MI-DETR: A small object detection model for mixed scenes

Abstract

Object detection is a fundamental task in computer vision with many applications. However, the complexity of real-world scenarios and the small size of objects present significant challenges to achieving rapid and accurate detection in mixed environments. To address these challenges, we propose a novel mixed scene-oriented small object detection model (MI-DETR) designed to enhance detection accuracy and efficiency. The backbone network of MI-DETR incorporates the Fast Fourier Transform, Channel Shuffle, and Orthogonal Attention Mechanism to improve feature extraction while reducing computational costs significantly. Additionally, we introduce a specialized small object feature layer and a Multi-Scale Feature Fusion (MSFF) module to strengthen the model’s feature fusion capabilities. Furthermore, we propose a novel loss function, Focaler-WIoU, which prioritizes high-quality anchor frames to enhance the detector’s performance. We validate the effectiveness of our model through experiments on small object detection datasets across three complex scenarios. The results show that the proposed MI-DETR model has 40% fewer parameters and 5% less computational effort than the former. On these three datasets, MI-DETR achieves accuracies of 70.2%, 34.5%, and 34.1%, respectively. It also achieves small object detection accuracies of 19.8%, 11.5%, and 12.6%, respectively. Additionally, its latency decreases by 0.9 ms, 1.0 ms, and 1.1 ms, respectively, outperforming other real-time detection models of similar size.