摘要
目标检测系统的测试数据生成对评估模型性能和发现潜在缺陷至关重要。现有方法在生成数据的多样性和真实性方面仍存在局限。该文提出了一种基于自动语义编辑的目标检测测试数据生成方法SemaGen,通过构建高质量语义对象库并结合自动化语义编辑策略,实现对图像的插入、删除和替换等高级语义操作。具体而言,该方法首先通过多重筛选机制构建语义对象库,确保对象的语义完整性和场景适应性;其次,利用场景复杂度量化模型,综合考虑背景占比、实例数量和空间分布等因素,实现编辑策略的自适应选择;最后,提出基于对象重要性的替换策略、迭代式删除方法以及考虑语义相似度的智能插入机制,确保生成图像的真实性和多样性。实验结果表明,SemaGen在三种对象操作任务上显著优于现有方法,生成的图像质量更高,FID得分更优,证实了该方法在生成数据质量方面的优越性。在目标检测模型测试中,SemaGen成功发现YOLO v11、SSD和Mask R-CNN等主流检测器在复杂场景下的性能缺陷,为目标检测测试用例生成提供了新的思路和工具。
Test data generation for object detection systems is crucial for evaluating model performance and identifying potential defects.Existing methods still have limitations in generating diverse and realistic data.We present SemaGen,a test data generation method for object detection based on automated semantic editing,which achieves advanced semantic operations such as insertion,deletion,and replacement through constructing high-quality semantic object libraries and combining automated editing strategies.Specifically,the proposed method first constructs a semantic object library through multiple screening mechanisms to ensure object semantic integrity and scene adaptability.Secondly,it utilizes a scene complexity quantification model that comprehensively considers background ratio,instance quantity,and spatial distribution to achieve adaptive selection of editing strategies.Finally,it proposes an object importance-based replacement strategy,an iterative deletion method,and an intelligent insertion mechanism considering semantic similarity to ensure the authenticity and diversity of generated images.The experimental results show that SemaGen significantly outperforms the existing methods on the three object manipulation tasks,generates higher quality images with better FID scores,and confirms its superiority in generating data quality.In object detection model testing,SemaGen successfully identifies performance deficiencies of mainstream detectors such as YOLO v11,SSD,and Mask R-CNN in complex scenarios,providing new insights and tools for generating object detection test cases.
作者
陈皓明
桂智明
刘艳芳
范鑫鑫
路云峰
CHEN Hao-ming;GUI Zhi-ming;LIU Yan-fang;FAN Xin-xin;LU Yun-feng(School of Computer Science and Technology,Beijing University of Technology,Beijing 100124,China;School of Computer Science and Engineering,Beihang University,Beijing 100083,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China;School of Reliability and Systems Engineering,Beihang University,Beijing 100088,China)
出处
《计算机技术与发展》
2025年第7期16-23,共8页
Computer Technology and Development
基金
复杂关键软件环境全国重点实验室自主课题(SKLSDE-2023ZX-17)。
关键词
目标检测
语义编辑
测试数据生成
深度神经网络
图像生成
object detection
semantic editing
test data generation
deep neural networks
image generation
