Projector compensation on non- at textured surfaces is a challenging task. Existing methods struggleto correct significant distortions. Our approach, IDNet, overcomes these limitations with incorporatingthree Inception-like modules, by which we capture multiscale features using deformable convolutionand non-local attention mechanisms. Specically, we employ multiscale deformable convolutions forgeometric distortions and an Inception-like module for photometric compensation. The addition ofa residual-based non-local attention module further enhances both processes. Experimental results show that IDNet performs as the state-of-the-art comparing with the existing methods.