LAUSR

With the help of the convolutional neural network (CNN) and generative adversarial network (GAN), image inpainting has achieved remarkable advances in performance. However, there are still some bottleneck problems to be solved, such as unnatural colors and heavy-weight network. In this paper, we propose a lightweight generative network for image inpainting using feature contrast enhancement. We achieve feature contrast enhancement based on dilated convolution and channel attention. Dilated convolution enlarges the receptive field of the convolution kernel with the same number of parameters, while channel attention successfully enhances feature contrast using adaptive weights. For loss function, we utilize color enhancement and symmetric mean absolute percentage error (SMAPE) losses to achieve image completion with high quality and natural color instead of the reconstruction losses. The proposed network needs only 4.5M parameters for image inpainting. Experimental results demonstrate that the proposed method outperforms state-of-the-art ones in terms of visual quality and quantitative measurements.