FSDN-DETR: Enhancing Fuzzy Systems Adapter with DeNoising Anchor Boxes for Transfer Learning in Small Object Detection

Abstract

The advancement of Transformer models in computer vision has rapidly spurred numerous Transformer-based object detection approaches, such as DEtection TRansformer. Although DETR’s self-attention mechanism effectively captures the global context, it struggles with fine-grained detail detection, limiting its efficacy in small object detection where noise can easily obscure or confuse small targets. To address these issues, we propose Fuzzy System DNN-DETR involving two key modules: Fuzzy Adapter Transformer Encoder and Fuzzy Denoising Transformer Decoder. The fuzzy Adapter Transformer Encoder utilizes adaptive fuzzy membership functions and rule-based smoothing to preserve critical details, such as edges and textures, while mitigating the loss of fine details in global feature processing. Meanwhile, the Fuzzy Denoising Transformer Decoder effectively reduces noise interference and enhances fine-grained feature capture, eliminating redundant computations in irrelevant regions. This approach achieves a balance between computational efficiency for medium-resolution images and the accuracy required for small object detection. Our architecture also employs adapter modules to reduce re-training costs, and a two-stage fine-tuning strategy adapts fuzzy modules to specific domains before harmonizing the model with task-specific adjustments. Experiments on the COCO and AI-TOD-V2 datasets show that FSDN-DETR achieves an approximately 20% improvement in average precision for very small objects, surpassing state-of-the-art models and demonstrating robustness and reliability for small object detection in complex environments.

Keywords

• object detection
• transformer
• transfer learning
• DEtection TRansformer
• fuzzy system
• adapter