Few-Shot Object Detection via Multi-Scale Channel Aggregation and Contrastive Prototype Learning.

Saved in:

Bibliographic Details
Title:	Few-Shot Object Detection via Multi-Scale Channel Aggregation and Contrastive Prototype Learning.
Authors:	Sun, Huadong¹ sunhuadong1209@163.com, Tong, Junwei² 18591625641@163.com, Zhao, Yexuan² dawsonjason2623@hotmail.com
Source:	IAENG International Journal of Computer Science. Jul2026, Vol. 53 Issue 7, p2866-2881. 16p.
Subjects:	Contrastive learning, Object recognition (Computer vision), Feature extraction
Abstract:	Few-shot object detection (FSOD) aims to recognize novel object categories with extremely limited annotated samples, where fine-tuning-based two-stage detectors often suffer from unstable prototype representations and insufficient feature discrimination. In particular, the mismatch of feature scales between support and query images, together with strong interference between base and novel classes, significantly limits the effectiveness of existing decoupling-based methods under low-shot settings. To address these challenges, we propose a robust two-stage FSOD framework that enhances both feature representation and category separation in a structured manner. Specifically, a Multi-Scale Channel Aggregation (MCA) module is introduced to integrate hierarchical local and global features, thereby improving cross-scale semantic consistency for object detection. In addition, a Contrastive Prototypical Feature Separation (CPFS) strategy is designed to reduce base-novel interference by jointly promoting intra-class compactness and inter-class separability in the prototypical embedding space, thereby enlarging class margins and stabilizing prototype estimation under limited supervision. Extensive experiments conducted on the PASCAL VOC and MS COCO benchmarks under multiple few-shot settings demonstrate that the proposed approach achieves competitive or superior performance across different novel class splits. Furthermore, statistical evaluation with repeated random sampling under generalized few-shot settings verifies improved robustness and reduced performance variance. These results indicate that the proposed framework provides a more reliable and discriminative solution for fewshot object detection. [ABSTRACT FROM AUTHOR]
	Copyright of IAENG International Journal of Computer Science is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database:	Engineering Source

Description
Abstract:	Few-shot object detection (FSOD) aims to recognize novel object categories with extremely limited annotated samples, where fine-tuning-based two-stage detectors often suffer from unstable prototype representations and insufficient feature discrimination. In particular, the mismatch of feature scales between support and query images, together with strong interference between base and novel classes, significantly limits the effectiveness of existing decoupling-based methods under low-shot settings. To address these challenges, we propose a robust two-stage FSOD framework that enhances both feature representation and category separation in a structured manner. Specifically, a Multi-Scale Channel Aggregation (MCA) module is introduced to integrate hierarchical local and global features, thereby improving cross-scale semantic consistency for object detection. In addition, a Contrastive Prototypical Feature Separation (CPFS) strategy is designed to reduce base-novel interference by jointly promoting intra-class compactness and inter-class separability in the prototypical embedding space, thereby enlarging class margins and stabilizing prototype estimation under limited supervision. Extensive experiments conducted on the PASCAL VOC and MS COCO benchmarks under multiple few-shot settings demonstrate that the proposed approach achieves competitive or superior performance across different novel class splits. Furthermore, statistical evaluation with repeated random sampling under generalized few-shot settings verifies improved robustness and reduced performance variance. These results indicate that the proposed framework provides a more reliable and discriminative solution for fewshot object detection. [ABSTRACT FROM AUTHOR]
ISSN:	1819656X