基于未知类别少样本学习的开放世界目标定位  

Open World Object Localization Based on Few Shot Learning of Unknown Categories

作　　者：王三泉 王璐[1] 储珺[1] 黄斌 WANG San-quan;WANG Lu;CHU Jun;HUANG Bin(School of Software Engineering,Nanchang Hangkong University,Nanchang 330063,China)

机构地区：[1]南昌航空大学软件学院,南昌330063

出　　处：《南昌航空大学学报（自然科学版）》2024年第3期33-44,71,共13页Journal of Nanchang Hangkong University(Natural Sciences)

基　　金：国家自然科学基金(62162045)。

摘　　要：开放世界目标识别中未知类别样本难以获取,这些样本无法有效参与训练,针对这个问题,提出基于未知类别少样本学习的开放世界目标定位方法。首先,利用伪标签生成模型,将背景区域中高目标性得分的目标区域标记为伪标签,为模型学习未知类提供少量样本,增强模型对未知类别的泛化能力;其次,设计两个分支共同学习目标特征,其一为目标性分支,其二为定位质量分支。前者专注于确定目标框内是否存在目标,后者则集中于评估目标框的质量,从目标存在性和目标定位质量两个维度提取和理解目标特征。同时,目标性分支联合训练真实标签与伪标签样本增强对未知类别的识别能力,定位质量分支通过学习高质量的真实标签样本并排除伪标签样本,以减少噪声干扰。最后在COCO数据集上对该方法进行有效性评估,结果显示,与其他方法相比,所提方法具有较为出色的检测性能。Unknown category samples are difficult to obtain in open world object recognition,and these samples cannot effectively participate in training.To address this issue,this study proposes an open-world object localization method based on few-shot learning for unknown categories.Initially,a pseudo-label generation model is employed to label regions with high objectness scores within background areas as pseudo-labels,providing a small number of samples for the model to learn unknown classes and enhancing its generalization ability towards unseen categories.Subsequently,two branches are designed to jointly learn object features:one focuses on determining the presence of an object within the bounding box(objectness branch),while the other concentrates on assessing the quality of the bounding box(localization quality branch).These branches extract and comprehend object features from two dimensions,i.e.object existence and localization quality.The objectness branch is trained jointly with both real and pseudo-label samples to enhance recognition of unknown categories,whereas the localization quality branch learns from high-quality real labels and discards pseudo-labels to minimize noise interference.An effective evaluation of this method on the COCO dataset demonstrates that it has excellent detection performance compared to other methods.

关 键 词：开放世界 目标检测 目标定位 伪标签 定位质量 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]