基于超分辨率模型与YOLO-V4的织物疵点检测

Fabric Defect Detection Based on Super Resolution and YOLO-V4

针对工业条件限制下采集的印花布数据集图像分辨率低、检测效果差等问题,课题组提出基于超分辨率模型SRGAN与YOLO-V4网络的织物疵点检测方法,并对SRGAN算法进行改进。课题组首先使用改进的SRGAN算法对原数据集进行超分辨率重构,提高图像分辨率;然后将重构图翻转变化与原图共同作为数据集输入YOLO-V4进行网络训练;最后通过YOLO-V4网络检测印花布表面疵点。实验结果表明:该方法可提高低分辨率织物图疵点检测效果,准确率高达90.29%,比超分辨率重构前提升了13.19%,能实现实时定位疵点的准确位置并输出疵点类别。

Aiming at the problems of low image resolution and poor detection effect of the collected printed fabric data set under industrial conditions, a fabric defect detection method based on super-resolution model(SRGAN) and YOLO-V4 networks was proposed, and the SRGAN algorithm was improved. Firstly, the improved SRGAN algorithm was used to reconstruct the original data set with super resolution to improve the image resolution;secondly, the inverted changes of the reconstructed image and the original image were input into the YOLO-V4 network as the dataset for training;finally, the surface defects of printed fabric were detected by YOLO-V4 network. The experimental results show that the proposed method can improve the effect of low resolution fabric image defect detection, and the accuracy rate is as high as 90.29%, which is 13.19% higher than that before super-resolution reconstruction. Moreover, it can realize the real-time positioning of defects and output the defect category.