High Fidelity Inpainting for Critical Applications Using Image Fusion

We propose a novel image restoration technique based on complementary information present in two or more images. While standalone image inpainting techniques fill the missing information artificially, use of spatiotemporal image sequence is suitable and reliable for critical applications, since it gathers the unavailable information in one image from other images in the sequence. There are two important issues to be dealt with before utilizing such sequences. Firstly, the images should be correctly registered in order to correctly access the undamaged images for in-painting the damaged one. Secondly, the illumination conditions should be taken care of, without which shadowing effects are generated since two images can differ in the illumination conditions, as well as sensor gain settings. In this work, a RANSAC based registration technique is used to perform efficient registration in presence of artefacts. Next, the dynamic range of the images are equalized to achieve illumination coherence. Further, we use fusion methods to incorporate information from all undamaged images into the damaged image. Cosmetic inpainting is used only if no other information is present. This work has uses in annotation removal from military and medical images and video, restoration of damaged video frames, etc. The results confirm the efficiency of the proposed method