基于改进U-Net的工件表面缺陷分割方法

Segmentation Method for Surface Defects of Industrial Components Based on Improved U-Net

针对当前深度神经网络模型在检测小缺陷目标时性能较差的问题,提出了一种基于改进U-Net的工件表面缺陷分割方法。该方法设计了一种仅下采样3次的U型网络,在保持图像特征分辨率的同时获得足够的感受野,有效解决神经网络多次下采样造成的小目标信息丢失问题;引入Dice损失和Focal损失组成的混合损失函数,通过增强分割损失权重并抑制背景信息来提高分割效果,有效解决小缺陷目标的低概率密度问题。通过在表面缺陷数据集上的大量实验和分析,结果表明该算法能够很好地细分出缺陷区域,并在分割精度与速度之间获得平衡。

Aiming at the problem that the current deep neural networks has poor performance in detecting small surface defects,this paper proposed a segmentation method for surface defects of industrial components based on improved U-Net.In order to solve the problem of small object information loss caused by multiple down-sampling operations in common neural networks,the proposed method designed an improved U-Net adopting only three down-sampling operations,which helps to obtain sufficient receptive field while maintaining feature resolution.In addition,the proposed method introduced a hybrid loss function composed of Dice loss and Focal loss.By enhancing the segmentation loss weight and suppressing the background information,the segmentation effect was improved,and the low probability density problem of small defect target was effectively solved.Through a large number of experimental analysis on surface defect datasets,the results show that the algorithm proposed in this paper can segment the defect area well,and achieve trade-off between segmentation precision and speed.