Specular Region Detection and Covariant Feature Extraction

Abstract

Endoscopy images pose a distinct set of challenges, such as specularity, uniformity, and deformation, which can obstruct surgeons’ observations and decision-making processes. These hurdles complicate feature extraction and may ultimately lead to the failure of a surgical navigation system. To tackle these obstacles, we introduce a Modified Maximal Stable Extremal Region (MMSER) detector that specifically targets fine specular regions. Subsequently, we ingeniously fuse the capabilities of MMSER and saturation region properties to precisely identify specular regions within endoscopy images. Furthermore, our approach harnesses the shared properties of covariant features and endoscopic imaging to detect features in intricate regions, such as low-textured and deformed areas. Surpassing contemporary methods, our proposed technique demonstrates remarkable performance when evaluated on the available CVC-ClinicSpec datasets. Our method has shown improvements in accuracy, recall, f1-score, and Jaccard index by 0.21%,25.42%,7.77% snd 11.77%, respectively, when compared to recent techniques. Owing to its exceptional ability to accurately pinpoint specular regions and extract features from complex areas, our approach holds the potential to significantly advance surgical navigation. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.