[Paper Review] License Plate Recognition with Compressive Sensing Based Feature Extraction

A. Jokic, Nikola Vukovic|arXiv (Cornell University)|Feb 7, 2019

Vehicle License Plate Recognition16 references6 citations

TL;DR

This paper proposes a license plate recognition system using compressive sensing for dimensionality reduction and feature extraction, enabling efficient classification with fewer samples. By applying a random Bernoulli measurement matrix and total variation minimization for reconstruction, the method achieves 99.08% average accuracy with only 96 compressed samples, demonstrating robust performance even at low sampling rates.

ABSTRACT

License plate recognition is the key component to many automatic traffic control systems. It enables the automatic identification of vehicles in many applications. Such systems must be able to identify vehicles from images taken in various conditions including low light, rain, snow, etc. In order to reduce the complexity and cost of the hardware required for such devices, the algorithm should be as efficient as possible. This paper proposes a license plate recognition system which uses a new approach based on compressive sensing techniques for dimensionality reduction and feature extraction. Dimensionality reduction will enable precise classification with less training data while demanding less computational power. Based on the extracted features, character recognition and classification is done by a Support Vector Machine classifier.

Motivation & Objective

• To reduce computational complexity and training data requirements in license plate recognition systems.
• To improve classification efficiency by leveraging compressive sensing for dimensionality reduction.
• To evaluate the performance of compressive sensing features in character recognition under varying sampling rates.
• To demonstrate robustness of the method using a multiclass SVM classifier on digit images.

Proposed method

• The system uses adaptive thresholding and connected component analysis for image segmentation to isolate individual characters.
• Each segmented character is resized to a fixed 16×16 pixel size to ensure scale invariance.
• Compressive sensing is applied via a random Bernoulli measurement matrix to generate m compressed samples from the N-pixel image, forming a measurement vector y = Ax.
• Image reconstruction is performed using total variation (TV) minimization to recover the original image from compressed measurements.
• The compressed samples are used directly as features for classification, bypassing full reconstruction.
• A multiclass SVM with error-correcting output codes and a one-vs-one coding scheme classifies the features into digit classes (0–9).

Experimental results

Research questions

• RQ1 Can compressive sensing features extracted from low-sampled images achieve high classification accuracy in license plate recognition?
• RQ2 How does the number of compressed samples affect recognition performance and reconstruction quality?
• RQ3 Is the proposed method robust to low sampling rates, even when image reconstruction is visually poor?
• RQ4 How does the performance of compressive sensing-based features compare to traditional feature extraction in terms of accuracy and computational cost?

Key findings

• With 64 compressed samples, the system achieved an average classification accuracy of 98.81%, with a minimum of 97.52% and maximum of 100% across 20 runs.
• Using 96 samples improved performance to an average accuracy of 99.08%, with a minimum of 98.02%.
• The method maintained high accuracy (98.2% average) even with only 32 samples, despite near-unrecognizable image reconstruction.
• The F1 score for digit '4' reached 100% with 96 samples, indicating perfect precision and recall.
• The classifier achieved over 99% F1 score for digits '0', '7', '8', and '9' with 96 samples, demonstrating strong class-specific performance.
• The use of a simple random Bernoulli matrix reduced computational complexity, making the method efficient for real-time applications.