基于MD-CBAM的多样性裂缝图像修复方法

Diversity crack image inpainting method based on mask distance convolutional block attention module

摘要:

大多数现有的桥梁裂缝图像修复方法为单一目标修复,无法根据孔洞周边的有效信息生成多种合理的填充内容且修复结果存在结构扭曲和纹理模糊的问题。本文提出了一种基于掩膜距离卷积块注意力模块(mask distance convolutional block attention module,MD-CBAM)的多样性裂缝图像修复网络,该方法主要由多样性结构生成器与纹理生成器组成。提出区域结构注意力以降低遮挡区域像素与有效像素的差异性,根据掩膜特征对注意力分数进行平均池化处理,提高模型对遮挡区域的推断能力。设计MD-CBAM模块用以在纹理生成阶段合成高质量的特征,该模块利用特征之间的距离信息与语义信息,有效增强了模型填充大孔洞的能力。实验结果表明,本文方法修复的图像具有更为明确的结构和更加合理的纹理,在各掩膜比例下峰值信噪比(peak signal-to-noise ratio,PSNR)和FID(Fréchet inception distance)均达到最优,其中PSNR在掩膜比例为[0.4,0.5)时增加了0.22—2.38 dB且结构相似度(structural similarity,SSIM)值达到最优。

Abstract:

Most of the existing bridge crack image inpainting methods are single target restoration,which cannot generate multiple reasonable filling contents based on valid information around the hole. Moreover,the inpainting results suffer from structural distortion and texture blurring.A diversity crack image inpainting network based on the mask distance convolutional block attention module (MD-CBAM) is proposed in this paper,which mainly consists of a diversity structure generator and a texture generator.The regional structure attention is proposed to reduce the difference between the pixels in the masked region and the valid pixels,and the average pooling is performed on the attention scores according to the mask features to improve the inference ability of the model to the masked area.The MD-CBAM module is designed to synthesize high-quality features in the texture generation stage.The module utilizes the distance information between features and semantic information to effectively enhance the capability of the model to fill large holes.The experimental results show that the inpainted image has a more definite structure and a more reasonable texture,and the peak signal-to-noise ratio (PSNR) and Fréchet inception distance (FID) reach the best at each mask ratio, where the PSNR increases by 0.22—2.38 dB at the mask ratio of [0.4,0.5) and the structural similarity (SSIM) value is optimal.