Title: ROI adaptive lossy compression in dermatology using deep learning and superpixels algorithm

Authors: Saida Lemnadjlia; Ahlem Melouah; Amel Slim

Addresses: Laboratory of Research in Computer Science (LRI), Department of Computer Science, Badji Mokhtar-Annaba University, P.O. Box 12, Annaba, 23000, Algeria ' Laboratory of Research in Computer Science (LRI), Department of Computer Science, Badji Mokhtar-Annaba University, P.O. Box 12, Annaba, 23000, Algeria ' Laboratory of Research in Computer Science (LRI), Department of Computer Science, Badji Mokhtar-Annaba University, P.O. Box 12, Annaba, 23000, Algeria

Abstract: Medical images play a vital role in diagnosing and monitoring diseases, yet storage challenges arise due to their size. The necessity for compression becomes evident, though traditional methods compromise image quality. This study proposes a solution by selectively degrading non-critical image portions, adjusting the compression ratio based on content. The relevance of each segment is determined using a deep learning (DL) and super-pixels combination. Initially, a super-pixels method groups homogeneous pixels, and then the image is split into partitions, each classified by a DL model. A compression factor guides the ratio and degradation. Evaluation on skin images reveals superior results (CR: 42.07, PSNR: 63.87, MSE: 1.15, SS: 97.62) compared to existing techniques, affirming the method's success.

Keywords: deep learning; super pixel; image compression; jpeg; adaptive; loss of data; medical images; quality factor; semantic.

DOI: 10.1504/IJCVR.2026.155536

International Journal of Computational Vision and Robotics, 2026 Vol.17 No.2, pp.247 - 271

Received: 03 Aug 2023
Accepted: 04 May 2024
Published online: 05 Aug 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article