【作者】 刘杰；

【导师】 罗芳；

【作者基本信息】 武汉理工大学 ， 软件工程， 2024， 硕士

【摘要】 通常深度学习模型假设训练集和测试集来自相同的分布,然而在实际的目标检测应用场景中,由于天气、照明、图像风格等多种因素的存在,可能会导致训练数据和测试数据的特征分布存在差异,不满足独立同分布条件,即使训练数据和测试数据所需检测的类别一致,也会因特征分布差异而出现域偏移现象,致使目标检测模型存在漏检和误检问题。为了解决此问题,本文提出一种基于多重域随机化的无监督域自适应跨域目标检测方法,主要研究内容如下:(1)针对目前基于生成式对抗的像素级无监督域自适应方法生成的合成域与目标域的拟合程度不足问题,本文提出了多重域随机化方法。首先,为缓解源域和目标域的域偏移,本文基于Cycle-GAN、CUT两种不同的非配对图像翻译模型实现多重域随机化,得到与目标域特征分布更为拟合的随机化域,使得基于多重域随机化的跨域目标检测模型能更有效地学习目标域的特征分布,提高模型的泛化性能和跨域检测能力;接着,对两种图像翻译模型得到的多重随机化域进行实验对比,分析多重随机化域与目标域的特征分布相似程度;最后,综合实验数据,确定最优的多重域随机化实现策略,为后续无监督域自适应跨域目标检测提供基础。(2)针对使用多域训练目标检测模型的过程中存在的多域训练困难、域偏向性、特征标签不一致、收敛速度慢等问题,本文利用经典两阶段目标检测模型Faster R-CNN构建了基于多重域随机化的无监督域自适应跨域目标检测模型MDR-Faster R-CNN,并利用多域域不变特征提取网络来解决多域训练困难问题;同时,设计了源域一致性损失函数来解决域偏向性及特征标签不一致问题;此外,设计了域随机化参数回调模块来提高模型训练的收敛速度。通过在多组跨域目标检测数据集上进行性能对比分析,验证了基于多重域随机化的跨域目标检测模型MDR-Faster R-CNN在跨域检测任务中的有效性。(3)针对目前无监督域自适应方法因依赖目标检测模型而缺乏灵活性的问题,本文进行了无监督域自适应方法的可移植性研究,使得无监督域自适应方法不依赖具体的目标检测模型,仅需少量修改便可对不同目标检测模型实现无监督域自适应,以提高无监督域自适应跨域目标检测方法的普适性;接着,将本文设计的基于多重域随机化的无监督域自适应跨域目标检测算法分别移植到单阶段的目标检测模型YOLOv7和基于Transformer的端到端目标检测模型可变形DETR中,构建MDR-YOLOv7和MDR-DefDETR模型;最后通过多组跨域目标检测任务验证了本文方法具有良好的可移植性。更多还原

【Abstract】 Usually,deep learning models assume that the training and testing datasets come from the same distribution.However,in practical object detection application scenarios,due to various factors such as weather,lighting,image style,etc.,there are differences in the feature distribution between the training and testing data,which do not meet the independent and identically distributed conditions.Even if the categories to be detected by the training and testing data are the same,domain shift may occur due to the differences in feature distribution,resulting in missed detections and false detections in the object detection model.To address this issue,this thesis proposes an unsupervised domain adaptive object detection method based on multiple domain randomization.The main research content is as follows:(1)This thesis proposes a multiple domain randomization(MDR)method to address the issue of insufficient fitting between the synthesized domain and the target domain generated by pixel level unsupervised domain adaptive methods based on generative adversarial methods.Firstly,to alleviate the domain shift between the source domain and the target domain,multiple domain randomization is achieved based on two different unpaired image translation models,Cycle-GAN and CUT,to obtain a more fitting randomization domain with the target domain feature distribution.This enables the object detection model based on multiple randomization domains to learn the target domain feature distribution more effectively,improving the model’s generalization performance and cross domain detection ability;Next,experimental comparisons were conducted on the multiple randomized domains obtained from two image translation models,analyzing the similarity of feature distributions between the multiple randomized domains and the target domain.The optimal implementation strategy for multiple domain randomization was designed to provide a foundation for subsequent unsupervised domain adaptive object detection.(2)In response to the difficulties of multi domain training,domain bias,inconsistent data labels,and slow convergence speed in using multi domain training object detection models,this thesis constructs an unsupervised domain adaptive object detection model MDR-Faster R-CNN based on multiple domain randomization using the classic two-stage object detection model Faster R-CNN,and uses a multi domain invariant feature learning network to solve the problem of multi domain training difficulties;Meanwhile,a source domain consistency loss function was designed to address domain bias and inconsistent data labels;In addition,a domain randomization parameter callback module was designed to improve the convergence speed of model training.The effectiveness of the unsupervised object detection model MDR-Faster R-CNN based on multiple domain randomization was verified through performance comparison analysis on several cross-domain object detection datasets.(3)In response to the lack of flexibility in unsupervised domain adaptation methods due to their reliance on object detection models,this thesis has conducted a recent study on the transplantability of unsupervised domain adaptation methods.This method does not rely on specific object detection models.With only a small amount of modification,unsupervised domain adaptation can be achieved for new object detection model to improve the universality of unsupervised domain adaptive cross-domain object detection;Next,the unsupervised domain adaptive object detection algorithm based on multiple domain randomization designed will be transplanted to the single-stage object detection model YOLOv7 and the Transformer based end-to-end object detection model deformable DETR,respectively,to construct the MDR-YOLOv7 and MDR-DefDETR models;Finally,the good transplantability of our method was verified through several sets of cross domain object detection tasks.更多还原