VResNet: A Deep Learning Architecture for Image Inpainting of Irregular Damaged Images

In computer vision, image inpainting is a famous problem to automatically reconstruct the damaged part of the image according to the undamaged portion of an image. Inpainting irregular damaged areas in the image is still challenging. Deep learning-based techniques have given us a fantastic performance over the last few years. In this paper, we propose VResNet, a deep-learning approach for image inpainting, inspired by U-Net architecture and the residual framework. Since deeper neural networks are extra hard to train, the superficial convolution block in U-Net architecture is replaced by the residual learning block in the proposed approach to simplify the training of deeper neural networks. To develop an effective and adaptable model, an extensive series of experiments was conducted using the Paris-Street-View dataset. Our proposed method achieved notable results, including a PSNR of 20.65, an SSIM of 0.65, an L1 Loss of 6.90, and a total loss (L Total) of 0.30 on the Paris-Street-View dataset. These outcomes clearly demonstrate the superior performance of our model when compared to other techniques. The paper presents both qualitative and quantitative comparisons to provide a comprehensive assessment of our approach. © 2024, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.