Optical method and system for rapid identification of multiple refractive index materials using multiscale texture and color invariants

• US 20050126505A1
• Filed: 12/10/2004
• Published: 06/16/2005
• Est. Priority Date: 12/10/2003
• Status: Active Grant

• Current Assignee: Woods Hole Oceanographic Institution
• Original Assignee: Woods Hole Oceanographic Institution
• Inventors: Gallager, Scott M., Tiwari, Sanjay
• Granted Patent: US 7,415,136 B2
• Time in Patent Office: Days
• US Class Current: 119/234
• CPC Class Codes: A01K 61/54 of bivalves, e.g. oysters or mussels
• CPC Class Codes: G02B 27/52 Phase contrast optics in microscopes G02B21/14
• CPC Class Codes: Y02A 40/81 Aquaculture, e.g. of fish

Abstract

An innovative optical system and method is disclosed for analyzing and uniquely identifying high-order refractive indices samples in a diverse population of nearly identical samples. The system and method are particularly suitable for ultra-fine materials having similar color, shape and features which are difficult to identify through conventional chemical, physical, electrical or optical methods due to a lack of distinguishing features. The invention discloses a uniquely configured optical system which employs polarized sample light passing through a full wave compensation plate, a linear polarizer analyzer and a quarter wave retardation plate for producing vivid color bi-refringence pattern images which uniquely identify high-order refractive indices samples in a diverse population of nearly visually identical samples. The resultant patterns display very subtle differences between species which are frequently indiscernable by conventional microscopy methods. When these images are analyzed with a trainable with a statistical learning model, such as a soft-margin support vector machine with a Gaussian RBF kernel, good discrimination is obtained on a feature set extracted from Gabor wavelet transforms and color distribution angles of each image. By constraining the Gabor center frequencies to be low, the resulting system can attain classification accuracy in excess of 90% for vertically oriented images, and in excess of 80% for randomly oriented images.

Claims

1. A flow-through imaging system for capturing images of high order refractive indices samples suspended in a liquid stream comprising:

a flow-through sample processor comprised of

a filtered liquid reservoir;

a sample container comprising

a sample inlet port for introducing said samples;

a top inlet port for combining said samples with a filtered liquid received from said reservoir;

a bottom inlet port for introducing a pressurized gas and a bottom outlet port for transporting a liquid stream of suspended samples to a first flow cell inlet port;

a sample flow cell for presenting a suspended liquid stream of said samples to a camera focal plane, said flow cell comprising

an imaging window for filming said samples in said liquid stream;

said first flow cell inlet port for receiving said suspended sample stream;

a second flow cell inlet port for introducing a sheath stream of filtered liquid from said reservoir, said sheath stream flowing around said sample stream and orienting said samples in said imaging window for filming;

a flow cell outlet port for transporting said filmed samples to a sample collector;

a sample illumination module on a proximal side of said flow cell comprised of

a brightfield illuminator comprising a xenon strobe light;

a first linear polarizer for filtering said strobe light; and

a condenser lens for producing polarized Koehler illumination for illuminating said samples;

a light receiving module on a distal side of said flow cell comprised of

an objective lens for collecting and magnifying light transmitted through said birefringent samples;

a full wave compensation or retardation plate for accentuating a speed difference between a plurality of slow and fast transmitted light rays, said compensation plate accentuating an interference color brightness of said transmitted light;

a linear polarizer analyzer rotated 90 degrees to said first linear polarizer, said analyzer excluding the vector sum of said slow and fast transmitted light rays;

a quarter wave retardation or compensation plate for circularly polarizing said transmitted light;

where said transmitted sample light passes from said objective lens through said full wave compensation plate, through said analyzer and then through said quarter wave retardation plate;

an ocular for receiving said transmitted sample light from said quarter wave compensation plate, producing a virtual image of said sample transmitted light and projecting said image onto a faceplate of a CCD camera chip; and

a digital CCD camera for filming a plurality of interference colors produced by said sample transmitted light;

2. A method for training a support vector machine for analyzing and identifying a digitized color image of a of high order refractive indices samples comprising the steps of:

identifying a of high order refractive indices sample population having P classes of unique species;

creating a plurality of image training sets for each of said P classes where each training

set has N images;

calculating a texture feature set vector for each image of each of said training sets,

said texture feature set calculation comprising the steps of

computing Gabor transforms for each of the hue (H), saturation (S) and value (V) channels of each image of said training sets and for S scales and R orientations; and

calculating statistical invariants comprising a mean and standard deviation for

each choice of said scales, said orientations and said HSV channels for each image of said training sets; and

applying a Fourier Transform magnitude transformation of each of said calculated feature set vectors so as to produce a rotation-invariant feature set;

calculating a color feature set vector for each image of each of said training sets, set color feature set calculation comprising the steps of

determining a first zero-delay cross-correlation between color distribution angle

pairs for each of said HSV channels;

determining a second zero-delay cross-correlation between color edge distribution

angles for color edge pairs for each of said HSV channels; and

determining a variance of the zero-delay cross-correlation between HSV standard

deviation color index pairs for each of said HSV channels; and

creating a feature training set matrix for each image training set, said matrix comprised of said texture feature set vectors and said color feature set vectors, said matrix having N×

P rows.

3. The method of claim 2 where P=6, N≧

60, S=4 and R=90.

4. The method of claim 2 further comprising running a leave-one-out optimization of the parameters C and gamma of the support vector machine for a feature training set comprising the steps of:

choosing an integer value of C between 5 and 100;

choosing a half integer value of gamma between 1 and 7;

initializing LOO error to 0;

removing one feature vector from said feature training set matrix;

training said support vector machine on remaining feature vectors of said feature training set matrix;

classifying said removed feature vector with said trained support vector machine;

checking the classification of said classified removed feature vector for error;

determining if said error is present and, if said error is found, increasing said LOO error by 1;

sequentially removing each of said remaining untested feature vectors, one vector at a time, and repeating said removing, said training, said classifying, said checking and said determining steps for each training set of said removed untested feature vectors;

defining a % LOO error as 100 times said LOO error at the end of said sequentially removing step divided by N×

P;

repeating all of the above steps for each integer value of C between 5 and 100 and each integer or half-integer value of gamma between 1 and 7; and

defining an optimum value of C and gamma which are the values of C and gamma that minimize % LOO error.

5. The method of claim 4 where P=6, N≧

60, S=4 and R=90.

6. A method for classifying a digitized color image of of high order refractive indices samples comprising the steps of

calculating a texture feature set vector for each image in a set of M digitized images of an unknown sample of of high order refractive indices sample species, said texture feature set calculation comprising the steps of

computing Gabor transforms for each of the hue, saturation and value channels of each image of said image set and for S scales and R orientations;

and

calculating statistical invariants comprising a mean and standard deviation for

each choice of said scales, said orientations and said HSV channels for each image of said image set; and

applying a Fourier Transform magnitude transformation of each of said calculated feature set vectors so as to produce a rotation-invariant feature set;

calculating a color feature set vector for each image of said image set, said color feature set calculation comprising the steps of

determining a first zero-delay cross-correlation between color distribution angle

pairs for each of said HSV channels;

determining a second zero-delay cross-correlation between color edge distribution angles for color edge pairs for each of said HSV channels; and

determining a variance of zero-delay cross-correlation between HSV standard

deviation color index pairs for each of said HSV channels; and

creating a feature set vector matrix for said unknown sample image set, said matrix comprised of said texture feature set vectors and said color feature set vectors, said matrix having M rows;

inputing optimized values for C and gamma in a support vector machine;

inputing a training feature set in said support vector machine and training said support vector machine;

inputing said unknown sample feature set matrix in said trained support vector machine and classifying said sample feature set matrix; and

providing classified output comprising a set of a pattern classes, each sample image associated with a pattern class, said image pattern class identifying each of said sample images.

7. The method of claim 6 where P=6, N≧

60, S=4 and R=90.