Improving Vehicle Identification Number Detection Accuracy with YOLOv5 and Histogram Equalization

Abstract

This study examines the effectiveness of different image preprocessing techniques for object detection models, using a dataset of VIN images from Roboflow. The dataset was segmented into training, validation, and testing subsets, encompassing a range of conditions such as noise, rain, varying lighting, and reflections. Model performance was evaluated through metrics including precision, recall, average precision (AP), mean average precision (mAP), error rate reduction, and frames per second (FPS).The baseline model, trained on the original dataset, achieved a precision of 97.9% and a recall of 95.7%, with an mAP@0.5 of 99.1% but a lower mAP@0.5:0.95 of 62.3%. Applying Histogram Equalization (HE) resulted in improved recall but reduced precision, with mAP@0.5:0.95 values remaining comparable to the original dataset. The HE+RGB preprocessing showed minor performance changes, with inconsistent improvements in recall and precision. Adaptive Histogram Equalization (AHE) notably improved model performance, reaching a precision of 98.8% and recall of 99.6%, with mAP@0.5 and mAP@0.5:0.95 values of 74.3%, 77.0% respectively. The CLAHE preprocessing technique outperformed all others, achieving the highest precision (99.4%), recall (98.6%), and mAP@0.5:0.95 (75.2% in training, 77.9% in validation, and 75.2% in testing), demonstrating the best balance of accuracy and generalization with minimal misclassifications. Overall, CLAHE emerged as the most effective preprocessing method, offering superior performance across all evaluation metrics.