| 摘 要: 为了简化查阅焊缝DCM(DICOM)图像时调整窗位窗宽的步骤,提出了Adaptive-MINMAX(AdaptiveMin-Max)算法。该算法处理的焊缝图像有较高的可读性,并基于该方法建立了一个焊缝缺陷数据集。随后,对YOLOv8s(You Only Look Once Version8-small)进行改进,提出了GU-YOLO(Global Upper Yolo)模型。通过更换主干网络的卷积模块为UpperConv,在其输出中增加了二次卷积结果,增强了网络提取小目标特征的能力;设计了GAMDetect(Global Attention Mechanism Detect)检测头部,在Detect模块中添加GAM注意力机制,以突出目标并抑制背景信息。实验结果显示,GU-YOLO在自建的缺陷数据集上获得的AP、AR、F1_score、mAP50指标分别提升了1.3百分点、13.9百分点、7.9百分点、15.3百分点,与其他相关模型相比,表现出较强的竞争力。 |
| 关键词: 机器视觉;DICOM;YOLOv8s;Adaptive-MINMAX;焊缝缺陷检测 |
|
中图分类号: TP391.41
文献标识码: A
|
|
| A Weld Seam Defect Detection Method Based on Adaptive-MINMAX and GU-YOLO |
|
HAO Hongkun, HU Haiyang, PENG Bo, ZHAO Minsheng
|
(School of Computer Science and Technology, Hangzhou Dianzi University, Hangzhou 310018, China)
haohongkun@hdu.edu.cn; huhaiyang@hdu.edu.cn; pengbo1@hdu.edu.cn; zhaominsheng@hdu.edu.cn
|
| Abstract: To simplify the steps of adjusting the window level and width when reviewing weld seam DCM(Digital Imaging and Communications in Medicine) images, the Adaptive-MINMAX (Adaptive Min-Max) algorithm is proposed. This algorithm enhances the readability of processed weld images, and a weld defect dataset is established based on this method. Subsequently, improvements are made to YOLOv8s (You Only Look Once Version 8-small), and the GU-YOLO (Global Upper YOLO) model is proposed. By replacing the convolutional modules in the backbone network with UpperConv and adding secondary convolution outputs to its results, the network's ability to extract features from small targets is enhanced. Additionally, the GAMDetect (Global Attention Mechanism Detect) detection head is designed by integrating the GAM attention mechanism into the Detect module to highlight targets and suppress background information. Experimental results show that GU-YOLO achieves improvements of 1.3 percentage points,13.9 percentage points, 7.9 percentage points, and 15.3 percentage points in AP, AR, F1_score, and mAP50 metrics respectively on the custom defect dataset, demonstrating strong competitiveness compared to other related models. |
| Keywords: machine vision; DICOM; YOLOv8s; Adaptive-MINMAX; weld defect detection |
