A stereo matching method based on freqwise-stereo and image-enhancement module under adverse illumination

Stereo matching has become an important approach for depth estimation due to its low cost, high accuracy, and long measurement range. Deep learning based stereo matching models rely on visual cues such as texture and geometric information. However, under adverse illumination conditions—including low light and overexposure—these visual cues diminish significantly, resulting in poor generalization performance in such scenarios. To address this issue, an image enhancement module is integrated into the network to improve image quality at the input stage, thereby enabling high precision stereo matching under challenging lighting conditions. In addition, widely adopted iterative-optimization matching paradigms struggle to capture both high-frequency details near object boundaries and low-frequency information in smooth regions, which further limits their generalization. Therefore, a stereo matching network, Freqwise-Stereo, based on a Frequency Fusion Cost Volume, is proposed. By incorporating multi-frequency feature enhancement, the network achieves improved generalization in both high-frequency and low-frequency regions. Building upon this architecture and the image-enhancement module, a stereo matching method specifically designed for adverse illumination environments is introduced. Finally, low-light and overexposure datasets were constructed to evaluate the proposed method. Experimental results demonstrate that the method yields more accurate disparity estimation across different scenarios and illumination conditions.