SpecuPro: An Integrated Method for Refined Specular Reflection Removal in Endoscopic Images
Jian Chen, Ali S. Shuaib, Ya Guo, Zhijian Zhang, Xinhong Song · 2024
The quality of endoscopic images is inevitably affected by varying degrees of specular reflections, which increases diagnostic difficulty. Although various methods for detecting and removing specular reflections have been proposed, most of the methods face challenges in achieving optimal segmentation accuracy and often perform poorly in restoring large reflective areas. In response, we introduce a method, SpecuPro, that integrates refined segmentation and restoration techniques for improved performance. Specular reflection regions are segmented by the proposed lightweight UNet variant, DSUNet++, while images are restored using an enhanced Criminisi algorithm. We evaluated our method using the publicly available real endoscopic dataset CVC-ClinicDB and compared it with methods from previous research. Quantitative metrics and qualitative visual results demonstrated that our method outperforms the existing methods in accurately segmenting specular reflections within a reasonable time frame and effectively inpainting large highlight areas.