软件导刊2025，Vol.24Issue(5)：200-206,7.DOI:10.11907/rjdk.241203
基于改进CenterNet少样本条件的增量化目标检测算法
An Incremental Object Detection Algorithm Based on Enhanced CenterNet under Few-shot Scenarios
摘要
Abstract
Deep learning models require a large number of annotated samples for training,but in practical scenarios,it is often difficult to ob-tain large-scale and high-quality annotated samples,and not having enough training samples can lead to poor model performance.At present,object detection algorithms targeting few sample conditions often suffer from missed and false detections due to insufficient training samples,which cannot effectively generate detection boxes for the target.To address the above issues,a few sample condition oriented incremental ob-ject detection algorithm is proposed based on the ONCE network model using the CenterNet method as a framework.The detector mainly con-sists of a feature extractor that combines attention mechanism and feature fusion,a decomposed attention encoder,an adaptive Gaussian ker-nel,and a locator.It adopts a two-stage training strategy and registers new classes through an incremental strategy.Among them,the feature extractor that combines attention mechanism and feature fusion can enhance the ability to extract and utilize features of base classes and new classes;The adaptive Gaussian kernel and decomposition attention method can enable the detector to obtain more accurate target center points,thereby making the generated target boxes more precise.The validation experiments on the COCO dataset and PASCAL VOC dataset showed that compared with other advanced methods,the proposed method improved the AP values of the new class,base class,and full class by an average of 7.7,2.5,and 1.3,respectively,and the AR values by an average of 15.0,21.7,and 17.6,respectively,on the COCO datas-et compared to the ONCE model;On the PASCAL VOC dataset,the average AP and AR values were improved by 2.4 and 21.2,respectively,compared to other methods,indicating better detection performance.关键词
注意力机制/特征融合/自适应高斯核/分解注意力Key words
self-attention mechanism/feature fusion/adaptive Gaussian kernel/decomposed attention分类
计算机与自动化引用本文复制引用
李航,贾可,金治成,李涵鑫,许昌源..基于改进CenterNet少样本条件的增量化目标检测算法[J].软件导刊,2025,24(5):200-206,7.基金项目
四川省科技计划项目(2023YFG0305) （2023YFG0305）
成都信息工程大学科研基金项目(KYTZ202156) （KYTZ202156）
