基于FAMGAN的轮胎X光图像缺陷检测

针对气泡缺陷特征和图像背景像素差异较小、检测困难的问题，以Skip-GANomaly为基础框架，提出了融合注意力机制生成对抗网络(FAMGAN),首先，生成器中编码器和解码器之间的跳连层由注意力特征融合模块(AFF)和注意力机制模块(CBAM)构成，提高了对目标特征的关注、减少了图像特征丢失；然后，在判别器中加入联合上采样模块(JPU),提高了模型检测图像缺陷的速度。最后，将本文提出的FAMGAN网络与近几年经典的生成对抗网络在自制的轮胎缺陷数据集上进行训练、测试和评估。实验结果表明，本文提出的网络对轮胎气泡缺陷检测的精度达到0.837,相比于Skip-GANomaly网络提高了近30%。

Defect detection of tire X-ray images based on FAMGAN

In response to the problem of small differences in blister defect features and background pixels in tire defect images, as well as difficulty in detection, Skip-GANomaly is adopted as the basic framework to propose the fusion attention mechanism generative adversarial network (FAMGAN). Firstly, the skip layer between the encoder and decoder in the generator consists of an attention feature fusion (AFF) module and a convolutional block attention module (CBAM) module, which improves the focus on target features and reduces image feature loss. Then, a joint pyramid upsampling (JPU) module was added to the discriminator to improve the speed of the model in detecting image defects. Finally, the FAMGAN network proposed in this article will be trained, tested, and evaluated on a self-made tire defect dataset with classic generative adversarial networks in recent years. The experimental results show that the proposed network achieves an accuracy of 0.837 for tire blister defect detection, which is nearly 30 percentage points higher than the Skip GANomaly network.