On The Deep Transfer Learning And Its Applications

Recently,deep learning has achieved a great success in the field of computer vision and shown amazing expressive abilities.Although these achievements are significant,most of them rely on a large amount of labeled data for getting acceptable performance.Therefore,it is important to learn features from unlabeled data.Transfer learning aims to transfer the learned knowledge in one scenario to another application scenario,which is an important exploration of deep learning landing.This paper focuses on the topic of transfer learning algorithms and their application in automatic driving and the main work is as follows:(1)For the closed set domain adaptive problem,we propose an iterative domain adaptive method,which gradually employs the pseudo-labels for training the target classifier.This algorithm provides a basis for directly optimizing the classification error,and the proposed method can even be applicable to both unsupervised and weakly supervised scenarios.The performance of this method is comparable to various state-of-the-art domain adaptive methods.In particular,it can perform well even for some difficult tasks.(2)For the open-set domain-adaptive problem,this paper improves the binary cross entropy loss function of OSDA-BP algorithm for picking up the potential unknown target samples.A symmetric and distance-based Kullback-Leibler loss function is proposed.The experiment on the office-31 dataset shows that the proposed algorithm can effectively improve the performance of the original OSDA-BP algorithm.(3)To deal with the problem of insufficient real-world data training for autonomous driving,a transfer learning algorithm is proposed,which can achieve the model transfer from the simulated environment to the actual real world.In this regard,this paper proposes a method to transfer the driving strategies based on simulation scenes to real road scenes without any labeled information.In order to verify the algorithm,we conduct real-world experiments for automatically driving the vehicle over the closed-loop real industrial Park Road.Without human interference,the vehicle can run a complete closed-loop,proving the effectiveness and reliability of the proposed solution.