A night semantic segmentation method based on low-illumination enhancement and edge optimization
By combining a low-light enhancement and inpainting network based on Retinex theory and local contrast enhancement with an improved semantic segmentation network, the performance degradation problem of the nighttime semantic segmentation model in low-light environments is solved, achieving improved image clarity and edge structure preservation, and enhancing the accuracy and robustness of nighttime segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUIZHOU UNIV
- Filing Date
- 2026-01-23
- Publication Date
- 2026-07-21
AI Technical Summary
Existing nighttime semantic segmentation models exhibit significant performance degradation in low-light environments, struggling to adapt to changes in illumination, exposure imbalances, and edge degradation. This leads to uneven changes in object appearance and structural texture, increasing the difficulty of boundary recognition. Furthermore, existing enhancement methods are prone to introducing blurring or over-smoothing and lack end-to-end optimization.
We employ a low-light enhancement and inpainting network based on Retinex theory, local contrast enhancement, and adaptive feature fusion. This network is combined with an improved semantic segmentation network and a channel attention module. End-to-end training is performed by jointly optimizing the loss function through multiple tasks, thus achieving deep coupling between image enhancement and segmentation tasks.
It effectively improves the clarity and edge structure preservation of nighttime images, enhances the robustness and segmentation accuracy of the model in low-light environments, and performs particularly well in boundary recognition and small target recognition, with adaptive capabilities.
Smart Images

Figure CN121962610B_ABST