Research On Image Classification Method Based On Deep Transfer Learning

With the continuous development of deep learning,deep convolutional neural networks have made breakthrough in general object classification tasks in recent years.This is mainly due to the classification method based on the deep convolutional neural network has more powerful feature extraction and image representation ability than the method based on manually defined features,and can obtain more accurate and stable classification effects.Although many image recognition methods have good performance when applied to training data and test data extracted from the same distribution,these methods are not applicable in practical scenarios and result in low performance.At the same time,in many practical applications,it needs to classify fine-grained objects.Existing models predominantly require extra information such as bounding box and part annotation in addition to the image category labels,which involves heavy manual labor.Through the in-depth study of image classification methods based on deep transfer learning,the joint adaptive method based on attention transfer mechanism is used to improve the domain adaptive image classification effect,and the hierarchical deep transfer learning method is used to improve the image classification effect of fine-grained micro datasets.The main contents of this paper are as follows:1.The research background and current status of deep transfer learning methods are expounded.In-depth analysis and research on transfer learning,including key technologies and important theories.The composition,characteristics and optimization methods of convolutional neural networks are described in detail.The common methods and characteristics of deep transfer learning in image classification are systematically analyzed.The advantages and disadvantages of several classical domain adaptation methods in the field of transfer learning are compared in detail.2.Based on the analysis of data distribution,we propose a joint balanced adaptive method based on attention transfer mechanism,which transfers feature representations extracted from the labelled datasets in the source domain to the unlabeled datasets in the target domain.First,we transfer the labelled source-domain space category information to the unlabeled target domain through attention transfer mechanism.The model uses attention information to improve image recognition accuracy by defining the convolutional neural network attention.Second,we introduce the prior distribution of the network parameters based on the target dataset and endow the layer with the ability to automatically learn the alignment degree which should be pursued at different levels of the network.Finally,we describe the input distribution of the domain-specific adaptive alignment layer by introducing cross-domain biases,quantitatively indicating inter-domain adaptation degree that each layer learns.This method not only obtains higher recognition accuracy,but also much better than the traditional manual feature methods.3.Fine-grained categorization is challenging due to its small inter-class and large intra-class variance.Moreover,requiring domain expertise makes fine-grained labelled data much more expensive to acquire.We propose a deep transfer learning model to efficiently transfer image features learned on large-scale labeled fine-grained datasets to micro fine-grained datasets.Firstly,we introduce a cohesion domain to measure correlation degree between source domain and target domain.Secondly,the source-domain convolutional neural network is adjusted according to its metrical feedback,in order to select task-specific features that are suitable for transferring to the target domain.Finally,we make most of perspective-class labels,which are inherent attributes of fine-grained data for multi-task learning,and learn all the attributes through joint learning to extract more discriminative representations.Experiments show that the model not only economizes training time effectively and achieves high categorization accuracy,but also verifies that the inter-domain feature transition can accelerate learning and optimization.