基于深度学习的型钢表面多形态微小缺陷检测算法

为了解决由于型钢表面缺陷形态多样、微小缺陷众多所带来的检测效率低与检测精度差的问题，提出一种基于可变形卷积与多尺度-密集特征金字塔的型钢表面缺陷检测算法——Steel-YOLOv3。首先，使用可变形卷积代替Darknet53网络部分残差单元的卷积层，从而强化特征提取网络对型钢表面多类型缺陷的特征学习能力；其次，设计了多尺度-密集特征金字塔模块：在原有YOLOv3算法的3层预测尺度上增加1层更浅层的预测尺度，再对多尺度特征图进行跨层密集连接，从而增强对密集微小缺陷的表征能力；最后，针对型钢缺陷尺寸分布特点，使用K-means维度聚类方法优化先验框尺寸并将先验框平均分配到4个对应预测尺度上。实验结果表明：Steel-YOLOv3算法具有89.24%的检测平均精度均值（mAP），与Faster R-CNN（Faster Region-based Convolutional Neural Network）、SSD（Single Shot MultiBox Detector）、YOLOv3和YOLOv5算法相比分别提高了3.51%、26.46%、12.63%和5.71%，且所提算法显著提升了微小剥落缺陷的检出率。另外，所提算法的每秒检测图像数量达到25.62张，满足实时检测的要求，可实际应用于型钢表面缺陷的在线检测。

Surface detection algorithm of multi-shape small defects for section steel based on deep learning

In order to solve the problems of low detection efficiency and poor detection precision caused by various surface defects and numerous small defects of section steel， a detection algorithm for surface defects of section steel， namely Steel-YOLOv3， was proposed on the basis of the deformable convolution and multi-scale dense feature pyramid. Firstly， the deformable convolution was used to replace the convolutional layers of part of the residual units in Darknet53 network， which strengthened the feature learning ability of feature extraction network for multi-type defects on the surface of section steel. Secondly， a multi-scale dense feature pyramid module was designed， which means that a shallower prediction scale was added to the 3 prediction scales of the original YOLOv3 algorithm and the multi-scale feature maps were connected across layers， thereby enhancing the ability to characterize dense small defects. Finally， according to the defect size distribution characteristics of section steel， the K-means dimension clustering method was used to optimize the scales of anchor boxes， and the anchor boxes were evenly distributed to 4 corresponding prediction scales. Experimental results show that Steel-YOLOv3 algorithm has a detection mean Average Precision （mAP） of 89.24%， which is improved by 3.51%， 26.46%， 12.63% and 5.71% compared with those of Faster Region-based Convolutional Neural Network （Faster R-CNN）， Single Shot multibox Detector （SSD）， YOLOv3 and YOLOv5 algorithms respectively. And the detection rate of tiny spalling defects is significantly improved by the proposed algorithm. Moreover， the proposed algorithm can detect 25.62 images per second， which means the requirement of real-time detection can be met and the algorithm can be applied to the online detection for the surface defects of section steel.