A Lightweight Hybrid Template-Matching–CNN Framework with Attention-Guided Fusion for Robust Small Object Detection

Small object detection in aerial and surveillance imagery remains challenging due to low resolution, occlusion, and background clutter. This study introduces a novel hybrid detection framework that fuses template matching with a deep learning detector (Faster R-CNN) through an attention-guided decision fusion mechanism. The novelty lies in (i) a dual-stage fusion pipeline that integrates precise structural cues from template matching with deep semantic features, and (ii) a custom scale-aware focal loss, adapted from Focal Loss to emphasize hard and small objects by dynamically increasing penalties for low-confidence predictions. Evaluated on a Pascal VOC subset (1000 images, 5 classes), the proposed system achieves an mAP improvement of 3.5% over the Faster R-CNN baseline and surpasses YOLO-Lite and R-CNN variants in precision and recall. The hybrid design adds only a minimal computational overhead (0.45 s/image vs. 0.42 s for Faster R-CNN), demonstrating favorable efficiency–accuracy trade-offs suitable for scalable deployment. These findings highlight the framework’s robustness, particularly in scenes containing occlusion, clutter, or visually small targets. Limitations regarding template dependency are discussed, along with future directions for automatic template generation and real-time video adaptation.