使用卷积自动编码器检测印刷电路板缺陷

Using Convolutional Autoencoders to Detect Defects in Printed Circuit Board

摘要: 在制造印刷电路板(PCB)过程中，由于诸如设备、环境、原料和人工操作等无法控制的影响，从而导致在生产各个环节中都有可能使产品存在缺陷，因此，检测并定位印刷电路板中的所有缺陷至关重要。传统人工目检存在检测效率低与检测缺陷不全等问题，为有效解决上述问题，提出了一种基于无监督学习的卷积自动编码器模型对印刷电路板的质量进行检测，并结合生成对抗网络将生成器作为卷积自动编码器模型的解码器。该网络模型仅通过无缺陷产品图像进行训练并学习无缺陷产品的特征，通过将缺陷图像重构为无缺陷图像，再与缺陷图像相减，获得包含缺陷信息的残差图，定位出缺陷位置。实验结果表明：该方法能够很好地识别印刷电路板缺陷，准确率达到96.15%，且具有较好的泛化能力和鲁棒性。

Abstract: In the process of manufacturing a printed circuit board (PCB), due to uncontrollable influences such as equipment, environment, raw materials and manual operations, which can cause defects in the product at all stages of production, it is essential to detect and locate all defects in the printed circuit board. In order to effectively solve the problems of low detection efficiency and incomplete detection defects in traditional manual eye inspection, a convolutional autoencoder model based on unsupervised learning was proposed to detect the quality of printed circuit boards, and the generator was used as the decoder of the convolutional autoencoder model combined with generative adversative network. The network model is trained and learns the features of defect-free products only through the image of defect-free products. By reconstructing the defect image into a defect-free image and subtracting the defect image, the residual map containing the defect information is obtained and the defect location is located. Experimental results show that the proposed method can identify PCB defects with an accuracy of 96.15%, and has good generalization ability and robustness.