A Lightweight CNN for Detail Enhancement and Color Correction of Low-Light Capsule Endoscopy

Wireless capsule endoscopy (WCE) is a non-invasive medical procedure that involves swallowing a small capsule equipped with a camera to capture images of the digestive tract. However, limitations in gastrointestinal structures and equipment performance may result in deficient illumination, thereby affecting diagnostic accuracy. In recent years, deep learning (DL) has been increasingly applied in the field of lowlight medical image enhancement. However, current DL methods often face challenges such as limited sensitivity to fine structures and color distortion. In this paper, we propose a novel method for addressing insufficient light in endoscopic images based on a residual neural network. We introduce a transform kernel (TFK) convolution to extract details from feature maps, making the recovered image more similar to real ones. Additionally, we propose a learnable image-guided enhancement block (IGEB), enhancing the input image in luminance and chroma distinctively at different stages of the network to improve the quality of output. The proposed approach significantly outperforms previous lowlight image enhancement (LLIE) algorithms both qualitatively and quantitatively, even with much fewer parameters. Furthermore, the restored images achieve better performance in some high-level tasks such as image segmentation.