HS-Former V3: An Uncertainty-Aware Hybrid Transformer–CNN Framework with Adaptive Frequency-Domain Learning and Edge-Aware Fusion for Low-Light Image Enhancement

improving low-light images is a very challenging vision task, since degradation in the low-light image is usually caused under complex low-light conditions that destroys the visibility of the image, sources of noise are often present, and the high-frequency structure information is lost. To address these problems, in this paper, we propose HS-Former V3: a hybrid Transformer-CNN setup, which aims to learn the modeling of global illumination together with the restoration of local structures. The setup contains a Swin Transformer V2 branch, a lightweight convolution branch, an Edge-Aware Cross Fusion (EACF) module, and a Dynamic Frequency-Selective Loss (DFL). Besides, we propose a masked-frequency pre-training strategy to further explore the effectiveness of unpaired low-light images and design an uncertainty estimation head to locate uncertain enhancement regions. The overall system is inspired by two deficiencies, i.e. deficiency of the traditional reconstruction objectives and the deficiency of features and efficiency of models. The evaluation protocol includes full-reference image-quality metrics, no-reference quality evaluation, computational complexity, corruption robustness, and downstream object detection. This structure will pave the way towards reproducible research on adaptive spectral supervision and edge-guided fusion, to be validated in a controlled successful implementation.