CAM-Net: Memory Efficient and Context-Aware Mamba for Image Inpainting

View at publisher Summarise with VIRAT
Authors: J Saloni, P Akhshan, S. Chandrakala
Cited by: 0
References: 25

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Achieving visually and semantically coherent image inpainting remains challenging, particularly for large missing regions with limited contextual cues. Although GAN-based models are widely used, they often lack contextual information with free-form holes, leading to significant structural inconsistencies, impaired global coherence and loss of fine-grained texture details. We propose CAM-Net, a context-aware, lightweight GAN architecture designed for improved and semantically consistent inpainting. In addition to the context-aware features, we explore long-range dependency features with linear time complexity given by Mamba to build a memory-efficient inpainting model. We evaluated CAM-Net on two widely used inpainting benchmarks: CelebA-HQ and Places2. The results demonstrate that the proposed model delivers slightly better inpainting quality with a great reduction in the parameters than the existing state-of-the-art approaches. The code implementation is available at https://github.com/saloni1126/CAM-Net .

Fields of Study & Subjects

Generative Adversarial Networks and Image Synthesis Advanced Image Processing Techniques Digital Media Forensic Detection

Publication details

Journal / source
International Journal of Pattern Recognition and Artificial Intelligence
Volume
40
Issue
09
Publication date
19 March 2026
Language
EN