基于小目标语义增强的机织物疵点检测方法  

Defect detection method of woven fabric based on small target semantic enhancement

作　　者：周星亚 刘可心 吴正香 夏克尔·赛塔尔[1] ZHOU Xingya;LIU Kexin;WU Zhengxiang;XIAKEER Saitaer(Xinjiang University,Urumqi,830047,China)

机构地区：[1]新疆大学,新疆乌鲁木齐830047

出　　处：《棉纺织技术》2024年第12期58-64,共7页Cotton Textile Technology

摘　　要：针对在机织物检测中,现有算法对于小目标的机织物疵点检测效果不佳,同时图像受噪声、类别不均影响导致算法学习困难、精度低的现状,提出使用一种可学习数据增强网络与YOLOv7直接相连,实现端到端训练,使输入图像能够在训练过程中得到转译,以提高图像的输入质量。设计了一种核选择注意力机制,并将其嵌入到YOLOv7的小目标检测头中,使网络更加关注小目标的特征。最后,将ELAN模块与一种基于像素的Transformer相结合,形成了ELAN⁃Transformer,并将其引入到YOLOv7的骨干网络中,克服ELAN模块在处理小疵点区域时的局限性,使网络对疵点区域的语义感知能力得到显著增强,提高了对疵点目标的准确性和鲁棒性。试验结果表明:在包含微小机织物疵点数据集上测试,该研究算法能更好地检出小目标机织物疵点,mAP@0.5达到95.2%,精度达到95.5%,召回率相比于原YOLOv7提升了2.8个百分点,满足纺织企业对机织物微小疵点的检测需求。In woven fabric detection,there were some disadvantages,including worse detection effect on woven fabrics for small targets in the existing algorithms,difficulity in algorithm learning due to the influence of noise and uneven categories on the image.A learnable data augmentation network was proposed to be directly connected to YOLOv7 to achieve end-to-end training,so that the input image can be translated in training process to improve the input quality of the image.A kernel selection attention mechanism was designed and embedded in small target detection head of YOLOv7 to make the network pay more attention to the features of small targets.Finally,ELAN module was combined with a pixel-based Transformer to form ELAN-Transformer,which was introduced into the backbone network of YOLOv7.The limitation of ELAN module in dealing with small defect areas was overcomed.The semantic perception ability of the network to the defect area was significantly enhanced,the accuracy and robustness of the defect target were improved.The test results showed that the algorithm can better detect the small target woven fabrics defects on the dataset containing micro woven fabric defects,with a mAP@0.5 of 95.2%,an precision of 95.5%,and a recall rate of 2.8 percentage points higher compared with the original YOLOv7,which could meet the detection needs of textile enterprises for the detection of small defects in woven fabrics.

关 键 词：目标检测 机织物疵点 YOLOv7 小目标语义 注意力机制 

分 类 号：TS101.9[轻工技术与工程—纺织工程]