Research On Image Classification Algorithmvia Transfer Learning

Image classification is a very common learning task in the field of computer vision.Traditional image classification tasks require a large amount of tagged data to train the model,but image annotation takes a certain amount of time and manpower.In practical applications,sometimes it is necessary classify categories that are not common or have not occurred.When the number of images with marks is small,and the traditional image classification model can not achieve satisfactory results,the idea of transfer learning can be utilized to utilize the marked data in the relevant field,that is,the source domain data to assist the target domain classification.In the transfer learning,there are some differences between the source domain and the target domain.How to reduce the difference between the source domain and the target domain in the transfer learning,and using the marked data in the source domain to assist the target domain to classify is the focus of this paper.Firstly,this paper studies the unsupervised domain adaptation problem in transfer learning.For the problem that the source and target domain probability distributions are inconsistent and the difference is large,a transfer learning method that simultaneously adapts the source and target domain edge probability distribution and conditional probability distribution is proposed.And in the adaptation process,the two different weight coefficients are respectively given,and at the same time,the category discriminant feature learning is performed in the adaptation process,and the distance between the same category samples is minimized in the feature space,and the different types of samples are Maximize the distance between each other to improve classification performance.Secondly,this paper studies the semi-supervised depth domain adaptation problem in transfer learning.For the case that the number of target domain samples in domain adaptation is small and some samples are marked,the method of semi-supervised domain adaptive learning using category semantic migration is proposed.Firstly,the depth domain is used to extract the source and target domain images,and then the CORAL method is used to reduce the domain difference between the source domain and the target domain.The entropy minimization principle is used to compare the source and target domains with high similarity.The categories are structurally aligned,so that the similarity relationship between the categories can be preserved in the feature space,the semantic migration between the sample categories can be performed,and the classification ability of the network to the target domain samples can be improved.Finally,this paper studies the zero-shot learning problem in transfer learning.For the Hubness problem and domain shift problem in zero-shot learning,a method of zero-shot learning using deep embedded model is proposed.Firstly,the feature extraction of known and unknown samples is performed by using the depth network,and then the sample features and corresponding category semantic representations are embedded in a common embedded space.In the embedded space,the feature embedding vector and the semantic embedding vector are minimized.The distance between the known class samples in the embedded space is classified by a linear classifier,which indirectly constrains the dispersion between different classes of samples,so that the known class samples have certain separability in the embedded space.In the prediction stage of unknown categories,the semantic embedded vectors of unknown categories are modified in the embedded space,and the k-nearest neighbors of the unknown class embedded feature vectors corresponding to the embedded semantic vectors are searched in the embedded space,and then the k neighbor vectors are used.The mean represents an unknown class semantic embedded vector such that the unknown class is closer to the unknown class in the prediction phase.