Yimeng Li

In all-weather Earth observation and complex unstructured environments, traditional single-modal remote sensing object detection often fails due to low illumination and strong background interference. While RGB–infrared fusion provides complementary information, existing methods are typically computationally intensive and struggle with dense small objects and modality discrepancies, limiting their deployment on resource-constrained platforms. To address these challenges, we propose ISDG-Net, a lightweight and efficient visible-infrared dual-modal object detection framework specifically tailored for edge deployment. ISDG-Net integrates four core components: (1) a channel-separated inverted bottleneck backbone (IBC-Conv) that reduces parameter redundancy while preserving modality-specific semantics; (2) a dynamic sparse attention module (DySparse) based on Bi-Level Routing Attention, enabling long-range dependency modeling with low computational cost; (3) an adaptive spatial fusion detection head (Detect-SASD) that aligns visible and infrared features at the pixel level to resolve semantic inconsistency and scale mismatch; and (4) a geometry-aware IoU selector (GIS) that mitigates over-suppression in crowded scenes by incorporating multi-dimensional geometric constraints into post-processing. Extensive experiments on the VEDAI, M3FD, and LLVIP datasets demonstrate the effectiveness and efficiency of ISDG-Net. It achieves 55.1% and 77.1% mAP@0.5 on VEDAI and M3FD, respectively, and 93.7% mAP@0.5 with 89.7% recall on LLVIP, while maintaining a compact model size of 4.2 M parameters and 11.3 GFLOPs. These results validate that accurate RGB–infrared detection is achievable under strict resource constraints, making ISDG-Net well-suited for deployment in edge-based remote sensing systems.

In drone aerial photography applications, small object detection is crucial. For instance, it enables locating missing individuals on the ground during search-and-rescue operations, identifying distant vehicles in traffic monitoring, and detecting early-stage pest infestations in agricultural fields. However, aerial images present a unique challenge: due to the high flight altitude of drones, targets occupy only a minimal pixel area. Combined with complex backgrounds and sparse features, small objects are easily obscured by surrounding environments. To address these issues, this paper proposes the EFPN-YOLO model based on YOLOv12n. First, we introduce the Feature-Sharing Convolution (FSConv) module, which extracts multi-scale features with low parameter requirements through shared convolution kernels and multi-scale sparse sampling. Second, by integrating deformable convolutions with a dual-channel attention mechanism, we develop the Enhanced Dual-Dimensional Calibration (EDDC) module, significantly improving spatial feature modeling capabilities and feature enhancement effectiveness. Finally, we construct the RC-FPN architecture, employing a bidirectional fusion structure and diagonal cross-layer skip connections to minimize information loss. Meanwhile, the Bottleneck structure in the C3K2 module is replaced with the RepViTBlock to construct the C3k2_RVB module, which enhances the multi-scale feature expression ability through a two-stage design of spatial and channel mixing. On the VisDrone2019 dataset, the model’s mAP50 improved from 33.9% to 40.7%; on the TinyPerson dataset, it rose from 13.9% to 19.2%; and on the NVIDIA Jetson Orin Nano 8 GB superplatform, the model achieved a frame rate (FPS) of 15. Experiments demonstrate that EFPN-YOLO excels in small object detection and holds significant practical value.