Abstract: Image inpainting is one technique of reconstruction defect areas by inferring information from the known context of defect images.For semantic image inpainting of large areas, there are still many problems in inpainting algorithms based on generation models, such as blur, artifacts, and poor visual similarity, especially for the complex background images and small datasets.To solve this problem, an image inpainting algorithm based on dense convolution generative adversarial networks is proposed.The generated adversarial network is the basic framework.Firstly, dense convolutional blocks are used to enhance image feature extraction, improve image repair capability, and avoid the problem of gradient disappearance caused by the network depth increa-sing in the generator network.Secondly, skip connection between the encoding and decoding structures is involved to avoid information transmission lost problems between network layers.After that, a total loss function, composed of the reconstruction loss, adversarial loss and TV loss, is used to optimize the network and enhance network stability.Finally, the proposed algorithm is validated on the CelebA dataset and Car dataset respectively, compared with three typical image inpainting algorithms.The effectiveness of the algorithm is proved in visual perception, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

Key words: Dense convolutional block, Generative adversarial networks, Image inpainting, Loss function, Skip connection