频域内面向目标检测的领域自适应  

Domain adaptation for object detection in the frequency domain

作　　者：李岳楠 徐浩宇 董浩[2] Li Yuenan;Xu Haoyu;Dong Hao(School of Electrical&Information Engineering,Tianjin University,Tianjin 300072,China;Tianjin Jinhang Institute of Technical Physics,Tianjin 300308,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072 [2]天津津航技术物理研究所,天津300308

出　　处：《红外与激光工程》2022年第7期442-450,共9页Infrared and Laser Engineering

基　　金：国家自然科学基金(61972281,61572352)。

摘　　要：近年来,基于深度学习的目标检测技术在机器人、自动驾驶和交通监控等领域有着广泛的应用。然而,由于训练集和测试集样本分布偏差的原因,将现成的预训练检测器应用到实际开放场景时通常会出现明显性能下降。针对该问题提出了一种频域内的领域自适应方法,利用离散余弦变换的频域能量集中特性,通过在频域内对少数重要频率系数进行处理,实现了面向目标检测的领域自适应,降低了对存储和计算资源的要求并减少了领域差异。该方法可以分为两个阶段:第一阶段使用无监督图像转换方式,将源域已标注的训练数据向目标域作转换;第二阶段采用基于对抗的领域自适应方法训练目标检测模型,对转换后的训练数据与目标域内的数据作特征适配。针对不同天气场景的目标识别实验表明:所提的频域内领域自适应方法在4种领域自适应对比算法中排名第一,与仅用源域数据训练的模型相比,mAP值提升了33.9%。Deep learning-based object detection technology has recently made significant progress and has a wide range of applications in robotics,autonomous driving,traffic surveillance,etc.However,due to the distribution discrepancy between the training and testing datasets,the off-the-shelf detectors pre-trained using the data in a specific domain often show apparent performance degradation when applied in wild scenarios.To address this problem,a domain adaptation method for object detection in the frequency domain is proposed.In light of the energy concentration property of the discrete cosine transform,the proposed algorithm conducts domain adaptation for object detection by processing only a few of the most significant frequency coefficients,which reduces memory and computing resource consumption and alleviates the domain shift problem.The proposed method consists of two stages.In the first stage,it translates annotated training data from the source domain to the target domain using unsupervised image-to-image translation.Adversarial domain adaptation is then applied to the object detection model to align the features of the translated data and the real data in the target domain.The experimental results of the object detection under different weather conditions show that the proposed method ranks first among the four testing algorithms.Compared with the object detection model trained with only source domain data,it can increase the mAP value by 33.9%.

关 键 词：领域自适应 目标检测 图像转换 频域 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]