Parallelization of license plate localization on GPU platform

Automatic license plate recognition which has many applications in intelligent transportation systems has three main steps, License Plate Detection (LPD), segmentation, and character recognition. The first step, LPD is the main step. This is because the accuracy of other steps depends on this step. While LPD has many challenging problems such as illumination changing, weather conditions, vehicle position, number of cars in each image, etc. In this paper, an algorithm for LPD is proposed. In the proposed technique, to remove noises from input images mean filtering is used. After that, the differences between the filtered image and input image are computed and it is converted to a black and white image. Finally, after applying edge detection and closing and opening morphological operations, license plate candidates are detected. The experimental results on a real Iranian dataset show that about 96.16% of license plates are localized. In addition, results show that the mean filter which has been used for noise removal is the most time-consuming kernel in the proposed algorithm. This important kernel on the GPU platform using CUDA parallel programming model is implemented. The experiments show that speedups of up to 14.54x and 10.77x for kernel- and application-level are achieved.

Automated summary

Automatic license plate recognition in intelligent transportation systems involves three main steps, including license plate detection, segmentation, and character recognition. The accuracy of the other steps relies on the success of the license plate detection step, which can be challenging due to various factors. A proposed algorithm in this paper uses mean filtering to remove noise from input images, computes the differences between the filtered image and the input image, and applies edge detection and morphological operations to detect license plate candidates. Experimental results on a real Iranian dataset show a 96.16% localization success rate.