Method for visual image detection

• US 20120224769A1
• Filed: 01/24/2012
• Published: 09/06/2012
• Est. Priority Date: 01/24/2011
• Status: Active Grant

• Current Assignee: Two Hat Security Research Corp.
• Original Assignee: Ashok Ramadass, Dilip Raghavan, Steven White
• Inventors: White, Steven, Ramadass, Ashok, Raghavan, Dilip
• Granted Patent: US 8,644,606 B2
• Time in Patent Office: Days
• US Class Current: 382/165
• CPC Class Codes: G06V 20/35 Categorising the entire scene, e.g. birthday party or wedding scene
• CPC Class Codes: G06V 40/10 Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands

Abstract

The method uses several steps to collect, analyze, compare, and flag an image for inappropriate content. The image is first collected based on detection and is analyzed via a database of inappropriate images. The collected image is scanned for skin-tone and/or texture, then the method eliminates the background of the image. The method then analyzes the curvature of the image. The next step is body part detection for inappropriate body parts. The image is then given a score based on all of the above steps. If the score is above a point, the image is flagged. If it is below a preselected point, the image is released. If the score is between the two points, it is withheld for further analysis.

Claims

1. A method for visual image detection using a computer readable medium comprising:

scan an example image for later matching image purposes and run the visual image detection process to calculate hash values for the example image;

input a new image for matching detection process;

scan new image for shape configuration detection wherein the shapes orientation and distance relation to each other is utilized when the image has a predetermined requirement of the shape relationships;

perform color detection in the feature descriptors for user determined colors and shades pre-selected from the color spectrum;

eliminate background sections of the image without feature descriptors are eliminated from future scanning;

eliminate curvature sections for curves on the edges of the feature descriptors where the noncompliant curvature sections of the features are eliminated;

determine hash values for the remaining features using the fast algorithm combined with the sift algorithm and comparing those values with hash values of images from a host database that were calculated in the same fashion;

select images that satisfy the user determined hash value similarity requirements.

2. The method according to claim 1, wherein the example image is a database of multiple example images.

3. The method according to claim 2, wherein the database comprises multiple images of nude body parts including breasts and genitalia.

4. The method according to claim 2, wherein the database comprises copyright protected images.

5. The method according to claim 1, wherein the input image is bilaterally filtered.

6. The method according to claim 1, wherein edge mapping has been performed.

7. The method according to claim 1, wherein the shape configuration detection is for facial shape configuration.

8. The method according to claim 1, wherein the color detection is skin-tone colors.

9. The method according to claim 1 further comprising detecting body limbs from the image following eliminating noncomplying curvature.

10. The method according to claim 9 further comprising eliminating non body limb area features form body part detection.

11. The method according to claim 1, wherein the hash value detection further comprising:

using the fast algorithm to extract features including corner features from the image and eliminates other edge features from the image;

using the sift algorithm which computes the extracted features from the fast algorithm;

clustering the feature descriptors into a range value using the K-means clustering function;

computing a unary feature for each descriptor;

creating a square matrix with the descriptor values;

selecting random projections from matrix that contains a bit hash value;

computing hash values and random projection values and serving into corresponding buckets.

12. The method according to claim 11, wherein the descriptor range is a 0 to 128 value.

13. The method according to claim 11, wherein the square matrix has a dimension of 128×

128.

14. The method according to claim 11, wherein number of random projections selected is 20.

15. The method according to claim 11, wherein the random projections are chosen so that the maximum number of squares is selected with reduced collisions.

16. The method according to claim 11, wherein the random projection outputs a 36 bit hash value.

17. The method according to claim 11, the resulting hash value is a 36×

20 value.

18. The method according to claim 11, wherein there are 20 buckets.