结合循环提取器与自蒸馏的目标检测方法

来源：计算机技术与发展杂志2024年第04期北京时间：

作者：仲林伟;陈丹伟

单位：南京邮电大学 计算机学院,江苏 南京 210003 Author(s): ZHONG Lin-wei;CHEN Dan-wei School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210003,China 关键词: 深度学习;目标检测;特殊天气;循环解纠缠;自蒸馏;域不变特征 Keywords: deep learning;object detection;adverse weather;cyclic-disentangled;self-distillation;domain-invariant representations 分类号: TP391. 41 DOI: 10. 3969 / j. issn. 1673-629X. 2024. 04. 011 摘要: 在深度学习时代,目标检测方法不断发展,且在良好的视觉环境中已经具有较高的水平。 但是,在特殊天气下,常规的目标检测方法的检测性能大幅下降,甚至失效,而特殊天气环境的行车安全一直是社会广泛关注的问题。 为了解决上述问题,该文主要设计了一个目标检测器的模型,即引入循环解纠缠、自蒸馏方法的改进 YOLO 模型。 在循环解纠缠模块,从输入图像中循环提取域不变特征,通过循环操作,可以在不依赖域相关注释的情况下,提高图像域特征和域不变特征的解缠能力;在自蒸馏模块,以提取到的域不变特征为教师对象,进一步提高泛化能力。 并且该检测器在只有一个源域进行训练的情况下,面对许多未曾训练过的目标域上仍然表现良好,提高了检测器在未知域的鲁棒性。 实验验证了模型在各种天气下城市场景目标检测的效果,实验数据表明,该方法优于基线约 8 百分点,相比基线方法获得了性能提升。 Abstract: In the era of deep learning,object detection methods are constantly developing and have reached a high level in a good visualenvironment. However,the detection performance?of conventional target detection methods in adverse weather conditions has significantlydecreased or even failed,and driving safety in adverse weather environments has always been a widespread concern in society. In order tosolve the above problems,we mainly design a model of the target detector,which is an improved YOLO model that introduces cyclic-disentanglement and self-distillation methods. In the cyclic disentanglement module,domain invariant features are extracted from the inputimage in a cyclic manner. Through cyclic operations,the ability to unwrap image domain features and domain invariant features can beimproved without relying on domain related annotations; in the self-distillation module,the extracted domain invariant features are usedas the teacher’s object to further improve generalization ability. Moreover,the detector performs well in many untrained target domainseven when trained in only one source domain, improving its robustness in the unknown domain. The experiment verifies the effectivenessof the model in detecting urban scene targets under various weather conditions. The experimental data show that the proposed methodoutperforms the baseline by about 8 percentage points and achieves performance improvement compared to the baseline method.

摘要：在深度学习时代,目标检测方法不断发展,且在良好的视觉环境中已经具有较高的水平。 但是,在特殊天气下,常规的目标检测方法的检测性能大幅下降,甚至失效,而特殊天气环境的行车安全一直是社会广泛关注的问题。 为了解决上述问题,该文主要设计了一个目标检测器的模型,即引入循环解纠缠、自蒸馏方法的改进 YOLO 模型。 在循环解纠缠模块,从输入图像中循环提取域不变特征,通过循环操作,可以在不依赖域相关注释的情况下,提高图像域特征和域不变特征的解缠能力;在自蒸馏模块,以提取到的域不变特征为教师对象,进一步提高泛化能力。 并且该检测器在只有一个源域进行训练的情况下,面对许多未曾训练过的目标域上仍然表现良好,提高了检测器在未知域的鲁棒性。 实验验证了模型在各种天气下城市场景目标检测的效果,实验数据表明,该方法优于基线约 8 百分点,相比基线方法获得了性能提升。

关键词：深度学习;目标检测;特殊天气;循环解纠缠;自蒸馏;域不变特征