Selective Large‐Kernel Modelling and Fine‐Grained Representation Learning for Efficient Aerial Object Detection
Year: 2026
Type: article
Abstract: ABSTRACT The prevalence of small objects and complex backgrounds presents considerable challenges for efficient object detection in aerial images. To address these issues, we propose AE‐DFINE, an enhanced framework based on D‐FINE‐S that improves detection performance through fine‐grained representation learning and contextual feature modelling. Specifically, a split interaction convolution (SI‐Conv) module is introduced into the backbone to strengthen fine‐grained feature extraction and improve the representation of small objects. In addition, a selective large Kernel module (SLKM) is incorporated into the hybrid encoder to perform lightweight large‐kernel modelling, enhancing informative contextual features while reducing interference from redundant background regions with limited computational overhead. By jointly leveraging SI‐Conv and SLKM, AE‐DFINE achieves a favourable balance between detection accuracy and computational efficiency. Extensive experiments conducted on the VisDrone and AI‐TOD datasets demonstrate that AE‐DFINE consistently outperforms D‐FINE‐S in terms of AP and while maintaining competitive computational efficiency.
Source: IET Image Processing
DOI: 10.1049/ipr2.70475
Institution: Dazhou Vocational and Technical College
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,740
OpenAlex ID: W7212373280