A Blur-Score-Guided Region Selection Method for Airborne Aircraft Detection in Remote Sensing Images

Airborne aircraft detection is of paramount importance for optimizing airspace management and enhancing flight safety and efficiency in both commercial and private sectors. High-speed airborne aircraft (over 800 km/h) often introduce motion blur and diminish the semantic correlation between the aircraft and its background. Conventional methods for stationary aircraft detection are inadequate for addressing these challenges. To overcome these issues, we propose BS-DETR, a novel transformer-based object detection model for airborne aircraft in remote sensing images. Our approach includes an improved tenengrad gradient algorithm to extract motion blur information and construct a Blur-Score map. We also introduce an adaptive feature fusion mechanism to integrate the Blur-Score map with multiscale features. In addition, an aircraft region selector (ARS) is employed to identify regions with a high probability of containing aircraft, thereby eliminating irrelevant background. We have established a comprehensive airborne aircraft dataset, including diverse aircraft models, cloud formations, and aircraft contrails. Experimental results on this dataset demonstrate that BS-DETR outperforms other state-of-the-art object detectors, highlighting the effectiveness of incorporating Blur-Score maps, and removing ineffective backgrounds.