Diverse receptive field based adversarial concurrent encoder network for image inpainting

Abstract

Image inpainting is nowadays demanding because of its wide applications such as removing the unwanted objects from the image or recovering the old corrupted photo. Existing approaches achieved superior performance with coarse-to-fine or progressive or recurrent architectures for image inpainting regardless of computational complexity. In these types, the disturbance at the first instance or first iteration may lead to semantically unambiguous results. Also, to inpaint the image with varying hole sizes it is desirable to focus on the diverse receptive fields without deeper network i.e, network with less number of parameters. Therefore, we have proposed a lightweight adversarial concurrent encoder architecture with a diverse receptive field for image inpainting. Here, the concurrent encoder is integrated with diverse receptive fields to benefit with lower computational complexity. The proposed method is compared with state-of-the-art (SOTA) methods on Places2 and Paris Street View dataset in terms of peak signal-to-noise ratio and structural similarity index. Along with the extensive results’ analysis and ablation study, the proposed method proves the effectiveness in terms of less computational complexity compared to existing SOTA methods.