Research And Implementation Of Domain Adaptation

In recent years,deep learning has made great progress in computer vision and image processing.In the application scenario of deep learning,the lack of labeled data is often encountered.It's a useful method to use the labeled data of source domain to supplement information for target domain.However,there may be domain shift between source and target domain.Domain adaptation is an effective method to solve the problem of model capability degradation caused by the different data distribution between domains.This thesis studies domain adaptation algorithm.We explore unsupervised domain adaptation algorithm and make innovations and applications from two aspects: classification and objection.The main work of this thesis is as follows:(1)Deep Adaptation Network(DAN)is one of the classical methods,which utilizes Multi-Kernel MMD(MK-MMD).However,DAN can still be improved in feature level transfer and it has different effects in different adaptation scenarios.To solve these two problems,we first combine Domain Confusion with MK-MMD to further improve the adaptability of models.At the same time,we explore the suitable weights of MK-MMD in different adaptation scenarios and the best combination of MK-MMD and domain confusion from both experimental and theoretical aspects.Besides,We apply the deep adaptation network combining domain confusion with MK-MMD in the task of vehicle classification.The comparative application experiments successfully prove the effectiveness of the proposed methods.(2)In order to relieve the performance decline of source training model in target domain,we combine domain adaptation with object detection.We improve the object detection model Faster R-CNN from three aspects: domain adaptation in feature map,domain adaptation in region of interest and the ideal of double ROI Pooling.The proposal model successfully narrows the gap between source and target domain and improve the accuracy of object detection in different scene.Besides,we apply the improved domain adaptation multi-scene object detection network in the task of multi-scene vehicle detection.The comparative application experiments successfully prove the effect and advantages of the proposed methods.