Study On Application Of Transfer Learning Based On Deep Convolutional Feature In Image Recognition

Image recognition is one of the hot research directions in the field of computer vision,and it is also an important foundation for machine learning and artificial intelligence.The image recognition system generally includes three steps,which are feature extraction,image expression and classifier modeling.The extraction of image features is the key to determine the quality of the whole recognition system.In the field of image processing,the development of image feature extraction can be divided into two phases:the traditional artificial design extraction features and the current stage of deep convolutional neural network autonomous learning features.However,the method of self-learning image features by means of deep learning is the support of massive amount of annotation data.On the settlement of effective data-poor issues,the latest solution is to combine deep learning with transfer learning,based on the existing methods of deep transfer learning,focuses on the problems in the application of image recognition,mainly improved and discussed the following two aspects:(1)When applying depth transfer learning methods for image recognition on small sample data sets,an image recognition method based on the two transfer-features fusion learning is proposed,aiming at lack of specificity in the general features extracted from the target dataset for the source model in depth transfer learning.This method is to enhance the representational ability of transfer features by introducing special features,the final experiment shows that this method can obtain higher average classification and recognition accuracy than the deep convolutional neural network trained directly on the small-scale data set and the DDC deep transfer learning based on pre-training method,verify the effectiveness of the two transfer-features fusion learning method,and indicates that the fusion feature has stronger representation ability.(2)In the migration learning,the source models of different network structures can learn different representation features,aiming at how to fully and effectively use these representation features,an image recognition method based on multi-source model of two transfer-features fusion learning is proposed;Four classical convolutional neural network models were selected for experiments.The results show that compared with the single source-based two transfer feature fusion learning method,the accuracy of image recognition obtained by this method is higher,and the semantic information contained in the image can be more fully and accurately reflected.