Research On Image Retrieval Based On The Aggregation Of Local Invariant Feature

Content-based image retrieval(CBIR), in which the image feature extracted from the image itself directly can be used to describe the characters of image and measure the similarity among images, has been a popular research problem in computer vision. More importantly,CBIR is helpful to build the objective, automatic and efficient image retrieval system which has great significance for managing the massive image data generated from the internet. General speaking, the features used in CBIR can be summarized as following types: color, texture, shape, and local invariant feature. Compared with others, local invariant feature has been more widely applied to Image registration, object recognition, image retrieval and other computer vision problems because they are distinctive, robust to occlusion and invariant to image transformations.This thesis mainly focuses on the aggregation method of local invariant feature based on Bag-of-wordsmodel and Fisher Kernel. First we have these methods analyzed and summarized deeply. Then we propose some effective improvement approach on existing aggregation methods and employ the modified algorithm in image retrieval. The main work and contributions of this thesis are as follows:1. We make a research on the key points of image retrieval based on local invariant feature which mainly includes the extraction technology of local invariant feature, image similarity measure based on local invariant features as well as the evaluation criterion of image retrieval.2. By diving into details of the theory of Bag-of-words model and the aggregation methods based on Bag-of-words model, we utilize some useful methods to solve the problems existing in Bag-of-words algorithm such as: 1) the Kd-Tree approximate nearest neighbor search algorithm which can improve the efficiency of the feature encoding procedure; 2) the ZCA Whitening method which can reduce the relevancy of coding coefficients among each dimensions; 3) the Power-law normalization which can smooth the image representation vector with reduce the influence of bursty visual elements; 4) the sparse autoencoder which can further remove the redundant information and make the image representation vector more discriminative. The experiment result shows that our modified algorithm achieves a large improvement on image retrieval.3. In this thesis we pay more attention to the aggregation methods based on Fisher Kernel and propose a novel aggregation method. Due to the performances of Gaussian Mixture Model would degrade in face of high-dimensional data, we employ the Sparse representation model to model the generative process of local invariant features. In order to add the local structural information into the image vector representation, we utilize Locality-constrained linear coding to calculate the representation coefficients. Besides these, we introduce a modified Whitening operation to reduce the dimensionality and the redundant information of the image vector representation. The experiments on Holidays and Oxford5 K demonstrate that our method achieves a substantial improvement on performance compared with other state of the art.