Research On Global Affine Invariant Method Of Image

With the development of computer and information technology, using computer visionsystem to assist or replace the human visual perception system can reduce the humanworkload or perform tasks that humans can’t complete, such as fields of image data retrieval,lunar exploration. When a computer vision system is used to achieve human one, it will facethe problem of identification and recognition of ‘observed’ Objects like human. Humans canuse invariant information to identify objects, while the computer can recognize targets byselecting and extracting objects for pattern recognition.As changes of the viewpoint and the distance will make the acquired images of thesame scene differ from each other, extracting features unaffected by the sensor attitude andposition from these images is not only a common problem during many image intelligentprocessing applications, but also a conundrum that cannot bypass in practice. The measuredtargets on the images acquired by image sensors satisfy the perspective transformation, whenthe target size is far smaller than the distance between target and sensor, a perspectivetransformation model can be approximated by an affine transformation model, and the affinetransformation can better describe the relationship between the image target in differentviewpoints and distances. In the technical fields of image retrieval, image digitalwatermarking, satellite remote sensing image processing, and target recognition, it oftenneeds to deal with extraction of features unaffected by geometric deformation from differentview images for the follow-up of other problems, so affine invariant feature has obviousapplication advantages, it has become a hot topic in these areas.With the application background of image processing and object recognition, this paperstudies the theory and method of affine invariant feature extraction based on multi-scaletheoretical framework and the affine geometry transform. It focuses on practical extractionmethod of affine invariant features that satisfy fine discrimination, strong anti non-affinedeformation ability, quick calculation speed, and wide application range. The main contentsof this paper are as follows:(1) Research on fast multi-scale auto convolution transform algorithmAiming at the great computation complexity of multi-scale auto convolution transform,this paper analyzes the impact of scale value on complexity in every transform computationphase. And a mathematical model that represents the relationship among computationalcomplexity, image size and scale value is established to analyze the proportion of thecomplexity of the Fourier transform. This Paper also studies minimum transform scalerequirements to maintain multi-scale auto convolution transformation value. As scale in therange of [-1,1] keeps an interval of1/Z+(Z+is an integer greater than1), this paperresearches on minimum transform frequency fast algorithm; for scale outside the range of [-1, 1], it studies minimum transform scale fast algorithm; According to the fact that partial scalewill meet1/Z+interval after transformation, it studies on the tandem rapid algorithm ofminimum transform scale and frequency. This paper analyzes the efficiency of themulti-scale fast algorithm and compares it with the improved efficiency of the actualrun-time algorithm by the computational complexity model. Also, application of rapidalgorithm in high-dimensional multi-scale auto convolution transformation has been studiedby establishing the computational complexity model of3-D multi-scale auto convolutiontransformation and analyzing the efficiency improved at a given scale in theory. Additionally,a fast algorithm of normalized affine moment invariants has been studied. Experimentalresults show that the transform speed of the proposed fast algorithm under various scales willbe increased to about2to6times. In the mean time, it still maintains the eigenvalueaccuracy that consistent with the original method. Particularly, the speed is increased to morethan2times when the minimum transform frequency fast algorithm is applied to thenormalized affine moment invariant.(2) Research on multi-scale histogram affine invariant extractionBased on the density function of multi-scale auto convolution transformation, this paperstudies on the normalization method that maintains the affine transformation relations ofdensity function, it proposes a histogram invariant feature extraction algorithm and then useshistogram to construct moments and entropy invariant, moreover, it generalizes thealgorithm to high-dimensional application. This paper derives the interchangeability,symmetry and standard form invariance of histogram feature, it analyzes the relationbetween scale values and eigenvalues and researches on the scales, gray threshold selectionmethod and the histogram features implementation. Relying on a two-dimensionalcoordinates, it makes a linear transformation at the two given factors and builds atwo-dimensional density function, furthermore, it studies how to convert functionconvolution value into histogram interval division for two-dimensional histogram featureextraction. This paper designs experiments to analyze the impact of scale values, graythreshold on two feature recognition rate and put features to test the classification effect ofbinary images, grayscale images and the perspective transformed images. Experiments showthat the recognition rate of two features performs better than MSA features of classic scalesin terms of anti non-affine deformation; especially, the two-dimensional histogram featurehas the best overall performance and fastest computational speed among others.(3) Research on affine geometry invariant extraction based on extended centroidThe extended centroid method obtains a series of affine invariant points from image byiterative affine zoning; it uses invariant points to construct geometric invariant feature. Thispaper studies the position relation between each pair of extended centroid obtained by affinezoning and the image centroid; it develops non-redundant affine zoning strategy and studiesaffine invariant feature extraction based on line length ratio and an extended image construction method. Also, it proposes a new extended centroid extraction method byincreasing the number of extended images, reducing the number of zoning, improving thefeature dimension and constructing line length ratio invariant. This paper makes experimentsto analyze the impact of the colinearity of the centroid, the affine invariance of extractedfeatures and spread-function on the stability of features, feature recognition ability andcomputational efficiency. Experimental results show that the cumulative error of theproposed zoning strategy is small, compared with the classical triangle, quadrangle area ratioinvariants, in the conditions of same feature numbers, the extracted feature has a betterability of robustness and classification.(4) Research on multi-scale affine geometry invariant extractionThis paper uses a multi-scale framework to construct a series of affine covariant image,develops an affine area ratio invariant feature extraction method based on extended centroid,utilizes covariant image to construct multi-scale area ratio invariant, proposes a compositepartitioning strategy to construct any number of invariant features by an affine zoning. Itresearches on the impact of scale values on covariant image and feature invariance, derivesthe minimum domain when feature maintains affine invariance. It achieves rapid calculationof invariant features. Covariant image intensity is determined jointly by the original imageintensity and the intensity distribution information, the covariant image intensity of binaryimage is not a single value, study of features can be used for the binary image recognition.This paper also studies on the method of invariant point extraction by using covariant imagefor the original image registration. Experimental results show that multi-scale affinegeometric feature performs better than MSA feature comprehensively in terms of anti-noise,anti-illumination changes, partial occlusion resistant, and anti-deformation. Under mostconditions, it has a better performance and a faster computational speed than MSA momentfeature. In terms of image registration, its accuracy is much higher than the original extendedcentroid method.This paper presents a fast affine invariant feature algorithm and three affine invariantfeature extraction methods; it has important application value in terms of target recognition,image auto-registration and so on.