混合Polarity采样的多尺度GAN图像修复算法  

Multi-scale GAN image inpainting algorithm with mixed polarity sampling

作　　者：陈刚 张丽英 姚剑 刘智勇[3] CHEN Gang;ZHANG Liying;YAO Jian;LIU Zhiyong(School of Artificial Intelligence,Guangdong Open University,Guangzhou 510091,China;School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430070,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]广东开放大学人工智能学院,广东广州510091 [2]武汉大学遥感信息工程学院,湖北武汉430070 [3]中国科学院自动化所,北京100190

出　　处：《微电子学与计算机》2025年第3期30-39,共10页Microelectronics & Computer

基　　金：国家自然科学基金(42271445);广东省软科学研究计划项目(2024A1010010001);广东开放大学重点科研团队项目(Kytd2302)。

摘　　要：针对图像修复任务中存在的纹理、局部细节丢失和全局结构一致性差等问题,传统的多尺度GAN图像修复算法无法很好地解决这些问题,提出了基于混合Polarity采样的多尺度GAN图像修复算法。首先,将可分离自注意力机制引入到网络的深层,能够捕捉像素之间的长距离依赖关系,提高了重建图像上下文感知性与全局结构性。其次,设计了具有混合Polarity采样能力的编码器。该编码器能够克服样本不平衡问题,以致能增强算法对缺陷区域的建模能力,从而更好地保持原始图像的局部细节,提高修复图像的质量。此外,在解码器中设计了自适应实例化层,增强算法对原图风格及轮廓的保留性能,进一步提升修复图像的感知质量。最后,将算法应用于公共数据集CelebA。实验表明,所提算法在峰值信噪比、结构相似性以及距离得分等指标上的表现均优于基线。In response to issues such as texture loss,local detail loss,and poor global structural consistency in image restoration tasks,traditional multi-scale GAN image restoration algorithms cannot effectively solve these problems.A multiscale GAN image restoration algorithm based on mixed Polarity sampling is proposed.Firstly,the separable self-attention mechanism is introduced into the deep layers of the network to capture long-distance dependencies between pixels,thereby enhancing context awareness and improving the global structure of reconstructed images.Secondly,an encoder with mixed polarity sampling capability is designed to address sample imbalance and improve the algorithm's modeling ability for defect areas,resulting in better preservation of local details in the original image and enhanced quality of the repaired image.Additionally,an adaptive instantiation layer is incorporated in the decoder to enhance the algorithm's performance in preserving the original image style and contour,further improving the perceptual quality of the repaired image.Finally,the algorithm is applied to the public dataset CelebA,and experimental results demonstrate its superior performance over the baseline in metrics such as peak signal-to-noise ratio,structural similarity,and distance score.

关 键 词：可分离自注意力 Polarity采样 感知质量 多尺度 图像修复 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]