Low-Light Image Enhancement Based on Light Effect Region Suppression

Aiming at the coupling problem of light effect interference and non-linear degradation in low-light images, this paper proposes a double-stage enhancement framework integrating hierarchical illumination decomposition and adversarial training. In order to avoid the overcorrection phenomenon of traditional enhancement methods to the light effect area, a light sensitive network is constructed to decouple the reconstruction of the illumination and reflection layers. And, a dynamic exposure regulation module is designed to learn the brightness mapping function of local regions implicit neural representation. Then, the local adaptive gamma correction is performed. Furthermore, a multi-scale adaptive transformer architecture is proposed, which adopts a deformable window mechanism to achieve long-range dependency modeling, to enhance the recovery ability of local degradation features through gate convolutional. In addition, the adversarial co-training strategy is introduced to guide the model to generate results that meet the physical light distribution by jointly using multi-scale discriminators and spectral constraint losses. Experimental results show that the proposed method, which can achieve balanced enhancement in dark areas, leads in SSIM indicators and most of PSNR is optimal.