Feature-Enhanced Faster R-CNN for Small Object Detection in Complex Scenes

This study proposes a feature-enhanced Faster R-CNN algorithm to address the challenges of dense targets, varying object shapes, and complex backgrounds in remote sensing image detection. First, the single-stage model YOLOv8 serves as a baseline for comparative evaluation, revealing the inherent advantage of two-stage detectors in detection accuracy. In scenarios with high-density small objects, the original Faster R-CNN achieves significantly higher accuracy than YOLOv8, though YOLOv8 offers faster inference speed. Based on this observation, the proposed method introduces targeted enhancements to address three key limitations of the original Faster R-CNN: To enhance feature extraction capability, the method integrates Inverted Residual Blocks and Depthwise Separable Convolutions into the ResNet50 backbone. This design strengthens low-dimensional feature representation through a three-phase process of channel expansion, feature extraction, and channel compression. The ReLU6 activation function further improves model robustness; To address insufficient multi-scale feature fusion, the method develops a densely connected feature pyramid network (Dense-FPN). Cross-layer skip connections enable feature maps from different backbone stages to interact densely with each FPN level, facilitating comprehensive fusion of shallow spatial details and deep semantic information; To improve localization accuracy, the method employs a cascaded detection head with a progressive IoU threshold adjustment strategy, refining the bounding box regression stepwise. Ablation studies confirm the effectiveness of each proposed module individually and in combination. Extensive experiments on the DOTA dataset show that the improved algorithm achieves a mean Average Precision (mAP) of 78.2%, representing a 5.3 percentage point increase over the original Faster R-CNN (72.9%). These results demonstrate that the proposed approach effectively resolves critical issues, including feature degradation, semantic inadequacy, and localization bias in remote sensing object detection.