基于生成对抗网络的木地板异常检测方法

Wooden floor anomaly detection method by Generative Adversarial Networks

在木板异常检测的过程中,由于木地板颜色跨度大,纹理背景与缺陷难以区分,同时存在缺陷占比非常小、缺陷的不可预见的问题,给基于图像的木板异常检测带来极大困难。文章提出一种新的基于记忆增强的生成对抗网络的无监督异常检测方法,仅使用大量无缺陷样本就可对其缺陷样本进行异常。该方法利用U-Net的跳跃连接结构,提高模型的重构效果;其次,提出一种注意力机制在判别器部分,生成结构性更复杂和细节更准确的图像;同时,加入记忆增强模块,使模型能够记忆正常数据的特征,从而使得异常数据的重构误差变大,判别性能得到增强。实验结果表明,本文提出的方法,与经典的GANomaly、skip-GANoamly、Anogan一些无监督检测模型相比,所提方法具有更高的AUC和F1评分,AUC可达0.962。

In the process of abnormal detection of wooden boards,due to the large color span of wooden floors,it is difficult to distinguish between texture backgrounds and defects,and there are also problems such as a very small proportion of defects and unforeseeable defects,which brings grcat difficulties to image-based abnormal detection of wooden boards.This article proposes a new unsupervised anomaly detection method based on memory enhanced generative adversarial networks,which can detect abnormal defects using only a large number of defect free samples.This method utilizes the skip connection structure of U-Net to improve the reconstruction effect of the model;Secondly,an attention mechanism is proposed in the discriminator section to gencerate images with more complex structures and more accurate details;At the same time,adding a memory enhancement module enables the model to remember the features of normal data,thereby increasing the reconstruction crror of abnormal data and enhancing the discrimination performance.The experimental results show 0.962that the proposed method has higher accuracy and Fl score compared to classic unsupervised detection models such as GANomaly,skip GANoamly,and Anogan.