JP2007226756A - Method and device for evaluating left image such as footprint using ridgelet transform - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and a device for evaluating a left image such as footprints by which personal authentication in criminal investigation is effectively performed by performing Ridgelet transform of the left image. <P>SOLUTION: In the Ridgelet transform, an edge on a straight line is effectively expressed, noise is effectively removed, magnification, reduction, parallel translation and rotation are easily performed in a transformation domain by nature of Radon transform for performing linear integration of a luminance value. Furthermore, a Radon domain is reproduced while raising resolution of a footprint image subjected to the Ridgelet transform stepwise and pattern matching is allowed by Hough transformation since the Radon transform is equal to a parameter space of the Hough transformation. The luminance value is differentiated in the arbitrary direction and applied to general Hough transformation by using the Ridgelet transform. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention belongs to a technical field related to a method and apparatus for identifying a residual image that enables effective personal authentication in a criminal investigation by performing Ridgelet conversion of a residual image such as a footprint.

  In the so-called remains image appraisal, for example, footstep images that have undergone Fourier transformation are classified according to circle, saddle shape, wave shape, mountain shape, straight line, and other patterns, and a footstep image database is constructed by assigning keywords, and a bandpass filter Template matching is applied. Further, the image registered in the database and the footprint image are manually collated. In recent years, computer systems have been introduced to automate keyword searches, but the classification of footprint images remains manual. Hereinafter, conventional techniques related to the present invention will be described.

Normalize images taken from a digital camera or image scanner in the creation of a shoe sole database that is useful for collating data from left footsteps and a convenient search method for this. There is an invention to code. (See Patent Document 1)
JP-A-2005-27951 (FIG. 1)

P. Chazal, J .; Flynn and B.M. Reilly has proposed a system in which footprint images are subjected to Fourier transform and searched by pattern matching. (See Non-Patent Document 1)
P. Chazal, J .; Flynn. and B. Reilly, “Automated Processing of Images Based on the Fourier Transform for Use in Forensic Science”, IEEE Trans. , Vol. PAMI-27,3, pp. 341-350 (Mar. 2005)

The Ridgelet transform is an E.I. J. et al. This is a signal processing technique proposed by Candes, which is obtained by performing Wavelet conversion after Radon conversion of a signal. (See Non-Patent Document 2)
E. J. et al. Candes, “Ridgelets: Theory and application”, Ph. D. dissertation, Dept. Statistics, Stanford Univ. , Stanford, CA (1998)

Since the Radon transform has the property of linearly integrating the luminance value, it is possible to effectively represent an edge on a straight line. Moreover, noise can be effectively removed. (See Non-Patent Document 3)
Minh N. Do and M.M. Vetterli, “The Fine Ridgelet Transform for Image Representation”, IEEE Trans. Image Processing, 12, pp. 16-28 (Jan. 2003)

Using the fact that the Radon transform is expressed in polar coordinates, the object can be easily rotated by extracting the subject from the image and then performing the Ridgelet transform. (See Non-Patent Documents 4 and 5)
M.M. Hasegawa and S.H. Tajima, “A Ridgelet Representation of Semantic Object Using Watered Segment”, Proc. of IEEE ISCIT 2004, Sapporo, Japan, (Oct. 2004) Makoto Hasegawa, Shinichi Tajima “Ridgelet expression of Semantic Object by Watershed division”, Journal of the Institute of Image Information and Television Engineers, 59, 5, pp. 786-790 (May 2005)

Since the Radon domain is equal to the Hough transform parameter space, it is possible to reproduce the Radon domain while gradually increasing the resolution of the Ridgelet transformed image and perform pattern matching by the Hough transform. (See Non-Patent Document 6)
Matsuyama Takashi, Komizu Yamato, “Hough transform and pattern matching”, Information Processing, 30, 9, pp. 1035-1046 (Sep. 1989)

R. Krishnapuram and D.K. Casasent proposes a method of recognizing an object of arbitrary shape by obtaining a correlation with a Hough parameter space as a template as well as detecting a straight line. (See Non-Patent Document 7)
R. Krishnapuram and D.K. Cassent, “Hough Space Transformations for Discrimination and Distribution Estimation”, Computer Vision, Graphics. and Image Processing, 38, pp. 299-316 (1987)

The Radon transform can differentiate a luminance value in an arbitrary direction by an operation in the transform domain. (See Non-Patent Document 8)
S. R. Deans, The Radon Transform and Some of It Applications, John Wileyans Sons (1983)

  In the Fourier transform, enlargement / reduction is performed by pre-processing, and for parallel movement, the amount of movement is made unchanged by using the energy of the conversion coefficient. In addition, since the rotation rotates the transformation domain, the correction timing does not match and flexible alignment with the pattern matching target image such as the shoe sole image is difficult. Furthermore, since the remains of images such as footprints remain in a poor environment, there are many defects and noises. This is a simulation stage using ideal images and has not yet been put into practical use.

  The Ridgelet transform is obtained by performing a Wavelet transform after performing a Radon transform on a signal. However, due to the nature of the Radon transform that linearly integrates luminance values, edges on a straight line can be effectively expressed and noise can be effectively removed. By utilizing the fact that the Radon transform is expressed in polar coordinates, the operation image in the transform domain can be easily scaled, translated, and rotated. Further, since the Radon transform is equal to the Hough transform parameter space, it is possible to reproduce the Radon domain while gradually increasing the resolution of the Ridgelet transformed footprint image, and easily perform pattern matching by the Hough transform.

  Personal identification based on remains images such as footprints is an effective means of criminal investigation, and even if the suspect denies the crime, the footprint image collected from the crime scene is identified, and if the characteristics match, the crime of the same person is authorized Although there are cases where it can be performed, the accuracy and speed required for this appraisal can be realized and utilized for information processing using a computer.

In the following, a method for Ridgelet conversion of footstep images among the remnant images, storing and searching in a database, a method for recognizing patterns by Hough conversion, a method for applying to general Hough conversion, and the configuration of the apparatus will be described. The configuration of the apparatus is shown in FIG.
The processing flow is shown in FIGS.

  It is possible to remove a noise by expressing a specific shape from a footprint image taken by a residual image extraction means (digital camera, image scanner, etc.) and stored in a database using a Ridgelet conversion means.

  The footprint image is given as shown in FIG. FIG. 5A is a footprint image left at the crime scene, and FIG. 5B is a shoe sole photograph of the corresponding shoe. In FIG. 5A, there are many noises and defects, and a plurality of footprints overlap. The shading due to the sole material is not seen in FIG. 5 (a), and the projections on the sole are displayed in white. That is, the correlation between the luminance values in the two images is small. Therefore, instead of directly matching luminance values, an image obtained by extracting a contour line using a differential filter is used.

ことができる。

Here, a Sobel filter is applied as a differential filter, and then the luminance value is binarized. FIG. 6 is an example of a contour image. The image shown in FIG. 6 is Ridgelet transformed. Represents a 2-dimensional region R ² of the coordinate image is given by the position vector x, to the luminance value f (x). Note that the luminance value f (x) immediately after the preprocessing is 0 or 1. Ridgelet of f (x)

be able to.

If ξ = (cos θ, sin θ), the Ridgelet transform D can be replaced with the equation of equation 4 using the Radon transform (equation of equation 3). That is, the Ridgelet conversion is equivalent to the Wavelet conversion after the Radon conversion of the grayscale image.

  FIG. 8 shows the result of Ridgelet conversion of FIG. Here, it is decomposed into four subbands by Wavelet transform. Further, the Haar basis is used as the basis function.

  FIG. 6 shows the result of reproducing the Radon domain by performing inverse Wavelet transformation on the Ridgelet domain.

  The result of reproducing the Radon domain using a part of the subbands in the Ridgelet domain is shown in FIG.

When the Radon domain in FIG. 9 is scaled in the vertical direction, the original image in FIG. 6 is scaled in the vertical and horizontal directions. Further, when the Radon domain is moved up and down in accordance with the declination angle θ (or ζ) using the equation of Formula 5, FIG. 6 moves in parallel in the vector a direction.

  6 can be rotated as shown in FIG. 11 by shifting the Ridgelet domain left and right.

  A method for recognizing a pattern by Hough conversion will be described. It is known that the Radon domain and the Hough transform parameter space are equal. That is, it is possible to detect a linear component using the Radon domain. FIG. 12 shows an example in which a linear component is detected. A Radon transform coefficient having a large absolute value is extracted, and inverse Radon transform is performed to detect a linear component. A straight groove in the shoe heel has been detected. It is conceivable to identify shoes by matching detected straight lines. Noise that does not fall on a straight line is detected. It is possible to remove noise by extracting only the linear component.

標準偏差とする。また、ｐはドメイン内における係数の個数である。

Using the Radon domain reproduced stepwise from the Ridgelet domain for each resolution, the correlation is calculated and pattern matching is performed. The Radon domain of the two images

Use standard deviation. P is the number of coefficients in the domain.

  By creating a logical sum image of shoe sole images that are candidates for pattern matching and extracting a logical product image of the images and footprint images, noise and unnecessary patterns in the footprint images can be removed. The image from which noise has been removed can be subjected to Ridgelet conversion again, and processing can proceed to matching at a high resolution.

  Here, the Radon domain is matched, but it is also possible to match the Ridgelet domain directly. The size of the image can be changed by vertically scaling the Ridgelet domain. It is also possible to rotate the image by shifting the Ridgelet domain left and right.

A method applied to general Hough conversion will be described. It is possible to differentiate the luminance value of the original image in an arbitrary direction in an arbitrary direction by an operation in the domain. (See Equation 7) By detecting the normal direction of the contour line and applying it to the generalized Hough transform, the position and normal direction of each point indicating the contour line of the footprint image are arranged in the R table. , Generalized Hough transform. An arbitrary shape can be detected more flexibly by the generalized Hough transform.

  The shoe sole pattern is classified by general Hough transformation to generate metadata and create a database. Pattern matching using metadata is also possible.

Ｒｉｄｇｅｌｅｔドメインの例 足跡画像（ａ）、靴底の画像（ｂ） Ｒａｄｏｎドメインの例 足跡画像（ａ）、靴底の画像（ｂ） 各解像度におけるＲａｄｏｎドメインの例 バイアス成分のみ（ａ）、第２サブバンドを含む（ｂ）、第３サブバンドを含む（ｃ） Ｒａｄｏｎドメイン内での操作による回転 ドメインの左シフト（ａ）、時計回りの回転（ｂ） Ｈｏｕｇｈ変換による直線検出 Diagram Remaining image coding flow Remaining image search flow Deformation flow of images by operation of Ridgelet domain Examples of footprint images Footprint images (a), shoe sole images (b) Example of pre-processed image Footprint image (a), shoe sole image (b)

Ridgelet domain example Footprint image (a), shoe sole image (b) Example of Radon domain Footprint image (a), shoe sole image (b) Example of Radon domain at each resolution Bias component only (a), second subband included (b), third subband included (c) Rotation by operation within Radon domain Left shift of domain (a), clockwise rotation (b) Straight line detection by Hough transform

Claims (5)

  A method for appraisal of a survivor image such as a footprint image, which is extracted from a subject and then subjected to Ridgelet transform to express a specific shape, remove noise, and can be stored and searched in a database.   2. The method for appraisal of a residual image according to claim 1, wherein the image is subjected to enlargement / reduction, parallel movement, and rotation by an operation in a conversion domain at the time of Ridgelet conversion.   The Radon domain is reproduced while gradually increasing the resolution of the Ridgelet transformed residual image in the method for identifying a residual image according to claim 1 or 2, and the footprint image can be pattern-matched with the shoe sole image by Hough transformation. A method for appraisal of a residual image.   The method of claim 1, wherein the luminance value of the original image is differentiated in an arbitrary direction by a Radon domain operation at the time of Ridgelet conversion, the normal direction of the contour line is detected, generalized Hough conversion is performed, and the metadata is A method of appraisal of a remnant image characterized by generating and searching.   Claim 1, 2, 3, or 4, wherein the residual image extraction means, Ridgelet conversion means, Radon conversion means, Wavelet conversion means, Hough conversion means, storage means, search means, matching means by straight line detection, A remnant image examination apparatus comprising reproduction means, enlargement / reduction means, movement means, rotation means, differentiation means, metadata creation by general Hough conversion and matching means.

Images