کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Illumination-Aware Capsule Endoscopy Enhancement: An Ablation Study Using Synthetic Paired Data

نویسندگان :

Nima Hosseini¹ Niloufar Hosseini² Ehsan Adibnia³

1- Islamic Azad University E-Campus 2- University of Sistan and Baluchestan 3- University of Sistan and Baluchestan

کلمات کلیدی :

Wireless capsule endoscopy،image enhancement،ablation study،synthetic data،illumination adjustment،deep learning

چکیده :

Wireless Capsule Endoscopy (WCE) is widely used for non-invasive examination of the small bowel; however, the acquired images frequently suffer from non-uniform illumination, low-light conditions, and contrast degradation. These limitations significantly reduce visual interpretability and hinder the performance of computer-aided diagnostic systems. A fundamental challenge in addressing this problem is the absence of paired low- and high-quality training data, which restricts the applicability of conventional supervised learning methods. In this paper, we present a two-stage enhancement framework tailored to PillCam SB2 imagery. First, a configurable Professional Degradation Module is introduced to synthesize realistic low-quality images by modeling domain-specific degradations such as illumination decay, vignetting, sensor noise, motion blur, and color shift. Second, a lightweight enhancement network incorporating an Adaptive Illumination Adjuster is trained using the generated paired data. The primary contribution of this work is a structured ablation study that quantitatively evaluates the relative importance of individual architectural and data-generation components. Experiments conducted on a large-scale synthetic dataset derived from 8,000 endoscopic images demonstrate that removing the illumination adjustment mechanism leads to the largest performance degradation, increasing the final validation loss by 11.1%, whereas excluding color-shift simulation has a comparatively minor impact. Although the proposed model achieves a PSNR of 24.16 dB and an SSIM of 0.809, the results also reveal persistent challenges in preserving natural color appearance. Overall, this study provides component-level insights that support evidence-based architectural design for WCE image enhancement and highlights illumination modeling as the dominant factor for performance improvement.