摘要
对布匹瑕疵图像中存在的复杂花色背景、小目标、尺度变化大、数量不均衡等问题进行研究,提出双路高分辨率转换网络的花色布匹瑕疵检测算法。利用双路高分辨率特征提取,消除复杂花色背景的影响;采用多尺度特征金字塔转换器,提高小瑕疵目标检测的准确率;设计自适应边界框生成器,指导初始锚定框设计;采用改进的聚焦损失解决少数类瑕疵样本准确率不高的问题。实验结果表明,该算法能有效提升布匹瑕疵检测的准确率及定位精度,优于当前主流的布匹瑕疵检测算法。
The problems of complex background,small object,large scale change,and unbalanced quantity in fabric defect images were studied.A two-way high-resolution transformer network based fabric defect detection algorithm was proposed.Two-way high-resolution feature extraction was used to eliminate the influence of complex backgrounds.Multi-scale feature pyramid transformer was used to improve the accuracy of small defect object detection.An adaptive bounding box generator was designed to guide the design of the initial anchor box.The focal loss was used to solve the problem of low accuracy of minority class ble-mish samples.Experimental results show that the proposed algorithm can effectively improve the accuracy and positioning accuracy of fabric defect detection,which is better than the current mainstream fabric defect detection algorithm.
作者
李辉
吕祥聪
申贝贝
陶冶
王俊印
LI Hui;LYU Xiang-cong;SHEN Bei-bei;TAO Ye;WANG Jun-yin(School of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China)
出处
《计算机工程与设计》
北大核心
2023年第9期2731-2739,共9页
Computer Engineering and Design
基金
智能感知与自主控制教育部工程研究中心开放基金项目(K100052021006)
国家自然科学基金项目(61702295)
山东省高等学校优秀青年创新团队计划基金项目(2019KJN047)。
关键词
布匹瑕疵
小目标
尺度变化大
不平衡分类
双路高分辨率
转换网络
目标检测
fabric defect
small object
large scale change
imbalance classification
dual high resolution
transformer network
object detection
