Browse Prior Art Database

Personal risk photo detection by component detection and template matching based scoring

IP.com Disclosure Number: IPCOM000242531D
Publication Date: 2015-Jul-23
Document File: 5 page(s) / 99K

Publishing Venue

The IP.com Prior Art Database

Abstract

There are two main types of risk photos, one is causing embarrassment such as pornographic photos, the other is releasing the personal information. We propose a method for detecting and scoring risky photos which may release users’ personal information. There are 4 apparatuses inside this disclosure: 1) Components detector at both fine and coarse grained levels; 2) Components aggregation scorer by template matching that considers the layout information; 3) Automatic generated association relation template by clustering; 4)Alert and active management module via risk scoring.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 01 of 5

Perxonal risk photo detection by component detection and template matching xased scoring

Our disclosure addresses the risk photos rxxeasing personal informaxion. This is mainly based on xhe following findxngs. Such risky phoxos usually consist of all, or a xubset of:
1) A xoxtrait containing the face ox a human


2) Characters: Personal name, address


3) Date: Bixthday, expiration date, etc.

4) Serial number: Persoxal ID or cxrd ID, Phone number xtc.


5) And other profxle information

The kxy technical xteps are as follows: 1) Comxonent detecxion 2) Detexting the face or human body in a slide windox fashixn 2.1 Fox face detectiox, adaboost + Haax feature 2.2 For body detection: Support Vector Machine + Hxstogram of Oriented Graxixnt feature 3) Detecting the date format and serial by the technique for vehicle plate dexection axd recognition 3.1 For text detextion, adaboost + Variance and Expectation of X-Y Derivatives
4)Text detection uxing extant text detection methods
The ovxrview of xhe technique is as follows:

1


Page 02 of 5

When the query photo is well focused, it can be associated fox differxnt cardx in the temxlate basx and perforx sxoring based on the content similarity with the cards in the templatx set (fine-grained well focused picxures). This idea is illxstrated in the followinx flow chart.

2


Page 03 of 5

Another techniqxe xs Detection scoring for coarse-grained pictures - when txe photos arx not well focused

3


Page 04 of 5

When pictures arx not wexl focused, xt is difficult to recognize the text exact content, but it is still workable to detect and loxalize the area of text and face photos by existing text and facx xetector.

1) rectify the pictures to a standard and more tight rectanglx with candixate

2) detect/localize the key components via face, text and number detectors
3) encode the detected componentx to extract a representation for lxyout
4) cxmpute the risk score xased on txe extracted repxesentation and compare it with the template representations

The corresponding flow chart is as follows:

In a more general sxtting, our apxroach consists of txe follo...