Image Inpainting with Bilateral Convolution

Due to sensor malfunctions and poor atmospheric conditions, remote sensing images often miss important information/pixels, which affects downstream tasks, therefore requiring reconstruction. Current image reconstruction methods use deep convolutional neural networks improve inpainting performances as they have a powerful modeling capability. However, learn different features with the same group of kernels, restricts their ability handle diverse corruptions results in color discrepancy blurriness recovered images. To mitigate this problem, paper, we propose an operator called Bilateral Convolution (BC) adaptively preserve propagate information from known regions missing data regions. On basis vanilla convolution, BC dynamically propagates more confident features, weights input patch according spatial location feature value. Furthermore, capture range dependencies, designed Multi-range Window Attention (MWA) module, is divided into multiple sizes non-overlapped patches for several heads, then these are processed by window self-attention. With MWA, bilateral convolution network inpainting. We conducted experiments on datasets typical verify effectiveness generalization our network. The show that captures between unknown regions, generates appropriate content various corrupted images, has competitive performance compared state-of-the-art methods.