Underwater image enhancement via boundary-aware gradient domain reconstruction and intuitionistic fuzzy set similarity measure-guided wavelet fusion for lowlight conditions

Due to the light absorption and scattering effects of water in dark underwater environments, along with the sensitivity limitations of imaging devices, captured images generally suffer from color distortion, low contrast, blurred details, and uneven illumination. These issues severely constrain the performance of underwater vision systems. Existing underwater image enhancement (UIE) algorithms are mostly designed for well-lit scenes and perform poorly in lowlight underwater environments, struggling to achieve a balance between image clarity and scene fidelity. Therefore, this paper proposes an UIE via boundary-aware gradient domain reconstruction and intuitionistic fuzzy set (IFS) similarity measure-guided wavelet fusion for lowlight conditions. The proposed method is overall divided into four progressive subtasks. First, a differentiated compensation mechanism is established. Combined with an iterative compensation model for pixel distribution optimization, precise color correction of the image is achieved through pixel clipping and grayscale stretching. Second, global brightness enhancement subtask is realized by means of virtual exposure image fusion. Third, a dual-branch parallel enhancement strategy is adopted to strengthen the dominant features of the image. For global contrast enhancement, the brightness-enhanced image is converted to the Lab color space. On one hand, contrast limited adaptive histogram equalization (CLAHE) is applied to the L-channel for contrast enhancement; on the other hand, a zero-symmetric adaptive compensation strategy is used to optimize the distribution of the a and b color channels. For detail enhancement, boundary-aware gradient domain reconstruction is employed to restore image details. Finally, the feature-enhanced image is decomposed into high-frequency (HF) and low-frequency (LF) subbands via wavelet transform. The LF subbands are processed through synthetic feature-weighted fusion, while the HF subbands are integrated using an IFS similarity measure-guided wavelet fusion approach. The subband information is then reconstructed through inverse wavelet transform to obtain the final enhanced image. Experimental results on six public datasets demonstrate that the proposed method achieves competitive or leading performance against 16 mainstream methods. In particular, it ranks first in the overall average rank and in detail metrics, and is significantly better than most comparison methods according to Friedman and Nemenyi tests, while remaining competitive in color-related metrics. Moreover, our method shows encouraging performance in downstream tasks such as image matching, image segmentation, and saliency detection, and exhibits strong generalization capability to complex degraded scenes including haze and sand dust.