基于修复对抗网络的烟包表面缺陷检测方法

Surface defect detection method of cigarette packets based on repair adversarial network

针对烟包表面缺陷检测方法存在识别准确率低、复用性差和检测条件高等问题,提出一种基于图像修复对抗网络的烟包表面缺陷检测方法。利用UNET网络改进上下文编码器构成修复网络,分别用扩张率等于2和3的空洞卷积提取不同粒度的特征信息,通过SK注意力机制对其赋予权重后融入到残差网络中构成判别网络,根据待测图像与修复图像的差值过滤掉缺陷特征比较明显的烟包,对差阈值附近的烟包用判别网络进行二次检测。在3000张烟包图像数据集上对比试验,结果显示,检测速度较快,准确率最高为97.5%,可以有效滤除表面缺陷的烟包。研究对包装盒表面缺陷的检测具有借鉴意义。

For the problems of low recognition accuracy,poor reusability and high detection conditions in the current detection methods of cigarette packet surface defects,an image repair adversarial network based detection method of cigarette packet surface defects was proposed.The UNET network was used to improve the context encoder to construct the repair network,and the hollow convolution with expansion rate equal to 2 and 3 were used to extract the feature information of different granularity,which was weighted by SK attention mechanism and integrated into the residual network to form a discrimination network,and the cigarette packets with obvious defect features were filtered out according to the difference between the image to be tested and the repaired image.The cigarette packet discrimination network near the difference threshold was detected twice.Compared with image data sets of 3000 cigarette packets,the results show that the method has a fast detection speed,the highest accuracy is 97.5%,and can effectively filter the cigarette packets with surface defects.The research has reference significance for the detection of surface defects of packaging boxes.