改进YOLOv3的轧钢表面缺陷检测方法  

Surface Defect Detection of Rolled Steel on Improved YOLOv3

作　　者：魏鑫 谈英姿[1] 罗晨 张刚 

机构地区：[1]东南大学自动化学院,江苏南京210000

出　　处：《工业控制计算机》2023年第8期85-87,共3页Industrial Control Computer

基　　金：国家自然科学基金(51975119)。

摘　　要：针对轧刚表面缺陷种类多样、形状多变导致检测效率低、精度差的问题,提出了一种改进YOLOv3的轧钢表面缺陷检测算法。首先,对骨干网络提取的特征采用PSA金字塔拆分注意力模块进行多尺度融合。其次,采用PAN结构代替FPN,使得浅层语义和深层语义的特征能充分融合。接着采用Decoupled_Head,将回归预测和逻辑预测分离以避免之间的干扰。最后,在损失函数方面,根据真实框大小赋予不同权值,以提高网络对小目标的检测效果。实验表明改进后的YOLOv3在NEU-DEU数据集上的平均检测精度为80.01%,比原始的YOLOv3提高了3.05%,且相较于YOLOx、YOLOv5等算法也有较大的检测精度优势。Aiming at the problems of low detection efficiency and poor accuracy due to the various types and shapes of rolled surface defects,an improved surface defect detection of rolled steel algorithm based on YOLOv3 is proposed in this paper.Firstly,the features extracted by the backbone network are multi-scale fused by the PSA module.Secondly,PAN structure is used to replace FPN,so that the features of shallow semantics and deep semantics can be fully integrated.The prediction part is changed from Head to Decoupled Head in order to separate regression prediction and logistic prediction to avoid the mutual interference.Finally,in terms of the loss function,different weights are given according to the size of the real frame to improve the detection effect of the network on small targets.Experiments show that the average detection accuracy of the improved YOLOv3 on the NEU-DEU dataset is 80.01%,which is 3.05%higher than the original YOLOv3,and compared to YOLOx,YOLOv5 and other algorithms also have greater detection accuracy advantages.

关 键 词：缺陷检测 YOLOv3模型 金字塔差分注意力 多尺度融合 检测头解耦 

分 类 号：TG115[金属学及工艺—物理冶金] TP183[金属学及工艺—金属学] TP391.41[自动化与计算机技术—控制理论与控制工程]