ALPR Neural Network Using Augmented Synthetic Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic license plate recognition (ALPR) systems are time-consuming, error-prone, and computationally intensive, especially when dealing with tilted, skewed, or blurred license plates, and may fail to detect suspect vehicles in real-time due to resource constraints in embedded devices.
Innovation Solution
An electronic device using an augmented dataset and a trained neural network model to generate synthetic license plate images with varied geometrical and visual parameters, allowing for efficient and accurate recognition of license plates even in challenging conditions, by applying OCR only on extracted plate images matched with the trained model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If OCR technique is applied on each license plate in captured images, then license plate numbers can be extracted, but the process becomes time-consuming and computationally intensive
Solution Approach 1:
The system performs preliminary actions by generating an augmented dataset of synthetic license plate images with various transformations (tilt, skew, blur, pixelation) before actual recognition. A neural network model is trained in advance on this augmented dataset, enabling fast matching during runtime without applying OCR to every captured plate, thus reducing cycle time while maintaining accuracy
Solution Approach 2:
The system creates synthetic copies of license plate images through data augmentation techniques, generating virtual training samples with different geometrical and visual parameters. These synthetic copies are used to train the neural network model, replacing the need for extensive manual labeling and enabling efficient recognition without processing every real captured image through time-consuming OCR
2Measurement precision
If OCR technique is applied on a large number of license plates, then all license plate numbers can be extracted, but computational resources are excessively consumed
Solution Approach 1:
The system creates synthetic copies of license plate images through data augmentation techniques, generating virtual training samples with different geometrical and visual parameters. These synthetic copies are used to train the neural network model, replacing the need for extensive manual labeling and enabling efficient recognition without processing every real captured image through time-consuming OCR
Solution Approach 2:
The system replaces the traditional OCR mechanical process with a neural network-based recognition system. Instead of applying OCR algorithms to every captured license plate (computationally intensive), the pre-trained neural network performs rapid pattern matching, significantly reducing computational resource consumption while maintaining extraction accuracy
3Adaptability or versatility
If conventional ALPR methods are used on tilted, skewed, or blurred license plates, then recognition can be attempted, but accuracy decreases and errors increase
Solution Approach 1:
The system applies parameter changes by transforming the synthetic license plate images with various geometrical parameters (tilt angle, skew angle, rotation) and visual parameters (blur level, pixelation degree, brightness, contrast). These parameter variations are applied during data augmentation to train the neural network model, enabling it to accurately recognize license plates under diverse and challenging conditions
Data Source
AI summary
An electronic device that captures a plurality of images by an image sensor. The electronic device includes control circuitry that generates a plurality of synthetic license plate images for each original license plate templates that are associated with a first hotlist of vehicles. The control circuitry generates a first augmented dataset of synthetic license plate images by an image transformation on the generated plurality of synthetic license plate images. The control circuitry trains a first neural network model using the generated first augmented dataset as a training dataset. The control circuitry detects a first vehicle in the captured plurality of images based on the extracted first license plate image portion and the trained first neural network model. The extracted first license plate image portion includes a first license plate identifier in the first hotlist of vehicles.