# US 2019/0244009 A1 · High recall additive pattern recognition for image and other applications

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2019/0244009 A1
- Document ID: US20190244009A1
- Status: Published application
- Kind: A1
- Application number: 15/889,513
- Filed: 2018-02-06
- Published: 2019-08-08
- Earliest claimed: 2018-02-06
- Assignees: Shutterfly, Inc.
- Applicants: Shutterfly, Inc.
- Inventors: Omer Moshe Moussaffi
- Web page: https://patentgrep.com/patent/US20190244009A1

## CPC

- G06N 20/10: using kernel methods, e.g. support vector machines [SVM]
- G06F 18/24: Classification techniques
- G06N 5/02: Knowledge representation; Symbolic representation
- G06V 10/764: using classification, e.g. of video objects
- G06V 40/161: {Detection; Localisation; Normalisation}
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

A computer-implemented method includes selecting a kernel and kernel parameters for a first Support Vector Machine (SVM) model, testing the first SVM model on a feature matrix T of n feature vectors of length m to produce false positive (FP) data set and false negative (FN) data set by a computer processor, wherein n and m are integer numbers, automatically removing feature vectors corresponding to the FP data set from the feature matrix T by the computer processor to produce a feature matrix T_best of size (n-size(FN))*m, retraining the first SVM model on the feature matrix T_best to produce a second SVM model, and checking if a ratio (T_best sample number)/(SVM support vector number) is above a threshold for the second SVM model on T_best. If the ratio is above the threshold, SVM predictions is performed using the second SVM model on the feature matrix T_best.

## Claims

### Claim 1

A computer-implemented method, comprising:
selecting a kernel and kernel parameters for a first Support Vector Machine (SVM) model;
testing the first SVM model on a feature matrix T to produce false positive (FP) data set and false negative (FN) data set by a computer processor, wherein the feature matrix T includes n feature vectors of length m, wherein n and m are integer numbers;
automatically removing feature vectors corresponding to the FP data set from the feature matrix T by the computer processor to produce a feature matrix T_best of size (n-size(FP))*m;
retraining the first SVM model on the feature matrix T_best to produce a second SVM model;
checking if a ratio (T_best sample number)/(SVM support vector number) is above a threshold for the second SVM model on T_best; and
if the ratio is above the threshold, performing SVM predictions using the second SVM model on the feature matrix T_best.

### Claim at position 2

2-14: (canceled)

### Claim 15 (depends on 1)

The computer-implemented method of claim 1, wherein the n feature vectors are based on faces or objects in images, the method further comprising:
classifying the faces or the objects in the images using the second SVM model on the feature matrix T_best.

### Claim 16 (depends on 15)

The computer-implemented method of claim 15, further comprising:
creating an image-product design based on the faces or the objects in the images classified by the second SVM model using the feature matrix T_best.

### Claim 17 (depends on 1)

The computer-implemented method of claim 1, further comprising:
if the ratio (T_best sample number)/(SVM model support vector number) is not above the threshold for the SVM model on T_best, repeating the steps of selecting, testing, automatically removing, retraining, and checking to find a third SVM model having (T_best sample number)/(SVM model support vector number) ratio that exceeds the threshold.

### Claim 18 (depends on 17)

The computer-implemented method of claim 17, further comprising:
performing SVM predictions using the third SVM model on the feature matrix T_best.

### Claim 19 (depends on 17)

The computer-implemented method of claim 17, further comprising:
if a SVM model having (T_best sample number)/(SVM model support vector number) ratio that exceeds the threshold is not found, selecting a fourth SVM model that yields a highest (T_best sample number)/(SVM model support vector number) ratio in the step of repeating the steps of selecting, testing, automatically removing, retraining, and checking; and
performing SVM predictions using the fourth SVM model on the feature matrix T_best.

### Claim 20 (depends on 1)

The computer-implemented method of claim 1, wherein the threshold for the (T_best sample number)/(SVM model support vector number) ratio is ten or higher.

### Claim 21 (depends on 1)

The computer-implemented method of claim 1, further comprising:
applying Grid Search on the feature matrix T to find optimal kernel parameters before the step of testing the first SVM model on a feature matrix T.

### Claim 22 (depends on 1)

The computer-implemented method of claim 1, wherein the feature vectors are based on faces or objects in images, the computer-implemented method further comprising:
classifying the objects and the faces in the images by performing SVM predictions on the feature matrix T_best.

### Claim 23 (depends on 22)

The computer-implemented method of claim 22, further comprising:
automatically creating designs for photo products based on the objects and faces classified by performing SVM predictions on the feature matrix T_best.

### Claim 24 (depends on 1)

The computer-implemented method of claim 1, further comprising:
selecting a feature vector v from the FP data set to add back to the feature matrix T_best to produce a feature matrix T_v;
retraining the second SVM model on the feature matrix T_v to produce a fifth SVM model;
calculating a benefit function to produce a benefit function value, wherein the benefit function is dependent of differences between true positives, true negatives, and numbers of support vectors generated on T_best by the second SVM model and on T_v respectively by the fifth SVM model;
adding the feature vector v to T_best if the benefit function value meets a predetermined criterion; and
performing SVM prediction using the fifth SVM model on the feature matrix T_best.

### Claim 25 (depends on 24)

The computer-implemented method of claim 24, wherein the n feature vectors are based on faces or objects in images, the method further comprising:
classifying the faces or the objects in the images using the fifth SVM model on the feature matrix T_best.

### Claim 26 (depends on 25)

The computer-implemented method of claim 25, further comprising:
creating an image-product design based on the faces or the objects in the images classified by the fifth SVM model using the feature matrix T_best.

### Claim 27 (depends on 24)

The computer-implemented method of claim 24, further comprising:
repeating the steps of selecting a feature vector v, training the second SVM model, performing SVM predictions, calculating a benefit function, and adding the feature vector v Page 6 to T_best by selecting and adding a different feature vector from the FP data set to feature matrix T_best, wherein SVM predictions are performed using the feature matrix T_best that gives a highest best function value.

### Claim 28 (depends on 27)

The computer-implemented method of claim 27, wherein all the feature vectors corresponding to the FP data set are evaluated by calculating a corresponding benefic function and determined to be added to the T_best or not depending on a value of the corresponding benefic function.

### Claim 29 (depends on 24)

The computer-implemented method of claim 24, wherein the feature vector v is not added to the feature matrix T_best if the value of the benefit function does not meet the predetermined criterion.

### Claim 30 (depends on 24)

The computer-implemented method of claim 24, wherein the feature vectors are based on faces or objects in images, the computer-implemented method further comprising:
classifying the objects and the faces in the images by performing SVM predictions on the feature matrix T_best.
