FDOD: A Novel Few-Shot Small Object Detection Framework Using Fractional Fourier Transform and Differentiation

Abstract

Few-shot small object detection poses a pivotal challenge in enhancing object recognition accuracy for autonomous driving. Limited by the scarcity of training samples, it struggles with inadequate semantic features and generalization capabilities across scales and domains. To address this, this paper introduces an efficient and lightweight Fractional Fourier Transform (FrFT) and Fractional Differentiation (FD) optimization for small object detection, abbreviated as FDOD, comprising four modules: meta-learning-based feature extraction, FrFT and FD optimization-based feature enhancement, alignment fine-tuning, and small object prediction. Firstly, a shared backbone network is employed to extract features from support and query samples; Secondly, the FrFT and FD optimization-based module is designed to decompose object features into phase and amplitude components, thereby enhancing the semantic representation of low-resolution small objects, and an FD optimization problem is formulated to dynamically optimize frequency-domain features, the combination of the two enables amplify non local textures and enhance edges in frequency bands for small object detection with few samples, and improving cross scale generalization ability; Thirdly, an adaptive attention-guided feature alignment strategy is adopted to refine the inter-class feature discrimination between support and query samples; Finally, an object bounding box is generated through the prediction head. We conducted extensive experiments on public datasets to validate FDOD. The experimental results show that its mean accuracy precision is improved by 11.7% and 8.6%, respectively, compared to the current leading model on PASCALVOC and MS-COCO, especially on small object samples such as pedestrians and bicycles, where the accuracy is improved significantly. At the same time, it demonstrated an efficient adaptive capability on KITTI, with a computational efficiency of $\mathrm {1.33\times }$ that of the current leading model.