Article 8 – Ornament Problem Suppression in Indonesian License Plate Recognition Systems

Samuel Mahatmaputra Tedjojuwono

Based on the original work of fast performance algorithm in detecting Indonesian license plate, the proposed work will solve the error found in the license plate localization process caused by plate like pattern within the image, which was called the ornament problem. Although not in all cases, this problem could exist when a car has banner, regular pattern, car’s front grill, that could miss understood by the system as license plate letters. The proposed work will implement filtering systems instead of machine learning approach. The filtering methods will follows three steps: detection filter based on the number of elements in the vector, based on the letter proportion of a license plate number, and based on the distance between detected letters. This approach will maintain the fast properties of the original algorithm and will increase the accuracy of localizing the license plate within the given image. Keywords— license plate, localization, computer vision.

In implementing ALPR (automatic license plate recognition) on a real traffic the algorithm implemented needs to be agile and fast. One of the processes of implementing ALPR system is the license plate localization process. The proposed work is based on the previous research that implements fast performance non-learning based algorithm [1]. However the algorithm still suffers from what was proposed in the related work as the ornament problem. The problem exist where contour pattern that has high similarity to a plate alphabet and number. This could negatively impact the license plate localization process. Figure 1 shows the result of wrongly localized license plate as a result of the front grill of the minibus mistakenly translated as the license plate letters. This is as a result of the vehicle’s grills contours have high similarity as the basic contours of a series of license plate letters. The proposed work presents an algorithm to suppress this ornament problem in LPR (license plate recognition) system using interconnected image segmentation and relates them to the proportion of the contours’ bounding box.

The ornament problem in a LPR systems happen when the localization process fail to detect the location of a license plate within a given image due to distraction from other ornaments of the cars. Furthermore, this falsely located license plate was the result of a strong presence of other letters in the sample image that are not part of the license plate letters. The problem usually takes place in a case of vehicle’s banners, front grills, logos, and any other regularly placed contours in a real case traffic

information systems. The propose systems implement filtering algorithm to suppress the so-called ornament problem [1] by dividing the system into three filtering functions. The three functions could be implemented separately, however through some of the experimental result we could see later on that combining the three will improve the localization process of license plate in Indonesian plate systems.

The proposed algorithm follows the following methods to suppress the ornament problems while still maintaining the fast and accurate properties of the original system:

  • Detection filter based on the number of elements in the vector
  • Detection filter based on the letter proportion of a license plate number
  • Detection filter based on the distance between detected letters.

In conclusion, implementing ALPR on a real live traffic system needs a robust algorithm to be able to handle input irregularity. One form of this arbitrary input is what was proposed to be called as the ornament problem. The algorithm proposed implements several filtering to solve the problem. One of which is to validate the ratio of every image segment bounding box width and height. The result is promising and could improve the accuracy of locating license plate letter within the given input as high as 90%. Moreover, as describe at the later part of the manuscript, an ideal size of the input image to the algorithm is yet to be identified. Future work would need to quantify what could be proposed as the reasonable input size, to minimize the missing information as a result of a “too” small image’s segment. Another input irregularity could come in a form of tilted image. Several approaches have been proposed and tested previously. However to the particular algorithm this approach are yet to be proven. Future work needs to identify and solve the tilted image segment problem and combine it with the current approach in order to increase the accuracy of license plate localization method.

For further details please refer to the following paper:

https://iopscience.iop.org/article/10.1088/1757-899X/185/1/012027/pdf