Choose proper image abstraction or feature for sign recognition

In order to develop a system that recognize and classify the different types of input patterns, a number of features or abstractions of a source image should be computed first and tested for the classification. However in practice, it is difficult to find out how many feature extractors should be applied to categorize many different types of images automatically.

In our approach, it can be easily understood that one single feature of an input image may not be applicable to categorize 328 different types of traffic sign patterns. So we may need to employ several different types of feature extracting algorithms to achieve this goal. Before leaving this assumption as a true, a new feature correlation visualization method is introduced in this work.

A special graph, called the feature correlation graph (FCG), is a 2-dimenional square image in which the width and height is the number of patterns to recognize, the coordinate of each pixel represents the correlation between features extracted from ith pattern and jth pattern, and the brightness of each pixel represent the correlation strength (0 represents the perfect matching and 255 for no correlation). So the brighter graph tells you that you have a good measure or classifier for the specified different types of input patterns.

Template correlation measures in 2D images for 328 traffic sign patterns. It uses the L2 norm (RMS) difference.

Templatematching.png Twit2.gif

X-Y profile correlation measures in 2D images for 328 traffic sign patterns. It uses the L2 norm (RMS) difference.

Image:xyprofileprojectioncorrelation.png

Canny edge template correlation measures in 2D images for 328 traffic sign patterns. It uses the L2 norm (RMS) difference.

Image:edgetemplatematching.png

Canny edge X-Y profile correlation measures in 2D images for 328 traffic sign patterns. It uses the L2 norm (RMS) difference.

Image:cannyedgexyprofilecorrelation.png

Line receptor correlation measures in 2D image for 328 traffic sign patterns. It uses the count of matching states between two line receptor crossing state.

Image:linereceptorcorrelation.png

Here is the project with source codes compiled with VC6 to generate FCG graphs.

Traffic Sign Color Processing

One things that is not considered in above approaches is the color feature of the traffic sign. Humans recognize the traffic sign mainly based on its color and shape features. Above features are focused on the sign shape, not its colors. Color analysis is very important and will be handled in details at Traffic sign color information processing.