Scholarlink

<p>Image inpainting is an important task in computer vision that focuses on restoring missing parts of an image. While recent advancements have been made in this area, challenges such as unrealistic fillings, fine-texture issues, and high computational demands persist. To address two main concerns–limited receptive fields and shallow understanding of images–we propose a novel sequential processing strategy, MADETFE (Mask-Aware Dual Expert with Transformers and FFC-based high-fidelity image inpainting). This strategy first uses a large-mask inpainting technique, LaMa, which leverages Fast Fourier Convolutions (FFC) to enhance the receptive field via a frequency-domain transformation, thereby addressing the problem of contextual fragmentation. Next, it integrates a Mask-aware Transformer (MAT) that leverages transformers to improve semantic comprehension, directly addressing issues related to shallow understanding and fine-texture artefacts. Our sequential processing strategy significantly boosts performance, achieving over a 50% enhancement in Learned Perceptual Image Patch Similarity (LPIPS) scores relative to the LaMa Model and nearly an 8% improvement in Paired Inception Discriminative Score (P-IDS) relative to the Dual Coupled Feature (DCF) model (2025). This approach demonstrates the effectiveness of multi-stage fusion methods in image inpainting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MADETFE : Mask aware dual expert with transformers and FFC based high-fidelity image inpainting

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Amit Chauhan,
  • Rounak Gera,
  • Rajni Mohana

摘要

Image inpainting is an important task in computer vision that focuses on restoring missing parts of an image. While recent advancements have been made in this area, challenges such as unrealistic fillings, fine-texture issues, and high computational demands persist. To address two main concerns–limited receptive fields and shallow understanding of images–we propose a novel sequential processing strategy, MADETFE (Mask-Aware Dual Expert with Transformers and FFC-based high-fidelity image inpainting). This strategy first uses a large-mask inpainting technique, LaMa, which leverages Fast Fourier Convolutions (FFC) to enhance the receptive field via a frequency-domain transformation, thereby addressing the problem of contextual fragmentation. Next, it integrates a Mask-aware Transformer (MAT) that leverages transformers to improve semantic comprehension, directly addressing issues related to shallow understanding and fine-texture artefacts. Our sequential processing strategy significantly boosts performance, achieving over a 50% enhancement in Learned Perceptual Image Patch Similarity (LPIPS) scores relative to the LaMa Model and nearly an 8% improvement in Paired Inception Discriminative Score (P-IDS) relative to the Dual Coupled Feature (DCF) model (2025). This approach demonstrates the effectiveness of multi-stage fusion methods in image inpainting.