结合感知注意力机制的生成式对抗网络图像修复

Generative Adversarial Network Image Inpainting Combined with Perceptual Attention Mechanism

针对现有经典图像修复算法修复结果存在的语义信息不合理、修复边界处易产生伪影等问题,本文结合注意力机制对其进行改进。第一层生成模型对图像进行编码解码操作,完成粗略修复;第二层生成模型结合感知注意力,完成具有更合理语义信息的精细修复;采用局部鉴别器和全局鉴别器对修复内容进行反馈优化。与其他两种主流修复算法基于CelebA数据集进行对比,PSNR值最大程度提升了1.34 dB,SSIM值最大程度提升了0.007。实验结果表明,用结合注意力机制算法修复后图像的语义结构以及纹理的完整性与原图更加接近。

There are unreasonable semantic information and artifacts at the repair boundary in the results obtained by the classic image inpainting algorithm.It is modified in this paper by combining with the perceptual attention mechanism.The preliminary result is generated by the first-layer generative model encoding and decoding the images.By combining with the perceptual attention mechanism,the second layer generative model generates a fine result the semantic information of which is more reasonable.Finally,the repaired image is fed back and optimized with the local discriminator and the global discriminator.A comparison between the improved inpainting algorithm with the other two mainstream ones is made based on the CelebA Dataset,showing that the peak signal to noise ratio increases by a maximum of 1.34 dB and the structural similarity increases by a maximum of 0.007.It is concluded that the integrity of semantic structure and the texture of the image repaired by the combined attention mechanism algorithm is much closer to the original image.