Interactive Selective Spatial Feature Extractor for Small Object Detection in Challenging Scenes

Abstract

Small object detection in computer vision has made great strides in accuracy and robustness. However, practical applications are still hampered by two main issues: inaccurate detection of small objects and the difficulty of deploying these models on resource- constrained devices due to their extensive parameters and high computational demands. To address these limitations, we propose the Interactive Selective Spatial Feature Extraction method. Which is designed for efficient small object detection, utilizing a modified C2f-Darknet backbone for robust feature extraction and an attention mechanism to focus on crucial spatial areas. The core innovation of this technique lies in its ability to extract and combine multi-scale information to capture fine-grained details and border content, and an internal interaction mechanism that allows communication and information sharing to selectively attend to the most relevant features while suppressing noise. Proposed a state-of-art system that is evaluated on a publicly available DOTA dataset and VisDrone dataset. The proposed methodology detects the accurate region.