Docstoc

4936481-adaboost-plate

Document Sample
4936481-adaboost-plate Powered By Docstoc
					Learning-based License Plate Detection on Edge Features

Presenter: Ho Wing Teng Authors: Mr. Ho Wing Teng, Mr. Yap Wooi Hen, Dr. Tay Yong Haur Computer Vision and Intelligent Systems (CVIS) Group Universiti Tunku Abdul Rahman, Malaysia

UNIVERSITI TUNKU ABDUL RAHMAN

Introduction
Presents license plate detection in unconstrained environment (cluttered scenes, changing illumination, in-plane and out-plane rotation of license plates) Adaboost learning-based method is implemented by Viola & Jones in their face detector.
[Viola, P. & Jones, M. (2004). Robust Real-Time Face Detection. International Journal of Computer Vision 57(2), 137-154.]

We present comparative results of our approach against two experiments:
– –

License plate image Intensity value. License plate image with edge detection

UNIVERSITI TUNKU ABDUL RAHMAN

LPD System Architecture

Original Image Convert to gray-scale image Canny Edge-detection License Plate Detector Result

UNIVERSITI TUNKU ABDUL RAHMAN

Image Pre-processing
Edge Detection – pre-process the original image before the detection – An edge represents the boundary of an object – used to identify the shapes and area of the particular object – remove some of the noises and make the license plate text edges more visible
UNIVERSITI TUNKU ABDUL RAHMAN

Canny Edge Detection (cont) Canny Edge Detection – Canny edge detector is one of the most popular edge detection methods – it provides optimal detection with no false detection, – better localization with minimum difference within the actual edge position and the detected edge

UNIVERSITI TUNKU ABDUL RAHMAN

Examples
Artificial License Plate without using canny edge detection.

Artificial license plate using canny edge detection.
UNIVERSITI TUNKU ABDUL RAHMAN

Test Samples

UNIVERSITI TUNKU ABDUL RAHMAN

UNIVERSITI TUNKU ABDUL RAHMAN

Viola & Jones Face Detection Architecture
Adaboost Learning Algorithm
Training Samples Feature Selection Features Cascading Final Classifier

Haar-like Features

1

T F

2

T F

3

T F

N

T F

UNIVERSITI TUNKU ABDUL RAHMAN

Feature Types used in Face Detection by Viola & Jones Haar-like Features

summation within the black region and from the pixels summation within the white region

• Feature can be computed from the subtraction of pixels
UNIVERSITI TUNKU ABDUL RAHMAN

Haar-like Features 384

288
Resolution Size

45 12

UNIVERSITI TUNKU ABDUL RAHMAN

White Region Black Region • Feature can be computed

from the subtraction of pixels summation within the black region and from the pixels summation within the white region

UNIVERSITI TUNKU ABDUL RAHMAN

Haar-like Features

• Variant of Features available to the adaboost training

UNIVERSITI TUNKU ABDUL RAHMAN

Schematic diagram of a 4 stages cascaded classifiers

Every processing node is a strong classifier Initially, large number of negative windows is rejected with very small processing time. Subsequent nodes eliminate additional negative windows with additional UNIVERSITI TUNKU ABDUL RAHMAN processing time

Result & Discussion

UNIVERSITI TUNKU ABDUL RAHMAN

Result & Discussion

UNIVERSITI TUNKU ABDUL RAHMAN

Conclusion and Future Works
Proposed Adaboost learning based method to construct LPD on Haar-like features Edge information has demonstrated better discriminative power compared to intensity information The total time costs only ~ 80 ms to process an image of 320 x 240 further improvement is necessary to reduce false alarm rate as well as increasing true positives
UNIVERSITI TUNKU ABDUL RAHMAN

How to calculate Integral Image
In the example to calculate the sum pixels in D, we need to determine the integral image of A, B, C, and D which is represented in number 1,2,3,and 4 in figure 3.5. Number 1 is the total of the pixel values in area A; Number 2 is the total of the pixels in area A and area B; Number 3 is the sum of the pixels in area A and area C; at last the number 4 will be the entire sum of the whole image which is the total of A+B+C+D. Therefore to get the total pixel in D, we can calculate with 4+1-(2+3).

UNIVERSITI TUNKU ABDUL RAHMAN

Formulae to calculate number of features available for adaboost training

Formula calculate number of features for an image size w x h

Figure 3.4 Formula calculate number of rotated features for an image size w x h

UNIVERSITI TUNKU ABDUL RAHMAN

•Table adopted from viola paper summarize the boosting algorithm for learning a query online.

UNIVERSITI TUNKU ABDUL RAHMAN

AdaBoost Training Process Diagram
Training Samples Positive Samples Negative Samples Assign to Training Samples with weights Start training Generate Large Feature Set

Calculate feature value on each training samples

Haar-like Feature

Update the weight in Training Samples yes The error < maxError

Calculate the Error value for each features Determine Threshold values Select the feature with lowest error Save the Selected feature To a text file

End training

No

UNIVERSITI TUNKU ABDUL RAHMAN


				
DOCUMENT INFO
Shared By:
Categories:
Stats:
views:57
posted:6/18/2009
language:English
pages:21