ABSTRACT The small dynamic range available in imaging devices, along with poor ambient light, often produces underexposed or overexposed images. In the realm of machine vision systems, the high dynamic range environment presents numerous challenges. As a result, they limit the progress and widespread use of related technologies. The article presents an image exposure correction methodology based on wavelet transforms and deep learning. The process uses a dual‐path mechanism. First, a reverse operation inverses the image, creating a new version. Wavelet decomposition of the input image and its input image inverts separates it into low‐frequency and high‐frequency components. The exposure correction network processes luminance‐related low‐frequency components. On the other hand, a residual network enhances the texture‐detail high‐frequency components. The inverse wavelet transform is then applied to the corrected low‐frequency and enhanced high‐frequency components to obtain an initial corrected image. A fusion network takes the original input image and the two intermediate corrected results from the dual‐path to produce the final output. Our method was validated on the public dataset ME. Quantitative results demonstrate that, compared to state‐of‐the‐art methods, our approach achieves average improvements of 1.015 dB and 0.037 in PSNR and SSIM, respectively, on the test dataset, confirming its superiority in enhancement and detail restoration.
Xu et al. (2026) studied this question.