Robust License Plate Detection For Surveillance Images Using Multi-Features Cascading Filtering

With the development of technology and improvement of people’s living standards,intelligent monitoring system is becoming more and more important in security and other fields.And license plate detection accounts for a large proportion in the intelligent monitoring system.Since that license plate is unique identity of vehicle,people pay more attention on it.So designing a fast robust license plate detection algorithm is very pivotal.This paper analyzed several existing license plate detection methods,most of which rely on color and texture information,while these information can be effected easily by light or other problems.What’s more,most of the existing methods can not detect blurred license plate well.We summarized several easily detectable characteristics of the license plate in China,and found common characteristics of clear and blurred license plate,finally worked out a robust license plate detection method for surveillance images using multi-features cascading.We designed pre-detection algorithm based on gradient feature using two characteristics.The first one is that there is a rectangular border with both clear and blurred license plate,and the second one is that the license plate have obviously split sections and division points.The positive samples were divided into three categories according to the clarity of the plate,and their normalized gradient features were extracted,while box sizes were saved.We trained linear SVM classifier using the extracted features and obtained a template representing weight values of different coordinates.We did convolution operation between the template and images to obtain template-matching coefficients,and took the areas having large coefficients.Since there were a lot of repeat counting which made the method slow,we changed convolution into bit "and" and bit "or" operation by an approximate of the template,which greatly reduced the time complexity of the algorithm.There were many "pseudo-plate" areas in the result of pre-detection,and these areas need to be removed.We found the biggest difference between these regions and the license plate is color,so we used main color extraction to excluding wrong areas.First,we converted the image from RGB color space into HSV color space.Then we extracted color histogram,and eliminated the influence of local small peaks using clustering method.Finally we determined three around main colors.If the main colors match the color pairs of license plate,the area is a license plate area.The areas meeting the conditions were merged into limited number of areas according to the degree of overlap,and the merging results are license plate areas.