知识嵌入引导的双分支融合增强开放词汇 目标检测

打开文本图片集

关键词：开放词汇目标检测；知识嵌入；标签匹配；双分支融合中图分类号：TP391.41 文献标识码：Adoi：10.37188/OPE.20253318.2929 CSTR：32169.14.OPE.20253318.2929

Abstract： To address the issues of weak understanding of new classconcepts，label confusion，and insuficient detection performance of new classes in open-set scenarios，a Knowledge Integration-guided Dualbranch Fusion Open-Vocabulary Object Detection （KI-DBFOVD） method was proposed in this paper. Firstly，a Knowledge Integration （KI） module was designed，where pseudo-labels generated by a VisionLanguage Model were embedded into the detector to learn about new class concepts. Subsequently，a Label Match （LM） module was introduced to refine the label matching process through multi-level threshold adjustment and independent matching between base and new classes，thereby aleviating the label confusion between base and new classes during detection. Finaly，a novel Dual-branch Fusion module（DBF） was constructed by fusing the traditional visual branch and the vision-language branch via geometric averaging.This fusion maintained the detection accuracy of base classes and more effectively detected and localized new class objects，then enhanced the overall detection performance of the KI-DBFOVD method. Ex- perimental results demonstrate that this method achieves a detection accuracy of 38.6% for new classes on the COCO dataset and 25.4% on the more challenging LVIS dataset，which contains a larger number of categories. These results outperform several mainstream methods and indicate that this approach is more suitable for different open-set scenarios..

Key words：open-vocabulary object detection； knowledge integration；label match；dual-branch fusion

1引言

目标检测是计算机视觉的一项基础任务，已广泛应用于环境资源监测[1]、交通安全监控[2]、工业产品检测[3等领域。（剩余21591字）