Few-shot object detection enhancement method based on SimAM attention and deformable convolution
Abstract
Few-Shot Object Detection (FSOD) aims to achieve accurate detection of novel categories using only a few annotated samples. Among existing approaches, the Two-stage Fine-tuning Approach (TFA) exhibits good generalization capability, yet its training strategy—which relies on fixed feature representations—limits the model’s adaptability to novel category features and scale variations. This paper proposes a feature adaptation enhancement framework for few-shot object detection. First, Deformable Convolutional Networks v2 (DCNv2) is introduced into the backbone network to dynamically adjust convolutional sampling positions and response weights, thereby improving the model’s ability to model geometric variations of objects. Second, a parameter-free SimAM attention module is embedded at the output stage of the Feature Pyramid Network (FPN) to adaptively recalibrate spatial and channel features, enhancing multi-scale feature representation. Furthermore, to address the issue of frozen multi-scale features during TFA fine-tuning, a lightweight adaptation strategy that only unfreezes the FPN layers is proposed, enabling the model to adjust the feature fusion process with limited novel class samples. Experimental results demonstrate that the proposed method outperforms the baseline model across various settings from 1-shot to 10-shot, achieving a 3.2% mAP improvement over the TFA baseline under the 1-shotcondition. This verifies the effectiveness of our method in few-shot object detection tasks.