Research on the Optimal Machine Learning Classifier for Traffic Signs

Wang, Boyu · 2022 · OpenAlex-citations

DOI: 10.1051/shsconf/202214403014

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the critical need for accurate traffic sign recognition in autonomous driving systems, motivated by projected significant increases in global passenger and freight traffic by 2030 and 2050. The primary research objective is to identify the optimal machine learning classifier for traffic sign classification by comparing three specific algorithms: Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Logistic Regression (LR). Additionally, the paper investigates how initial image preprocessing techniques, specifically binarization and sharpening, impact classification accuracy. The methodology utilizes a dataset of fifteen distinct traffic sign classes (Class0 to Class14) stored in Portable PixMap (PPM) format. The study employs Python and the OpenCV library for data processing. Two preprocessing methods were evaluated: image binarization, which sets pixel gray values to 0 or 255 based on a threshold, and image sharpening using a Laplacian operator to enhance edge details and contrast. The classifiers were configured with specific parameters: LR used Multinomial Logistic Regression with the stochastic average gradient solver; MLP utilized an incremental algorithm to determine the optimal number of neurons (settling on 50 neurons in one hidden layer); and SVM employed the LinearSVC method with a one-vs-the-rest strategy for multi-class classification. The models were trained on a dataset of over 15,000 images and validated using cross-validation techniques. The results indicate that image binarization had negligible impact on performance, yielding accuracy rates around 95% similar to unprocessed images. In contrast, image sharpening significantly improved classification accuracy across all models, boosting performance by approximately 2%. Without sharpening, SVM achieved the highest accuracy at roughly 98%, while LR and MLP performed considerably worse. After applying sharpening, SVM reached 99.8% accuracy, LR achieved 99.2%, and MLP approached 99%. Although cross-validation suggested LR performed best in that specific statistical test, SVM demonstrated more stable and accurate results during standard testing. The study also noted that MLP struggled with distinguishing similar classes, such as speed limit signs for 100 and 120, particularly under low brightness conditions. The study concludes that SVM is the optimal classifier for this specific traffic sign dataset due to its consistent high performance and stability, particularly when combined with image sharpening. LR is identified as a strong alternative, performing nearly as well as SVM when images are sharpened. MLP was found to be the least effective, requiring additional processing to reduce loss and showing lower accuracy for small attribute sets. The authors acknowledge that the training data may not fully represent real-world driving conditions, suggesting future research should incorporate more diverse environmental scenarios to improve generalization.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• sign visibility legibility