Plate Number Localization Using a Semi-Hybrid Genetic Algorithm and a Novel Binarization Technique

What is it about?

License plate (LP) number detection and recognition are very important modules in all intelligent transportation systems. The processes of detection and recognition of the LP symbols are carried out in several steps. The most important and difficult step is the binarization step, especially if the localization is based on the connected component analysis technique where each symbol should be separated from its neighbors and from the LP frame. The connection of symbols to their neighbors or to the LP frame may be due to physical reasons such as bolts, nuts or dirt, or it may be due to illumination conditions and camera to LP pose. The problem has been partially solved in a previous system by allowing the genetic algorithm to skip a limited number of connected symbols but when this number is relatively high it results in missing the LP or incorrectly localizing some of LP symbols. To overcome the illumination conditions that cause symbols to be connected to each other some researchers give different solutions based on the LP application by changing the problem variables to maximize the detection rate based on the application domain. By introducing a more general solution for the binarization technique a wider range of the application domain is achieved. This is accomplished by introducing a new variable window adaptive technique that tries to separate the connected symbols that are output from a fixed size adaptive technique. Another important criterion in the localization process is the duration of the process which should be performed in a period less than 0.1 seconds to support real-time detection and recognition. The duration of the detection depends mainly on the genetic algorithm (GA) which selects a sequence of L-objects from a large number (M) of connected component objects based on an objective distance from the geometrical-representation of the prototype LP. By partitioning the whole population Z into G (≈M/L) subgroups where each subgroup deals with a subset of the M objects, the search space for the GA is reduced to approximately M×(3.5L) raised to the power of (L-1) (linear function of M) instead of (M-L/2) raised to the power of L (exponential function of M) in case of a single group. Hence a substantial improvement is achieved by the subgrouping and the limitation of allocated objects for each subgroup which based on a hashing function that is proposed to give LP symbols hashing values near each other while being far from values for non-LP objects. Many other enhancements of the GA components are introduced to increase the system accuracy such as the objective distance parameters and the geometrical representation matrix that model the LP prototype. Since the introduced binarization technique solves the symbol connection problem related to illumination conditions, further enhancement is introduced in the skipping ability of the GA to overcome physical problems by performing an exhaustive search after the completion of each GA run to reach the optimum state of the skipping positions (based on some heuristic rules) to separate physically connected symbols

Featured Image

Photo by Maria Lysenko on Unsplash

Photo by Maria Lysenko on Unsplash

Why is it important?

By the introduced enhancements we improve both accuracy and speed of the LP detection process and make it independent on the vehicle identification application without resorting to select different system parameters for each application category as reported in the research paper of the creator of the used public dataset.

Perspectives