Complete Recovery and Health Status Detection of Roller Tank Lugs Using Image Inpainting Based on Unordered Image Stitching

Abstract

To address the challenge of inaccurate detection of the health status of roller tank lugs in the case of occlusion, this article proposes a nonlearning-based image inpainting method for roller tank lugs. The approach employs an unordered image stitching algorithm to effectively recover the occluded roller tank lugs and facilitate accurate detection of their health status. In terms of occluded region extraction, this article proposes an extraction algorithm based on a binary tree model, which effectively identifies and extracts the occluded regions by estimating the nonoverlapping areas in the reference image and subsequent multiframe images, thereby accurately extracting the images of the polyurethane wheels on roller tank lugs. This article presents an unordered image stitching technique that does not require image sorting and can directly perform stitching, effectively addressing the image distortion issues caused by cumulative stitching errors in traditional unordered stitching methods. The experimental results show that the proposed algorithm outperforms the traditional algorithm in both qualitative analysis and quantitative metrics in terms of occluded region extraction and image stitching. Compared with traditional image inpainting algorithms, the proposed method can recover the occluded portions of roller tank lugs more efficiently, with an average error rate of just 1.57% in the calculation of wear of its completed images. These results indicate that our approach meets the practical needs of engineering applications.