IDNet: An inception-like deformable non-local network for projection compensation over non-flat textured surfaces

Projector compensation on non-flat, textured surfaces represents a formidable challenge in computational imaging, with conventional convolution-based methods frequently encountering critical limitations, especially in image edge regions characterized by complex geometric transformations. To systematically address these persistent challenges, we introduce IDNet, an innovative framework distinguished by its multi-scale receptive feature extraction modules. Central to our approach are multi-scale deformable convolution modules that dynamically adapt to geometric distortions through intelligently flexible sampling positions and precise offset mechanisms, which significantly enhance processing capabilities in intricate distortion regions. By strategically integrating non-local attention mechanisms, IDNet comprehensively captures global contextual information, thereby substantially improving both geometric and photometric compensation accuracy. Our experimental validation demonstrates that the proposed method achieves comparable compensation performance to existing approaches, particularly in the most challenging and geometrically complex edge regions of projected images.