Entropy-Guided Deep Repair for Missing Depth Information in Depth Images

This article proposes a new entropy-guided deep repair strategy that effectively addresses the issue of missing depth information in raw depth images acquired from depth sensors. Firstly, the raw depth and color images are preprocessed to provide a connected hole map and grayscale image for subsequent tasks. Then, the concept of information entropy is introduced to identify invalid points that should be filled first by evaluating depth information in effective neighborhoods. Next, the depth values of invalid points are predicted based on color and gradient information guidance. In this step, color information ensures the accuracy of prediction, and gradient information is utilized to further improve the precision of repair. In this way, missing information in the raw depth image is effectively repaired. Finally, comparative experiments on the Middlebury dataset demonstrate that the proposed method exhibits better robustness and accuracy than other intelligent methods, providing new ideas and methods for relevant research in depth image processing.