Improved Small Object Semantic Segmentation Under Low-Light Condition

Limited lighting conditions in real-world industrial production environments significantly affect the performance of segmentation models in small-object scenarios. To enhance segmentation accuracy for small objects, we introduce GS-Seg, a lightweight and generalpurpose segmentation model which follows an encoderdecoder architecture, a shallow-layer feature extractor is integrated into a minimal backbone to capture fine-grained details. This module strengthens small-object learning by incorporating extracted features during the decoding process. Additionally, multi-scale feature fusion was utilized to enrich abstract semantic representations, enhanced segmentation accuracy in complex environments. The performance evaluation indicates that GS-Seg surpassed current State-Of-The-Art (SOTA) methods in segmenting small objects, achieved superior mIoU and edge F1 scores. Furthermore, the lightweight network architecture ensures compatibility with edge-device deployment, supporting mainstream NVIDIA and Rockchip processors.