Improved self-calibration image enhancement algorithm based on attention mechanism and its application in maritime low-light images

In order to solve the problems of the insufficient brightness enhancement, low sharpness, and color distortion of existing maritime low-light image enhancement algorithms, this paper proposes an algorithm based on improved self-calibrated illumination (SCI) learning.

An attention mechanism is introduced to enhance the unevenly illuminated regions in low-light images to different degrees; an illumination adjustment module is constructed to secondary the intermediate outputs of the illumination learning process; a denoising module is introduced to ameliorate the problem of noise in dark regions being amplified with the enhancement of the brightness; and batch normalization (BN) is changed to batch channel normalization (BCN), which utilizes channel and batch dimensions to adaptively combine the normalized outputs. Image quality is then evaluated in both subjective and objective aspects.

Experiments are conducted under three testing datasets. The results show that when compared with the original unimproved algorithm, the improved algorithm not only improves the image brightness, but also provides higher color richness and no color distortion, while improving the standard deviation by an average of 20.01%, reducing the natural image quality evaluation by 9.16%, and improving the average gradient and information entropy by 23.68% and 6.46% respectively.

The improved algorithm represents a breakthrough in image visual quality, enabling the better enhancement of maritime low-light images in different environments.