A New Method for Uneven Illumination Correction of Wireless Capsule Endoscopy Images

Uneven illumination (UI) in Wireless Capsule Endoscopy (WCE) images poses a significant challenge for accurate diagnosis and image analysis. To address this issue, we propose a novel method, named Triplet Clustering Fusion Autoencoder (TCFA), which accounts for different levels of UI severity. To achieve this, we introduce a distortion level encoder utilizing the triplet loss function to classify the distortion severity. Additionally, we propose a Variational Cross-Attention module to efficiently correct UI by extracting both global and structural local information. Furthermore, we leverage a pre-trained network to extract relevant structural features from WCE images. Our method demonstrates significant improvements in image enhancement tasks, as evidenced by extensive experiments conducted on WCE images from the Kvasir-Capsule dataset.