Cross-Domain Object Detection Based On Style Transfer And Guidance Mechanism

Object detection is an important part of machine vision,and deep learning theory has made great progress in this field in recent years.Most of the existing object detection methods assume that their training data and test data come from the same dataset,that is,obey the same distribution.However,in practical application scenarios,this assumption is difficult to satisfy.If the training data and test data do not meet the same distribution assumption,the performance of the object detection model will drop sharply.In order to solve this problem,this paper carries out research on cross-domain object detection based on the theory of transfer learning.The problem solved in this study is that there are domain differences between the training data and the test data in cross-domain object detection,which cannot satisfy the hypothesis of the same distribution.This paper proposes that the main reason for the inconsistent data distribution between the source domain where the training data is located and the target domain where the test data is located is the large differences in image style,object appearance and object size between these two domains,which have not been taken into account in the previous cross-domain object detection methods.In order to alleviate the performance degradation of the object detection model caused by differences in image style,object appearance and object size between the source domain and the target domain,this paper proposes two ideas.1)At the data level,based on the existing style transfer model,this paper focuses on the quality and diversity of the generated images,and trains the style transfer model that meets the specific task.And by adjusting the content of the source domain data and the style of the target domain data,the generated data has the style of the target domain data,while the objects on the image still maintain good semantic characteristics,which makes the distribution of the source domain data and the distribution of the target domain data closer.2)At the model level,this paper proposes two guidance modules:high-level feature module and multi-scale feature extraction module.The high-level feature module uses its high-level semantics to guide the learning of the lowlevel network layers in the model.The features extracted by the multi-scale feature extraction module contain rich object position information,and are fused to different network layers to make the network learn effectively.At the same time,the squeeze-excitation module and feature fusion module in this paper make the features learned by the entire model more discriminative.In this paper,the proposed method is tested on multiple public object detection datasets.In the test of using VOC2007-trainval and VOC2012-trainval as the source domain dataset,BDD100K,Clipart lk,Comic2k and Watercolor2k as the target domain datasets,respectively.The proposed method can reach 22.51%,40.7%,39.05%and 55.02%mAPs in BDD100K,Clipartlk,Comic2k,and Watercolor2k,respectively.In the test of using UCAS-AOD as the source domain dataset and NWPU VHR-10 as the target domain dataset,the mAP of the proposed method achieves 67.7%.In the test of using URPC2019 as the source domain dataset and ChinaMM2019 as the target domain dataset,the mAP of the proposed method achieves 56.33%.The experimental results verify the effectiveness of the proposed method.