基于改进FCOS的钢带表面缺陷检测算法

Algorithm on Surface Defect Detection of Strip Based on Improved FCOS

摘要: 针对现有钢带表面缺陷检测所存在的检测效率低、适用范围有限等缺陷，提出一种基于改进FCOS的钢带表面缺陷检测算法。该算法使用含形变卷积的卷积神经网络提取缺陷特征，使用关键点特征融合增强检测模型输入，并使用中心采样策略选取训练样本优化模型训练，最后使用东北大学钢带表面缺陷公共数据集进行训练和评估。本文算法在东北大学钢带表面缺陷公共数据集上平均检测精度为74%，检测速度为31.4 FPS。

Abstract: Aiming at the defects of low detection efficiency and limited applicable scope in strip surface defect detection, a steel strip surface defect detection algorithm based on improved FCOS was proposed. The algorithm uses convolutional neural network with deformable convolution to extract defect features, uses key point feature fusion to enhance the detection model input and uses the central point sampling strategy to select training samples to optimize model training. Finally, the proposed algorithm is trained and evaluated on public dataset NEU-DET, Northeastern University surface defect database. On NEU-DET dataset, the mean average precision of this algorithm achieves 74% and the detection velocity is 31.4 FPS.