Image Reprocessing of Occluded Object

Abstract

The occlusion in images, where parts of an object in an image are hidden or obstructed by other elements, often hampers their interpretability and utility in various applications. This paper aims to address the challenge of occlusion using a learning approach. This paper proposes a technique that uses Convolutional Neural Networks (CNNs) to inpaint (reconstruct) missing or occluded regions within images with the help of sur- rounding information of the occluded area. The model discussed in the paper was trained on the Place4Net dataset, which focuses on different scenes such as urban, rural, indoor, and outdoor environments. Our method aims to accurately restore occluded regions, thereby recovering the visual information lost due to occlusion. The typical convolution operation, which covers all the pixels of the image, is slightly modified and is termed Partial Convolution, where features are extracted from the desired (non- occluded) regions. The effectiveness of our method was evaluated through user studies, where participants provided feedback on the reconstructed images. The model in this paper was tested on the old images that were collected from the people in the vicinity, and the collected feedback, quantified using a Mean Opinion Score (MOS), indicated that the majority of participants found the restored images to be of good visual quality, demonstrating the utility and robustness of our approach.