Feature Optimization Method Based On Deep Learning And Label Correlation

How to extract the features to depict the images with rich semantic information has aroused the increasing attention.Feature extraction plays an important role in image classification and retrieval.Traditional methods use hand-crafted features,such as SIFT,to represent images,which have been widely applied in many domains.In recent years,Deep learning has been popular due to its excellent performance.Furthermore,a convolutional neural network has been shown its superior performance as a feature extractor tool,based on the ability to learn representation from raw data.Hence,CNN has been commonly applied to image classification and retrieval.At the same time,distance metric learning has achieved impressive performance in the field of image processing by using feature space transformation strategy.In this paper,the CNN-F is employed to extract image features for the image hashing method,which has been shown the superior performance of image retrieval compared to the manual feature SIFT.The experimental results on the NUS-WIDE dataset demonstrate the effectiveness and advantages of our proposed method.On this basis,in the multi-instance multi-label setting,the label correlation is estimated during the training procedure,and the optimal metric for classifier is learned to project bags with positively correlated labels to different feature space,such that their mahalanobis distances can be reduced as much as possible.Finally,the experiments on MSRC v2 have demonstrated the effectiveness and superiority of the proposed method based on label correlation.