DFRE-Net: a Dual-branch Cross-feature Fusion with Regional Enhancement Network for Image Inpainting

초록

Image inpainting is a critical task in computational intelligence, particularly in AI-driven big data environments, where it involves synthesizing missing image content based on available information to restore occluded or damaged regions. As a learning-based fusion technique, image inpainting leverages AI to intelligently reconstruct missing or corrupted areas. To overcome the limitations of existing approaches, where convolutional neural networks (CNNs) struggle with global semantic coherence and Transformers lack sensitivity to fine local details. In this paper, the authors introduce DFRE-Net, a dual-branch cross-feature fusion network with regional enhancement for high-fidelity image inpainting. The encoder adopts a computationally intelligent dual-branch architecture: a CNN branch with multiscale residual connections extracts high-frequency details, whereas a Transformer branch captures long-range contextual relationships. The decoder incorporates a Cross-feature Fusion Module (CFFM), utilizing multidilation depthwise separable convolutions for adaptive local-global feature integration, and a Regional Enhancement Module (REM) with mask-guided optimization to reduce foreground-background interference. Experimental results demonstrate that the proposed method outperforms state-of-the-art models on the Paris StreetView, Places365, and CelebA-HQ datasets, validating the effectiveness of cross-architecture feature fusion in addressing complex occlusion inpainting tasks.

키워드