Image Retrieval Based On Local Invariant Moment Of Sketch

The popularity of portable camera devices has contributed to the rapid growth of image data,and the demand for efficient image retrieval methods has become increasingly urgent.Sketch images have a great advantage in describing the concept of blurred images in human mind.With the promotion of touch-screen mobile phones and computers,users can easily draw and enter the sketch images that need to be queried.These all facilitated the development of sketch-based image retrieval method.In this thesis,aiming at the problems in sketch retrieval,such as the database image preprocessing is rough,the extracted contours contain too much noise,the sketch feature extraction method is less robust to the deformation,the matching search method is not optimized,and the user interaction is poor.The image preprocessing,sketch feature extraction and image matching are mainly researched.The main contents and innovation of this thesis include:1.A method which draws sketch outline from color image based on Snake and Significant figure is proposed.Aiming at the problem of ignoring or not carefully preprocessing the image in database,the apparent contrast method is firstly used to determine the general position of object in the image.Then the initial contour is drawn with gradient information.Furthermore,snake model is iteratively used to get accurate contour.The proposed method can get sketches which have simple structure,prominent main part and less pixel noise from color image.The result sketches are similar to user hand-painted sketches.2.A sketch feature extraction method based on local geometry moment invariant is proposed.In existing sketch feature extraction methods,the robustness to the sketch description is not good.Aiming at the problem,the coordinate system of image is determined by the geometric characteristics of image.Furthermore,geometry moment invariant is calculated as eigenvector of image blocks which are divided from over all image based on the coordinate system.In the method,moment character descriptor which has great robustness to the changes in the images is used to describe the sketch.It not only improves the effect of sketch recognition,but also improve the matching accuracy of large-scale image retrieval.Compared with Shape Context,Histograms of Oriented,GIST feature and Edgel Index structure methods,the retrieval accuracy of proposed method has increased 12 percentage points.Experiments show that it also has a better recognition effect on the images after translation,scaling,flipping transformation and deformation.3.An image retrieval and optimization of results method based on fuzzy theory and ant colony Algorithm is proposed.Most of the sketch matching methods are simple and poor interaction with user.Aiming at the problem,the order of the images after retrieval is sorted according to the membership function.Then,the image correlation degree matrix is updated according to the results of use's feedback.Finally,the matrix is used to optimize the retrieval results.Experiments show this method has a certain improvement in the accuracy of the image retrieval results for different sketch feature extraction methods.