A Hybrid Gaussian Model Matching Method Based on Improved Dozing Distance

Abstract

The invention proposes an algorithm system suitable for image classification and image retrieval. The algorithm mainly includes two modules of image modeling and image matching. In image modeling, in order to improve the description ability of features, the algorithm firstly uses deep convolutional neural network to extract the deep features of images, and on this basis, uses Gaussian mixture model to model and represent images. In image matching, aiming at the efficiency of the bulldozing distance algorithm and its sensitivity to noise, an accurate and efficient improved bulldozing distance algorithm is proposed. Gaussian noise and sparse constraints are introduced in the algorithm design process. Simultaneously, in designing the improved bulldozing distance algorithm, aiming at the defect of existing geodesic distance measurement algorithm design, the present invention designs three kinds of geodesic distances after considering the Riemann geometric structure of Gaussian distribution, in order to more accurately Measures the distance between Gaussian distributions.

Description

technical field [0001] The present invention relates to the technical fields of computer vision, probability statistics, and Riemannian geometry. It specifically utilizes deep convolutional neural network features to model mixed Gaussian models, and proposes an efficient and robust measure for the similarity measurement between mixed Gaussian models. algorithm. Background technique [0002] In the process of image processing, image feature representation, as the first step of image processing, plays a very important role. The feature representation of the image mainly expresses the image as one or more matrices in the mathematical sense through a series of algorithms. Image feature representation methods at this stage are mainly divided into two categories: features based on handcrafted features and features based on deep convolutional neural networks. The former is simple and efficient to calculate, including Scale-invariant feature transform (SIFT), Histogram of Oriented...

AI Technical Summary

Problems solved by technology

However, the traditional EMD algorithm still has the problems of computational efficiency and noise sensitivity. This method proposes an improved bulldozing distance algorithm (The Improved-Earth Mover’s Distance-I-EMD) to address these shortcomings.

Images