Infrared–Depth Drogue Target Detection via Frequency-Domain Enhancement and Decoupled Gated Fusion

An enhanced RT-DETR with frequency decoupling and orthogonal regularization for UAV infrared small target detection

Summary Infrared small target detection in unmanned aerial vehicle (UAV) imagery is pivotal for low-light surveillance. However, existing frameworks suffer from feature redundancy, suboptimal context modeling, and gradient vanishing during tiny target localization. To address these bottlenecks, we propose an enhanced real-time detection transformer tailored for drone-based infrared scenarios. First, the Ortho-Block module leverages orthogonal regularization to eliminate redundancy and purify target representations from background clutter. Second, the AIFI-HiLo module decouples scene semantics via high- and low-frequency attention within intra-scale interactions to capture dense targets. Furthermore, a dual-stream GLSA mechanism with lightweight deformable convolutions adapts to extreme scale variations. Finally, the NWD-SIoU loss introduces a dynamic scaling factor to provide smoother gradients and accelerate localization convergence. Experimental results on HIT-UAV (Harbin Institute of Technology Unmanned Aerial Vehicle dataset) show a 4.8% mAP50 improvement over the RT-DETR baseline, while evaluations on VisDrone2019 confirm robust cross-scene adaptability.

Object detection algorithm based on infrared-visible dual-modality feature fusion

CFM-YOLO, an infrared–visible dual-modality detection algorithm based on YOLOv11, is proposed to improve UAV object detection under adverse illumination and complex aerial backgrounds. The network is redesigned from three aspects: cross-modal feature extraction, lightweight feature fusion, and small-object-oriented detection. First, a Cross-Modality Fusion Mamba (CFM) module is introduced to promote channel-level interaction between visible and infrared features and to model long-range spatial dependencies with selective state-space modeling. Second, a lightweight feature fusion network is used to improve multi-scale information transmission while limiting redundant computation. Third, a P2 detection layer, Ghost convolution, and Focal-WIoU loss are incorporated to enhance small-object localization and alleviate the effect of imbalanced bounding-box samples. Quantitative experiments on the DroneVehicle dataset show that CFM-YOLO achieves 81.6% mAP@0.5 and 60.3% mAP@0.5:0.95, improving over the YOLOv11n-dual/base baseline by 8.4 and 5.5 percentage points, respectively. Qualitative results on the LLVIP dataset further indicate that the proposed method can reduce several missed detections in low-light pedestrian scenes. These results suggest that CFM-YOLO provides a competitive trade-off between detection accuracy and computational cost for UAV-based infrared–visible object detection.

CIFI-YOLO: A SWaP-Aware Object Detector for UAV Optical Sensors via Cross-Iterative Fusion and Gated Attention

Operating optical sensors on uncrewed aerial vehicles (UAVs) requires balancing high precision and severe Size Weight and Power constraints of edge hardware. Traditional detection algorithms often fail in high altitudes because small targets hide in complex ground clutter and sensor noise. CIFI-YOLO is presented as a hardware-compatible redesign optimized for edge sensing. Three innovations for aerial sensors are introduced within the YOLOv8-based framework. Environmental interference is suppressed by a gated spatio-channel attention module (GSCA-M). High-frequency spatial details lost during deep feature extraction are recovered through a cross-iterative fusion (CIF) scheme. Multiscale context aggregation is provided by an information enhancement module (IEM). Consistent performance gains are demonstrated through evaluations on the VisDrone2019, UAV-DT, and CODrone datasets. Compared with YOLOv8s, absolute AP50 gains of 6.1%, 5.5%, and 7.7% points are achieved by CIFI-YOLO on VisDrone2019, UAV-DT, and CODrone, respectively. The parameter count is reduced by 65.0%, and SRAM memory bottlenecks are alleviated by CIFI-YOLO. A latency of 9.08 ms is achieved on the RTX 3080 Ti evaluation platform, corresponding to approximately 110 frames/s and indicating the potential for real-time edge deployment.

Small-Target Detection via Fusion of Visible and Infrared Image Features

Visible–infrared small-target detection is challenged by weak single-modality representation, modality discrepancy, and the quadratic cost of dense cross-modal attention. We propose TFFB, a feature-level fusion detector that combines spatial feature compression (SFC), cross-attention modality enhancement (CME), and iterative cross-modal enhancement (ICME) to balance information exchange and computational efficiency. To further improve localization, we introduce Focaler-SIoU for small-box regression. On Anti-UAV300, TFFB improves the middle-fusion baseline from 76.2%/43.7% to 81.5%/48.6% in mAP@0.5/mAP@0.5:0.95, and TFFB with Focaler-SIoU reaches 83.6% and 50.2%, respectively. The results indicate that compact cross-modal interaction can strengthen visible–infrared UAV detection while keeping computational costs moderate.

Frequency–Spatial Collaborative Gated Attention Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) is challenged by low signal-to-noise ratios and complex background clutter. Existing methods remain insufficient in capturing spectral discrepancies and fusing dual-domain features. To address these limitations, we propose FSGANet, which improves frequency-domain clutter suppression through wavelet priors and adopts gated attention to blend dual-domain features, thereby significantly enhancing IRSTD performance. In particular, the model consists of three modules: 1) frequency-spatial collaborative gated attention (FSCGA) module, a dual-domain fusion attention module that captures multiscale spatial features and global frequency-domain information, and blends dual-domain representations to construct target features while suppressing irrelevant noise; 2) dynamic frequency refine (DFR) module, which employs compression, dynamic weight assignment, and recovery to further amplify the spectral signals of targets and attenuate clutter spectral components; and 3) encoder-aligned wavelet prior (EAWP), a wavelet-transform-based prior spectral cue that guides the frequency-domain selection block in FSCGA to precisely capture foreground–background spectral discrepancies. FSGANet contains 1.79-M parameters and 32.33 GFLOPs. Extensive experiments on four public datasets demonstrate the superior detection performance and efficiency of our model. The code is available at https://github.com/xiaodacheng01/FSGANet

TDSF-NET: a real-time target-driven sparse fusion network for RGB-infrared small object detection