- The paper presents a multi-stage automatic number plate recognition (ANPR) system integrating image preprocessing, number plate localization, character segmentation, and character recognition.
- The methodology employs specific techniques like Sobel vertical edge detection for localization and vertical projection for segmentation to isolate characters effectively.
- Using a dataset of Croatian vehicle images, the Random Forest Classifier achieved the highest character recognition accuracy at 90.9%, outperforming k-NN, Neural Network, and SVM due to its robustness and multi-class capability.
The paper "Automatic Number Plate Recognition using Random Forest Classifier" presents a system for automatic recognition of vehicle number plates, which is crucial for intelligent transportation systems (ITSs) aimed at traffic management, monitoring of traffic violations, and maintaining security in areas with restricted access. The proposed methodology involves dividing the process into several distinct stages: Pre-Processing, Number Plate Localization, Character Segmentation, and Character Recognition.
Methodology:
- Pre-Processing:
- The methodology starts with the conversion of RGB images to grayscale to reduce computational complexity and enhance contrast, employing Contrast Limited Adaptive Histogram Equalization (CLASH).
- Noise is minimized using a bilateral filter with a kernel size of 5, effectively preserving the edge details.
- The grayscale image is then binarized and dilated to emphasize the edges and mitigate false positives during edge detection.
- Number Plate Localization:
- The paper employs Sobel vertical edge detection instead of Canny edge detection to focus on vertical edges while ignoring irrelevant horizontal structures inherent in the vehicle, such as windshields and bumpers. This method efficiently distinguishes the region with the highest concentration of edges, which corresponds to the number plate.
- Character Segmentation:
- Upon localization, the extraneous noise around the characters is removed, employing vertical projection of Sobel edge-detected images to eliminate non-character noise regions, ensuring only the character region peaks are retained for segmentation.
- Character Recognition:
- Random Forest Classifier (RF) is utilized for recognizing characters extracted from the segmented number plate. RF leverages an ensemble of decision trees, enhancing the system’s resilience to noisy data due to the robustness of ensemble predictions.
- Various other classifiers including k-nearest neighbor (k-NN), Neural Network, and Support Vector Machines (SVMs) were evaluated, but RF achieved the highest accuracy at 90.9%.
Results and Discussion:
The application, developed in Python and supported by libraries such as OpenCV, numpy, and scikit-learn, was tested using a dataset of Croatian vehicle images. With Random Forest Classifier, the system demonstrated superior performance, effectively reducing ambiguity between characters such as 'G' and '6', and achieving an accuracy of 90.9% — the highest among the classifiers tested. The system also successfully handled ambiguities associated with indistinct characters due to proper noise management and robust character segmentation.
The paper points out that Random Forest outperforms SVM in this application due to its intrinsic multiclass capabilities and ensemble nature, which provide robustness in noisy environments and parallelizable training, unlike SVMs. The paper illustrates that each classifier (k-NN, NN, SVM, and RF) has distinctive strengths, yet RF’s integration of multiple decision trees for consensus results in optimal overall recognition accuracy.
Conclusion:
The presented system demonstrates an efficient automatic number plate recognition pipeline integrating preprocessing, robust edge-based localization, effective segmentation, and comprehensive character recognition. The Random Forest Classifier emerges as the most effective tool within the comparative framework for this task, capitalizing on its multi-class classification and noise resilience capabilities. The method's accuracy and efficiency make it suitable for deployment in real-world ITS applications.