<p>In general, the extraction of the vehicle plate is a previous step of plate recognition, and it actively studied for several decades. Plate localization is used in various security and traffic applications. In this paper, the proposed method is efficient to localize a plate for the multinational countries. The proposed method consists of three levels. The first level is the preprocessing<strong> </strong>that contains several steps. The digital camera capture images have been taken about twenty meters from the car with zooming two to three meters. Images are resampled using the zooming technique (bilinear interpolation) that makes the dimension of image (1024 x 768) pixels. The resampled images are resized to (940x 730). These images are converted to grayscale (green channel), and the bilateral filter is applied for removing noise. The second level is plate detection that adopts morphological operations, image subtraction, and vertical edge detection (Sobel). At the last Connected component analysis and Hough transform are used. The third level is the Deskew plate that notifies the plate may skew so that Hough transform is used to detect the largest line segment. Then, the images are rotated using bilinear interpolation. About 860 images are tested for different countries (Iraq, Belarus, Armenia, Hungary), and the accuracy is 98.99 % for extraction of the plate and 100% for the Deskew plate. Thus, the proposed system shows high efficiency in achievement.</p>
In the classification of license plate there are some challenges such that the different sizes of plate numbers, the plates' background, and the number of the dataset of the plates. In this paper, a multiclass classification model established using deep convolutional neural network (CNN) to classify the license plate for three countries (Armenia, Belarus, Hungary) with the dataset of 600 images as 200 images for each class (160 for training and 40 for validation sets). Because of the small numbers of datasets, a preprocessing on the dataset is performed using pixel normalization and image data augmentation techniques (rotation, horizontal flip, zoom range) to increase the number of datasets. After that, we feed the augmented images into the convolution layer model, which consists of four blocks of convolution layer. For calculating and optimizing the efficiency of the classification model, a categorical cross-entropy and Adam optimizer used with a learning rate was 0.0001. The model's performance showed 99.17% and 97.50% of the training and validation sets accuracies sequentially, with total accuracy of classification is 96.66%. The time of training is lasting for 12 minutes. An anaconda python 3.7 and Keras Tensor flow backend are used.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.