基于改进型U-Net卷积神经网络的磁片表面缺陷检测

Defect Detection of Magnetic Disk Based on Improved U-Net Convolutional Neural Network

为了对磁片质量检测过程中常见的缺角、划痕、脏污缺陷准确地分割,利用U-Net卷积神经网络的编码解码功能,提出了一种改进的U-Net网络的磁片缺陷图像分割算法。该方法采用深度可分离卷积来减小计算量与模型参数量,结合注意力机制sSE block提炼图像特征图,提高模型的准确率。实验结果表明,所提出算法在磁片缺陷检测中网络的输出图像失真更小,针对缺角、划痕、脏污缺陷检测取得了良好的表现,网络检测结果的准确率(AC)分别达到98.23%、97.25%、96.57%,与原始网络相比提高了1.1%~2.86%,平均交互比(MIoU)分别达到了84.72%、77.36%、75.81%,提高了1.5%~3%,图像分割的效果良好。将改进后的网络在车间现场进行测试,误报率小于5%,漏报率为0。

To accurately segment the common defects such as unfilled corner, scratch and dirty in the process of magnetic disc quality detection, an improved image segmentation algorithm based on the encoding and decoding function of U-NET convolutional neural network is proposed in this paper. It is pointed out that this method uses deep separable convolution to reduce the amount of computation and model parameters and combines with the attention mechanism sSE block to extract the image feature map to improve the accuracy of the model. Experimental results show that the proposed algorithm has smaller output image distortion of network in the disk defect detection, and has achieved good performance for lack of Angle, scratches and dirt defect detection. The network accuracy(AC) of online test result is 98.23%, 97.25% and 96.57% respectively, which has increased by 1.1% ~ 2.86% compared with the original network, and the average interaction than(MIoU) reaches 84.72%, 77.36% and 75.81% respectively, increasing by 1.5% ~ 3% with good image segmentation effect. Using the improved network to conduct test in the workshop site, the false alarm rate is less than 5% and the missing alarm rate is 0.