Research On Domain Adaptation Methods Based On Deep Learning In Weak Dataset Scenarios

Deep learning is widely concerned because of its excellent performance.However,deep neural networks usually require a large number of labeled samples to train a large number of model parameters.In practical application,it is expensive to collect a large number of labeled samples,and only three kinds of weak datasets defined in this paper can be obtained:1)only have a large number of unlabeled samples and no labeled samples,2)only have a small number of labeled samples,3)have a small number of labeled samples and a large number of unlabeled samples.However,it is difficult to train a reliable deep neural networks directly using weak datasets with insufficient supervision information.In this paper,by studying the corresponding domain adaptation methods,we use weak dataset and auxiliary training set containing a lot of supervision information to train reliable deep neural networks.Firstly,for the weak dataset scenario where the training samples set only contains a large number of unlabeled samples,this paper proposes an unsupervised domain adaptation method based on Minimizing Output’s Differences of Classifiers with Different Responsibilities(MODCDR),constructing one generator and two task specific classifiers which work in the source domain and the target domain respectively.Based on the above structure,MODCDR can not only generate discriminant features,but also achieve categorylevel feature distribution alignment when training the generator to minimize the difference of prediction probability vectors outputted by these two classifiers.In addition,the two classifiers can be used to obtain reliable pseudo-label for self-training.The experimental results show that MODCDR has superior performance compared with the popular unsupervised domain adaptation methods,proving that MODCDR can achieve categorylevel domain adaptation simply and efficiently by combining discriminant feature generation,category-level feature distribution alignment and self-training.Secondly,for the weak dataset scenario where the training samples set only contains a very small number of labeled samples,this paper proposes a supervised domain adaptation method based on Condition Category Alignment(CCA),which makes the feature distribution of source domain and target domain be conditionally optimized according to their own states.In the training phase of CCA,the estimated intra-class distance is minimized and the estimated inter-class distance is maximized only when the estimated intra-class distance of a batch of samples is large enough and the estimated inter-class distance is small enough.In addition,CCA uses the difference between the estimated maximum intra-class distance and the estimated minimum inter-class distance of a batch of samples to adjust the strength coefficient of optimizing feature distribution.The experimental results show that CCA performs well in the weak dataset scenario,which proves that CCA achieves better category-level domain adaptation by conditionally optimizing feature distribution.Finally,for the weak dataset scenario where the training samples set contains a large number of unlabeled samples and a small number of labeled samples,this paper proposes a new semi-supervised domain adaptation method based on Feature Distribution Alignment Based on Category Prototype(FDABCP).FDABCP use the classifier based on cosine similarity to learn the category prototype,reducing the negative impact of mixing the supervision information of the source domain and target domain.FDABCP uses training samples not only to adjust the classifier and generate discriminant features,but also to construct two classifiers learning the source and target domain category prototypes to achieve the category-level feature distribution alignment.The experimental results show that FDABCP has superior performance in the weak dataset scenario,which proves that FDABCP can further realize the value of training samples to achieve the category-level domain adaptation by achieving the category-level feature distribution alignment directly.