Multi-level Image Feature Extraction Based On Deep Structure And Research On Image Retrieval Application

Picture information is presented as carriers gradually hot with the rapid development of the Internet, the number of online images showing explosive growth, especially in the recent rapid development of mobile Internet terminal. With the emergence of a large number of image resources, people demand from the image retrieval and the traditional way with a character as key to retrieve by using the label tag pictures of models increasingly just cannot meet the requirements for the new Internet. To retrieve similar images through the semantic of image is the development direction of image retrieval in the future. How to extract the semantics of the picture itself, and be retrieved by its extracted information is the main content of this paper.In this case, from the traditional way of visual words mode, a combination of human habits for picture identification and the Convolutional Neural Networks is proposed as multi-level feature for image retrieval methods. This paper focuses on extracting the region of interest, feature extraction of region and image retrieval method. In the image feature extraction part, paper use convolutional neural networks to extract feature, by network mode ZF5 with pool5 layer replaced by SPM structure. The seventh layer, the activation value vector of fc7 is adopted as the image feature. In this paper, we follow these steps to get the feature representation of a single image. First we adopt selective-search method to generate the original candidate proposals. The original sets would be filtered by multi-ways as size, location, score etc. In order to adjust to some meaningless images, we also add the original image to the selected set. Experiments shows that, in all dimensions by PCA, the new feature combination outperforms the accuracy of the single-level feature a little in Holidays dataset. With a KD-Tree index structure, the accuracy would promote two presents more. And in the object-rich dataset VOC, multi-level feature shows a great promotion than single-level feature. The multi-level features extracted from complex picture could be viewed as appropriate representation for various kind of pictures and ensure image retrieval accuracy rate.