Paper #1

Efficient Object Detection in Remote Sensing Images Using Quantitative Augmentation and Competitive Learning

Huaxiang Song, Junping Xie, Yan Zhang, Yang Zhou, Wenhui Wang, YingYing Duan and Xinyi Xie

Abstract: Object detection in remote sensing images (RSIs) is crucial in Earth observation. However, current approaches often overlook key characteristics of RSIs, resulting in models that fail to balance accuracy and computational efficiency. To the authors’ knowledge, these limitations stem from the inherent scarcity and complexity of RSI samples, which cannot be fully resolved by solely modifying the model architecture. To address these challenges, we propose QACL-Net, an object detection method built on the Faster R-CNN framework, which significantly enhances the performance of CNN-based detectors for RSI recognition while maintaining fast inference speeds. QACL-Net incorporates several innovative techniques. Firstly, we introduce the quantitative augmentation (QA) strategy to address RSI sample scarcity. Secondly, we propose the equal-quadrate mosaic (EQM) algorithm to improve the effectiveness of the traditional mosaic technique for RSI detection. Thirdly, we implement the competitive learning (CL) strategy to resolve the problem of redundant feature fusion in the feature pyramid network. Crucially, the proposed enhancement techniques are integrated into three plug-and-play modules. To evaluate the proposed method, we develop two variants of QACL-Net by utilizing an EfficientNet-B0 and EfficientNet-B3 backbone model for the detector architecture, respectively. Extensive experiments on two widely used RSI datasets demonstrate that QACL-Net outperforms 31 advanced methods since 2022 on the DIOR20 dataset. Specifically, QACL-Net-B3 achieves a 6.9% improvement in accuracy on the challenging DIOR20 dataset. Additionally, QACL-Net-B3 reduces model size by 33% and increases inference speed by 17% compared to the baseline model. In summary, our work highlights the significant impact of RSI sample scarcity, noisy backgrounds, and feature fusion redundancy on object detection performance. Theoretically, our approach can be seamlessly integrated with other detection models, as the QA, EQM, and CL modules require only minimal modifications to the model structure.

Keywords: Competitive Learning; Equal-Quadrate Mosaic; QACL-Net; Quantitative Augmentation; Remote Sensing Object Detection.