Objective The endoscope, a key instrument for target observation, information analysis, and operation execution in confined spaces, offers non-contact and non-invasive capabilities. It is widely used in clinical evaluations of the digestive tract and in the inspection and diagnosis of confined environments such as industrial pipelines via nondestructive testing. However, non-uniform illumination from the endoscopic light source, combined with the smooth surfaces of human organs or pipelines, often causes specular reflection. This leads to glare in the endoscopic field of view, which obscures target details and severely compromises imaging quality and diagnostic accuracy. Traditional glare suppression methods primarily rely on RGB three-channel cameras for image processing. However, these approaches are computationally intensive and limited to only three spectral bands, thereby constraining the available spectral information. As a result, they are inadequate for accurate online analysis of histochemical component differences at specific wavelengths. To overcome these limitations, this study proposes a novel glare elimination method that integrates endoscopic hyperspectral imaging with multi-exposure fusion.

Methods To effectively suppress glare in hyperspectral images while preserving their rich spectral and spatial information, this study first analyzes the image and spectral characteristics of the central glare region and the peripheral non-glare region. Based on this analysis, a threshold mask is constructed to identify abnormal pixels in the hyperspectral data cube—such as those exceeding the camera’s saturation threshold or exhibiting low overall signal intensity—for subsequent removal. Then, the maximum intensity of each pixel across the entire spectral band is calculated, and the spectrum of each pixel is normalized. Finally, the spectral intensity curves of the remaining pixels in each image are averaged and fused. For comparative purposes, three existing glare suppression methods were selected. Image quality was evaluated using the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), while spectral fidelity was assessed based on the mean absolute error (MAE), root mean square error (RMSE), and spectral angle mapper (SAM).

Results and Discussions The endoscopic hyperspectral imaging system captured nine hyperspectral images at different exposure times. Three RGB bands (640, 532, and 471 nm) were selected to generate pseudo-color images for visualization. At high exposure times, the central area reached the maximum intensity value of 4095, indicating overexposure and a consequent loss of detail in these bright regions. In contrast, the peripheral areas exhibited higher intensity values and retained more structural detail (Fig.2). As the exposure time decreased, the overexposed area in the center gradually diminished, allowing previously obscured details to be recovered. However, noise and burrs became more pronounced in the periphery (Fig.3), which could obscure fine details and required appropriate removal. After normalizing each pixel, the overall contrast of the hyperspectral images was significantly enhanced. Details that were previously lost in the glare region at high exposure times, as well as those obscured in the peripheral regions at low exposure times, were effectively restored (Fig.4). Finally, in comparison with other methods, the hyperspectral image obtained through average fusion demonstrated superior clarity and contrast in structural details across the entire field of view. It also achieved relatively high PSNR and SSIM scores (with the latter approaching 1, as shown in Tab.1), indicating that the structural information was highly consistent with the reference image. Moreover, the proposed method offered faster computational speed. In terms of spectral fidelity, our method exhibited a higher similarity to the original spectrum in both the central and peripheral regions ( Fig.6). Quantitative evaluations based on different spectral similarity metrics (Tab.2) confirmed that our method achieved the lowest index values across different positions, demonstrating its superior spectral reconstruction accuracy. This approach not only effectively suppressed central glare but also preserved structural and spectral information at the edges.

Conclusions This study tackles the challenge of glare in endoscopic imaging by introducing a novel elimination method based on hyperspectral and multi-exposure image fusion. Qualitative and quantitative comparisons with existing techniques demonstrate that the proposed approach not only effectively suppresses specular reflections but also enhances textural details across the entire field of view. Moreover, it exhibits superior performance in preserving spectral fidelity. This work provides a valuable tool for the spectral analysis of target surfaces in confined spaces and offers a promising solution for accurate, online tumor diagnosis.