Published Paper Details:

Image inpainting is a challenging computer vision task that focuses on reconstructing missing or occluded regions in images, preserving semantic and structural consistency. This paper presents a novel deep learning-based inpainting framework combining a generative adversarial network (GAN)-based generator and a Context-Aware Network (CANet) to address the limitations of traditional convolutional approaches. The GAN generator employs an encoder-bottleneck-decoder architecture, effectively capturing global image features and generating plausible content for masked areas. Meanwhile, CANet enhances the inpainting quality by incorporating context-sensitive attention mechanisms through Context-Aware Convolutions (CAB) and Cross-Scale Context Attention (CSCA), enabling better feature propagation and spatial coherence. The proposed approach is trained and evaluated on the large-scale CelebA face dataset with centrally masked regions to simulate occlusions. Quantitative and qualitative results demonstrate the model's capability to produce visually realistic and semantically consistent inpainted images, significantly outperforming baseline masked inputs. This framework not only achieves effective restoration for face images but also generalizes to other natural images, as shown by experiments on external datasets. The study underlines the importance of attention-based modules in image restoration tasks and paves the way for future improvements, including handling irregular mask shapes, higher resolution outputs, and real-time applications in photo editing and augmented reality.