关键词: 小样本目标检测, 基于度量, 区域生成网络, 非线性分类器, 特征聚合

Abstract: Abstract: In existing few-shot object detection, the Region Proposal Network (RPN) is usually trained on base class data to generate new class candidate boxes. However, new class data is relatively scarce compared to the base class. Introducing new classes may lead to the presence of different complex backgrounds, causing the RPN to misclassify background as foreground, resulting in the omission of high Intersection over Union (IoU) value candidate boxes. To address the above issues, a few-shot object detection method that based on improved region proposal network and feature aggregation was proposed, namely IFA-FSOD (Few-Shot Object Detection Based on Improved RPN and Feature Aggregation) . Firstly, an improvement was made based on the RPN by incorporating a metric-based non-linear classifier. This classifier was designed within the RPN to compute the similarity between features extracted by the backbone network and the features representing the new class. This enhancement aims to increase the recall rate for candidate boxes of the new class, ultimately filtering out candidate boxes with IoU. Then, a feature aggregation module based on the attention mechanism was introduced in the region of interest alignment (ROI Align). By designing grids of different scales, more comprehensive information and feature representation were obtained, and alleviate the lack of feature information caused by different scales. Experimental results show that compared with the QA-FewDet method, IFA-FSOD improves nAP50 by 4.5percentage points under Novel Set 3 10shot on the new class of PASCAL-VOC dataset; compared with the FsDetView method, under the settings of 10 shot and 30 shot, IFA-FSOD's mAP increased by 0.2 and 0.8 percentage points respectively on the new class of the COCO dataset. IFA can effectively improve the detection performance of target categories in the case of few-shot, and solve the problem of missing candidate boxes with high IoU values and incomplete feature information capture.

Key words: few-shot object detection, metric-based, Region Proposal Network(RPN), nonlinear classifier, feature aggregation