← 研踪 PaperTrailCVF · CVPR 2026:cvf:999ce060a177eaec0dd7a8fcd58c8294 · 记录 #1
BDNet:Bio-Inspired Dual-Backbone Small Object Detection Network
来源日期（仅到年份） 2026
来源记录日期 2026
雷达观测 2026-11-05 02:53
摘要
In remote sensing images, small objects often exhibit low color contrast and blurred edges, leading to suboptimal feature extraction. Physiological studies indicate that the LGN/V1-V2-V4 pathway provides color-opponent sensitivity and hierarchical enhancement for color information extraction, whereas the V1-V4 pathway exhibits strong orientation selectivity for edge extraction. Integrating these complementary visual signals in the V4 region can substantially improve target discrimination. Motivated by these findings, we propose a dual-backbone network (BDNet) to enhance feature extraction for small objects. BDNet adopts a parallel architecture to capture fine-grained features from color and edge cues. Specifically, the color-extraction backbone simulates the color-opponent mechanism in LGN/V1 via a Color Antagonism Module (CAM) to amplify color differences, and further mimics the chromatic processing hierarchy in V2 using a Visual Cortex Hue Enhancement Module (VCHM) to enrich hue representations. Together, these two modules alleviate low color contrast. The edge-extraction backbone simulates the orientation selectivity of receptive fields in V1 through an Orientation Selective Module (OrSM) to select and enhance salient edges, thereby reducing edge blurring from fragmented edge responses. Finally, the two feature types are fused via a Feature Fusion Module (FFM) that emulates integration in V4, yielding a comprehensive feature representation. Experiments demonstrate that BDNet outperforms state-of-the-art methods on the VisDrone2019, NWPU VHR-10, and AI-TODv2 datasets, providing a bio-inspired solution for small-object detection in remote sensing images.
论文发现出处
这些页面提供论文线索，不代表同行评审或质量认证。
观测到的版本
平台记录号代表本站观测到的元数据变化，不是作者修订号。来源日期仅到月份或年份时，按该月或该年首日存档。
记录 #1 · 2026 · 最新观测版本
BDNet:Bio-Inspired Dual-Backbone Small Object Detection Network
In remote sensing images, small objects often exhibit low color contrast and blurred edges, leading to suboptimal feature extraction. Physiological studies indicate that the LGN/V1-V2-V4 pathway provides color-opponent sensitivity and hierarchical enhancement for color information extraction, whereas the V1-V4 pathway exhibits strong orientation selectivity for edge extraction. Integrating these complementary visual signals in the V4 region can substantially improve target discrimination. Motivated by these findings, we propose a dual-backbone network (BDNet) to enhance feature extraction for small objects. BDNet adopts a parallel architecture to capture fine-grained features from color and edge cues. Specifically, the color-extraction backbone simulates the color-opponent mechanism in LGN/V1 via a Color Antagonism Module (CAM) to amplify color differences, and further mimics the chromatic processing hierarchy in V2 using a Visual Cortex Hue Enhancement Module (VCHM) to enrich hue representations. Together, these two modules alleviate low color contrast. The edge-extraction backbone simulates the orientation selectivity of receptive fields in V1 through an Orientation Selective Module (OrSM) to select and enhance salient edges, thereby reducing edge blurring from fragmented edge responses. Finally, the two feature types are fused via a Feature Fusion Module (FFM) that emulates integration in V4, yielding a comprehensive feature representation. Experiments demonstrate that BDNet outperforms state-of-the-art methods on the VisDrone2019, NWPU VHR-10, and AI-TODv2 datasets, providing a bio-inspired solution for small-object detection in remote sensing images.
查看此版本解读 →查看此版本原文 ↗