.png){kind=icon}
.png){kind=icon}
Return
Lightweight PAG-YOLO Algorithm for Small Object Detection in Remote Sensing
.png){kind=icon}
PRE
AI
J
R
X
H
X
Y
DOI：10.3788/LOP252087.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Objective Small object detection in remote sensing imagery remains difficult because targets are tiny, densely packed, and often confused with cluttered textures or low-contrast backgrounds. These properties lead to feature dilution after down-sampling, occlusion-induced ambiguity, and high false positives triggered by look-alike structures (e.g., rooftops, markings). To address these issues while keeping models lightweight for real-time scenarios, we propose PAG-YOLO, a compact detector that follows a shallow extraction-deep focusing-pre-head refinement paradigm. The goal is to expand the effective receptive field for early textures, selectively amplify small-object responses in high-level features, and deliver context-aware, detail-preserving representations to the detection head without increasing computational burden. Methods PAG-YOLO is built upon the YOLOv12 baseline and introduces three lightweight modules that form a complementary loop. First, a PINWHEEL WEIGHTED COnvolution (PW_WConv) is inserted into the backbone's shallow stages to capture directional edges and textures crucial for tiny targets. PW_WConv parallelizes four axial convolutions (horizontal/vertical and their counterparts) with a pointwise branch; outputs are concatenated and fused by 1x1 convolution, enlarging the effective receptive field while preserving cost. This pinwheel design improves sensitivity to slender, orientation-specific cues commonly seen in aerial imagery. Second, an adaptive coordinate dynamic spatial attention (ACDSA) is placed at the backbone output. It uses coordinate-guided pooling to encode position and drives a compact kernel generator to produce multi-scale dynamic spatial kernels. The fused responses emphasize small-object regions and suppress texture clutter, with residual injection for stability. Third, a Global-Local Channel and Spatial Attention (GLCAS) is inserted before the detection head. It computes global and local channel dependencies in parallel, couples them with a lightweight spatial attention map, and fuses all paths using learnable weights, ensuring the head receives features that integrate global semantics with local textures. Training uses stochastic gradient descent (SGD) on 512 pixel x 512 pixel inputs for 300 epoch with batch size 32, adhering to YOLOv12n's recommended hyperparameters for fair comparison. Metrics include precision (P), recall (R), mAP@0.5, and mAP@0.5:0.95, alongside Params and GFLOPs. Results and Discussions On DOTA v1.0, ablations show each module contributes complementary gains. Full PAG-YOLO reaches P=80.7%, R=64.3%, mAP@0.5 is 72.7%, mAP@0.5:0.95 is 48.2%; removing PW_WConv causes the largest drop (mAP@0.5 down arrow to 68.9%), underscoring the value of early directional texture capture. Excluding ACDSA mainly hurts high-IoU alignment (mAP@0.5:0.95 down arrow to 47.1%), while omitting GLCAS weakens contextual suppression in dense scenes (mAP@0.5 down arrow to 70.7%). The added modules keep parameters and FLOPs close to the baseline, with only a minor frame rate (FPS) trade-off. On VisDrone2019, PAG-YOLO delivers mAP@0.5 37.6% and mAP@0.5:0.95 23.2%, improving YOLOv12n by +5.1 percentage points and +4.5 percentage points under similar to 2.7x10(6) Params and similar to 6.2 GFLOPs, outperforming several small-object-oriented peers at similar cost. On DOTA v1.0, it attains mAP@0.5 72.7% (+4.9% vs. YOLOv12n). Category-wise, ships, basketball courts, large vehicles, and harbors see the largest gains; residual gaps on slender, highly oriented classes (e.g., bridges) reflect the current HBB head's lack of rotation modeling. Qualitatively, PAG-YOLO yields tighter boxes and fewer merges in dense clusters and maintains recall under low contrast, consistent with the intended roles of PW_WConv (shallow contours), ACDSA (deep spatial focusing), and GLCAS (global-local fusion). Conclusions PAG-YOLO demonstrates that targeted, lightweight enhancements at three strategic points of the pipeline can substantially boost small-object detection in complex remote sensing scenes while preserving real-time feasibility. On VisDrone2019 and DOTA v1.0, it improves mAP@0.5 over YOLOv12n by +5.1 percentage points and +4.9 percentage points, respectively, at roughly 2.7x10(6) parameters and 6.2 GFLOPs. The ablations validate the complementary nature of PW_WConv (directional texture retention), ACDSA (coordinate-guided dynamic spatial focusing), and GLCAS (global-local attention fusion). Future work will pursue lightweight oriented heads (OBB), angle regression and matching, extended aspect-ratio priors, tiling for ultra-high resolutions, and deployment-friendly training (weak/semi-supervised) to further strengthen generalization across rotated or long-tail categories.
Keywords:
YOLOv12
small object detection
axial convolution
spatial attention mechanism
lightweight network
Journal
L
IF:
1
Papers:
505
Citations:
0