LCANet: A Lightweight Faster R-CNN Framework with Selective P2 Coordinate Attention for Tiny Object Detection in Aerial Images

Abstract

Accurate detection of tiny objects in aerial images remains a challenging problem in remote sensing due to extreme scale variations, complex backgrounds, dense object distributions, and the limited spatial representation of small targets. Although modern deep learning-based object detectors achieve remarkable detection accuracy, many rely on computationally intensive backbone networks and complex attention mechanisms that restrict their deployment on resource-constrained platforms such as unmanned aerial vehicles (UAVs) and edge devices. Lightweight detection frameworks reduce computational complexity but often suffer from degraded localization performance for tiny objects because high-resolution spatial features are gradually lost during feature extraction.
To address this limitation, this paper proposes a lightweight object detection framework that selectively integrates Coordinate Attention (CA) into the highest-resolution feature pyramid level (P2) of a Faster R-CNN architecture employing a ResNet-18 backbone and a Feature Pyramid Network (FPN). Unlike conventional approaches that introduce attention modules across multiple feature levels, the proposed method enhances only the P2 feature map, where the fine spatial detail required for tiny-object localization is best preserved, by factoring channel attention into two 1D feature-encoding vectors along the horizontal and vertical directions. This selective integration strengthens directional and positional feature representation while preserving the computational efficiency of the lightweight detector, and it leaves the Region Proposal Network (RPN), RoIAlign, and Faster R-CNN detection head entirely unchanged.
The proposed framework is evaluated on the DOTA v2.0 aerial-image dataset using a structured patch-based partition of 8,734 training patches and 1,000 validation patches under Task 2 (horizontal bounding box) settings. The baseline ResNet-18-FPN Faster R-CNN model converges smoothly over a 10-epoch schedule, reaching a final training loss of 0.5775 and a validation mAP@50 of 73.71%. Integrating the proposed P2-selective Coordinate Attention module reduces the final training loss to 0.3782 and raises the validation mAP@50 to 86.20%, an absolute improvement of 12.49 percentage points, while training and validation loss curves demonstrate faster convergence and improved optimization stability throughout training. Qualitative evaluation on high-density and ultra-dense test scenes further confirms that the proposed model recovers small, densely packed targets missed by the baseline.
These findings indicate that selectively enhancing the highest-resolution feature representation is an effective and computationally efficient strategy for improving tiny-object localization without modifying the Faster R-CNN detection pipeline. The proposed framework therefore provides a practical solution for lightweight aerial object detection in real-world remote sensing and UAV applications.