Research On Fabric Defect Detection Algorithm Based On Deep Learning

Fabric defect detection is a key stage in textile post-processing for end-product quality control.Inspired by the successful application of deep convolutional neural networks in industry,researchers have tried to use optimization methods based on deep learning techniques to improve the detection performance of the models.However,most of these existing deep learning-based fabric defect detection methods fail to achieve a good trade-off between the four aspects of detection efficiency,accuracy,generalizability,and lightweight deployment.This paper conducts research on fabric defect detection algorithms based on object detection models in the field of deep learning,with the aim of exploring a lightweight and efficient method for fabric defect detection.The following research is carried out in this paper:(1)For the problem that there are large differences in object characteristic attributes between natural scene images and fabric images,this paper proposes a pre-processing method for fabric datasets.Firstly,the visualization of the defect distribution(aspect distribution and aspect ratio distribution of the defect bounding box)is used to analyze and summarize two technical difficulties in this vision task due to the influence of the defect object itself.Then this paper focuses on the problem of inappropriate ratio of the model’s predefined anchors,and uses K-means clustering algorithm to achieve the rectification of the model’s anchors parameters.The experimental results show that the pre-processing method can improve the detection performance of the model to a certain extent.(2)To address the four aspects of detection efficiency,accuracy,generalizability,and lightweight deployment that most current detection methods cannot balance,this paper proposes an advanced fabric defect detection model,which is called Hook Net.It takes Efficient Det as the baseline model and contains both the feature aggregation module PDAM and the biased feature pyramid structure Hook FPN.The PDAM establishes cross-channel interactions in a local and global manner to capture more contextual information.The Hook FPN focuses on low-level features,reducing the model complexity.In addition,Alpha-GIo U loss is used to further improve the localization accuracy of the bounding box.Extensive experimental results show that the detection performance of the proposed model has exceeded current open-source mainstream object detection models,and has the state-of-the-art(SOTA)effect.(3)Considering the necessity of deploying the network model on the end-side devices in defect detection task,the backbone of Hook Net proposed in(2)is designed to be lightweight and the final model deployment is verified based on the research content.Firstly,the MBConv structure unit reconfiguration is performed for the backbone of Hook Net,and the effectiveness of the designed lightweight structure unit GE-MBConv is demonstrated through experiments.The final model,GE-Hook Net,is then deployed in an NVIDIA Jetson Nano embedded development board,using Tensor RT for accelerated inference.The experimental results show that the GE-Hook Net model proposed in this paper can be better applied to end-side devices and basically meets the requirements of real-time defect detection tasks under low-speed transmission.