Dense Residual Dense Block Super-Resolution for Enhancing Vehicle Detection in Remote Sensing

Abstrak

Accurate vehicle detection in complex traffic environments remains a challenging task, particularly when dealing with low-resolution imagery and small-scale objects such as distant vehicles and pedestrians. Limitations in image quality, caused by hardware constraints, transmission noise, and real-time processing requirements, often lead to significant performance degradation in conventional object detection frameworks. To address these challenges, this paper proposes an integrated super-resolution–assisted vehicle detection framework that combines a transformer-based SR preprocessing module with a YOLOv12-based object detector. The proposed SR module is designed to exploit global contextual information and long-range dependencies, enabling more faithful reconstruction of fine-grained structural details that are critical for small-object recognition. The enhanced high-resolution representations are subsequently fed into the YOLOv12 detector, which offers a favorable trade-off between detection accuracy and computational efficiency across multiple object scales. Extensive experiments conducted on multiple benchmark datasets, including VEDAI-VISIBLE, VEDAI-IR, DOTA, and xView, demonstrate that the proposed approach consistently outperforms existing CNN- and GAN-based SR methods as well as baseline detection models. Quantitative and qualitative results confirm notable improvements in detection accuracy, particularly for small and densely distributed vehicles, while maintaining practical inference efficiency. These findings highlight the effectiveness of integrating transformer-based SR as a preprocessing stage for robust vehicle detection, contributing to more reliable intelligent transportation and traffic monitoring systems.

Identifier

• OpenAlex: https://openalex.org/W7161940814
• DOI: https://doi.org/10.1109/access.2026.3695534