关键词: 图像修复, 上下文信息多样聚合, 编解码信息融合, 掩码匹配鉴别器

Abstract: To effectively solve the defects of structural distortion and blurry texture when repairing large and irregular semantic missing area images by using the existing algorithms, a diverse aggregation image restoration algorithm based on contextual information is proposed. First,the information on the damaged image is extracted by the encoder to estimate the missing content. Thereafter, the context information from various receptive fields is merged through the multi-information aggregation block to enhance the structure and texture information of the missing area. Then, the original image features are restored through the decoder. Finally, the mask matching discriminator is adopted to perform discrimination training on the generated image,and the model is optimized by combining the counter loss, reconstruction loss,perception loss and style loss to promote the generator to synthesize clear textures. The proposed algorithm is trained and tested on the public data set. The experimental results show that the proposed algorithm can obtain clearer and more reasonable structure and texture details than state-of-the-art when inpainting randomly irregular and large missing areas. Its objective indices such as peak signal-to-noise ratio and structural similarity are superior over the compared algorithms.

Key words: image inpainting , diverse aggregation of contextual information , encoding and decoding information fusion , mask matching discriminator