Automatic License Plate Localization Using Intrinsic Rules Saliency

This paper addresses an intrinsic rule-based license plate localization (LPL) algorithm. It first selects candidate regions, and then filters negative regions with statistical constraints. Key contribution is assigning image inferred weights to the rules leading to adaptability in selecting saliency feature, which then overrules other features and the collective measure, decides the estimation. Saliency of rules is inherent to the frame under consideration hence all inevitable negative effects present in the frame are nullified, incorporating great deal of flexibility and more generalization. Situations considered for simulation, to claim that the algorithm is better generalized are, variations in illumination, skewness, aspect ratio and hence the LP font size, vehicle size, pose, partial occlusion of vehicles and presence of multiple plates. Proposed method allows parallel computation of rules, hence suitable for real time application. The mixed data set has 697 images of almost all varieties. We achieve a Miss Rate (MR) = 4% and False Detection Rate (FDR) = 5.95% in average. Also we have implemented skew correction of the above detected LPs necessary for efficient character detection.This paper addresses an intrinsic rule-based license plate localization (LPL) algorithm. It first selects candidate regions, and then filters negative regions with statistical constraints. Key contribution is assigning image inferred weights to the rules leading to adaptability in selecting saliency feature, which then overrules other features and the collective measure, decides the estimation. Saliency of rules is inherent to the frame under consideration hence all inevitable negative effects present in the frame are nullified, incorporating great deal of flexibility and more generalization. Situations considered for simulation, to claim that the algorithm is better generalized are, variations in illumination, skewness, aspect ratio and hence the LP font size, vehicle size, pose, partial occlusion of vehicles and presence of multiple plates. Proposed method allows parallel computation of rules, hence suitable for real time application. The mixed data set has 697 images of almost all varieties. We achieve a Miss Rate (MR) = 4% and False Detection Rate (FDR) = 5.95% in average. Also we have implemented skew correction of the above detected LPs necessary for efficient character detection.