Research On Image Retrieval And Image Classification Based On Feature Coding

In this highly informative times, with the high speed development of computer vision, and a sharp rise in the number of images every day, how to effectively process the images and how to get the pictures people interested in, which is an important subject. This paper focuses on two aspects of image retrieval and image classification, image retrieval is finding target image from the mass images. However, how to use effective retrieval methods to reduce computational complexity and ensure retrieval accuracy, this is not only a difficulty but also an emphasis, this paper proposed an image retrieval reranking algorithms. Classification in this paper is fine-grained image categorization, this is greater difficult than the common image classification. Because fine-grained categorization refers to the task of classifying objects that belong to the same basic-level class(e.g. different bird species), and they share similar shape or visual appearances. The main research results in this paper are as follows.1. A new method called image retrieval reranking is proposed. This method firstly segment one image using SLIC super pixel segmented method. We extract RGB and LBP feature in each super pixel block respectively, each super pixel block is encoded in both feature space respectively and by combining the codes in different feature space, a new combined code of a super pixel is achieved. Finally we calculate the similarities between images by codes of images to get the first ranking. We use two ways to achieve image retrieval reranking. The first one is called Image retrieval reranking based on weighted SVM classifier. We use Boosting optimization algorithm to train multiple classifiers and choose better classifiers to get the last strong classifier for reranking. The second one is called Image retrieval reranking based on MIL(Multiple Instance Learning) algorithm. We can train and test images using mi-SVM. Finally we use the classification probability for reranking.2. A new method called fine-grained image categorization based on Fisher Vector is proposed. This method firstly build GMM(Gaussian Mixture Model) according to the feature points. Then we get templates in training samples, and we use Fisher Vector to represent these templates. We use two ways to achieve the coding of images: The first one is called image coding method based on SPM(spatial pyramid model). This method in order to get the response map by matching each template and an image, then we can get the distributionin the space of the pyramid, and pool all templates distribution to achieve the coding of one image. The second one is called image coding method based on the dictionary of templates. This method firstly build dictionary by using K-means algorithm with the Fisher Vector of all templates, then get the coding of image traversal box on the dictionary. Then we pool the coding of each size of templates to achieve the coding of one image. Finally we use the SVM classifier to classify images.