-paced development of digital image processing technology and smart manufacturing,the fabric and clothing production industry is also actively transforming and upgrading.Image segmentation technology can divide fabric and clothing images into different regions and identify different parts of the fabric.Image retrieval technology can automatically search for fabric images that are similar to the target image.Defect detection technology can automatically recognize and locate defects in the fabric.The application of these technologies can enhance productivity and integrity.However,traditional image processing technology has problems such as low efficiency and high error rates.Therefore,researching image processing technology in intelligent fabric production lines can not only accelerate the intelligent process of production lines but also provide reference and inspiration for the intelligent upgrade of other industries.Currently,fabric and clothing-related images often have problems such as noise interference or irrelevant feature interference,unknown pattern categories and defect types,and poor real-time operation,which make accurate and efficient segmentation,retrieval,and defect detection of fabric and clothing images challenging.This thesis focuses on the difficulties and challenges encountered in the image processing tasks of the intelligent production lines for fabrics,and explores the research on relevant image processing algorithms using deep learning technology.The main content and contributions are as follows:(1)An unsupervised image segmentation algorithm based on improved U-Net and cooperative learning is proposed to address the problems of large data size and lack of annotation in image segmentation.This algorithm jointly learns pixel labels and feature representations after inputting the target image,uses the improved U-Net network to extract image features fully,obtains the initial segmentation result using Argmax classification,and obtains the final refined segmentation result using super-pixel optimization.In addition,batch normalization is used to optimize the label quantity between backbone network and Argmax.The efficacy of this algorithm is validated by experiments.(2)A deep hashing algorithm based on multi-dimensional feature fusion is put forward to address problems such as irrelevant feature interference and noise interference in image retrieval.This algorithm uses multi-scale and multi-direction pooling to achieve multi-dimensional feature fusion and uses residual pyramid pooling after the backbone network to further preserve multi-scale features,mitigating interference from noise and extraneous features.The accuracy and efficacy of this method are verified through experiments.(3)A defect detection algorithm based on data augmentation and normalizing flow is proposed to address problems such as existing unknown defect types and low defect occurrence frequency in defect detection.The normalizing flow model with self-attention mechanism can achieve accurate probability density estimation of the latent space of the image.And the likelihood calculated from data augmentation can calculate a more robust defect score.In addition,this model can be directly applied to defect localization through gradient backpropagation.The algorithm was finally experimentally validated to be able to accurately determine and locate defects in fabric clothing images.Building upon the aforementioned research content,this thesis develops a system that utilizes image processing algorithms for intelligent fabric production lines and applies the intelligent algorithms designed in fabric and clothing image processing tasks.System tests show that the system is stable and can improve the efficiency and accuracy of image processing tasks.