Jichang Guo

Semantic segmentation in nighttime conditions remains challenging due to severe noise and reduced visibility, limitations exacerbated by traditional processing of 8-bit sRGB images. To address this, we introduce the NRSS dataset, the first publicly available nighttime RAW image dataset specifically designed for semantic segmentation. It contains images from diverse typical nighttime scenes, with pixel-level annotations and a 14-bit dynamic range. We further propose an efficient RAW Image Processing network (RAWIPNet), optimized jointly and end-to-end with downstream segmentation models. Our RAWIPNet leverages a Dilated Sparse Self-Attention block (DSSA) to suppress irrelevant noise and a multi-scale feed-forward network (MSFFN) to enhance contextual feature representations. Extensive experiments demonstrate that our method achieves superior nighttime semantic segmentation performance on the NRSS dataset compared to state-of-the-art methods. Moreover, evaluations across RAW datasets from different scenes and downstream tasks confirm the robust generalization of our method. The dataset will be made available here.

Detecting small objects in unmanned aerial vehicle (UAV) remote sensing imagery is essential for real-world applications such as forest pest monitoring and smart city management. However, existing object detection algorithms struggle with dense objects, significant scale variations, and cluttered backgrounds. To address these challenges, we propose a cross-layer feature fusion YOLO framework (CF-YOLO) based on YOLOv11s. CF-YOLO integrates three key modules: (1) a split-block attention module (SBAM) that improves contextual perception by capturing both global semantic relationships and fine local details; (2) a cross-layer multi-scale feature fusion (CMFF) module that effectively combines shallow spatial details with deep semantic features, enhancing the localization and recognition of small objects; and (3) a multi-branch downsampling module that preserves high-resolution shallow information through diverse downsampling paths, providing richer inputs for feature fusion. Evaluations on the VisDrone benchmark demonstrate that CF-YOLO outperforms YOLOv11s, improving precision by 7.57%, recall by 6.04%, and mAP@0.5 by 8.07%, while maintaining a comparable number of parameters. The proposed method significantly improves the detection accuracy for small objects in UAV remote sensing images without increasing the number of parameters.