OCR-Free USDOT Number Localization via Sliding Window Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated camera-based systems for recognizing USDOT numbers on commercial vehicles face challenges in localization accuracy due to variations in font, size, and poor image quality, especially under low contrast and noisy conditions, leading to low detection performance.
Innovation Solution
The system employs an OCR-free approach using sliding windows and a trained classifier to detect the 'USDOT' text, with an offline phase creating a dataset of positive and negative samples to account for font and size variations, and an online phase performing preprocessing and non-maximum suppression to enhance localization accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If OCR-based localization is used to identify USDOT numbers, then the system can handle high quality images with clear text, but detection performance deteriorates in noisy, low contrast images captured at night or under poor lighting conditions
Solution Approach 1:
The patent replaces the OCR-based localization approach with a machine learning classifier that uses sliding windows to detect USDOT tags. Instead of relying on optical character recognition which fails in poor image quality, the system trains a classifier on positive samples (containing USDOT tags) and negative samples (without tags) to robustly identify tag locations even in noisy, low contrast nighttime images captured by NIR cameras.
Solution Approach 2:
The patent transforms the localization problem from text recognition to classification by changing the detection parameters. Rather than attempting to recognize characters through OCR, the system converts image patches into feature vectors and applies a trained classifier that has learned to distinguish USDOT tag regions from other text regions based on visual patterns, enabling reliable detection under varying illumination and image quality conditions.
2Adaptability or versatility
If the system accounts for variations in font and size of USDOT tags, then the localization becomes more robust, but the complexity of the detection system increases
Solution Approach 1:
The patent performs preliminary action by pre-training a machine learning classifier offline using extensive positive and negative sample datasets that capture variations in font, size, and appearance of USDOT tags. This pre-training phase creates a robust model that can handle diverse tag appearances without requiring complex real-time adjustments during actual detection, thereby managing system complexity while maintaining high adaptability.
Solution Approach 2:
The patent uses copying by creating numerous synthetic positive and negative samples that replicate various font styles, sizes, and visual characteristics of USDOT tags. These copied samples are used to train the classifier, enabling it to recognize the diverse real-world variations of USDOT tags without requiring the system to explicitly model each variation, thus handling complexity efficiently.
3Adaptability or versatility
If sliding windows of different sizes and aspect ratios are used to detect USDOT tags, then the system can accommodate variations in tag appearance, but the computational complexity increases
Solution Approach 1:
The patent implements dynamics by using sliding windows of different sizes and aspect ratios that dynamically adapt to the varying dimensions and proportions of USDOT tags in the image. The trained classifier processes these dynamic window configurations efficiently, allowing the system to handle tag size variations without requiring manual specification of fixed window parameters, thereby balancing adaptability with computational feasibility.
Data Source
AI summary
Methods and systems for localizing numbers and characters in captured images. A side image of a vehicle captured by one or more cameras can be preprocessed to determine a region of interest. A confidence value of series of windows within regions of interest of different sizes and aspect ratios containing a structure of interest can be calculated. Highest confidence candidate regions can then be identified with respect to the regions of interest and at least one region adjacent to the highest confidence candidate regions. An OCR operation can then be performed in the adjacent region. An identifier can then be returned from the adjacent region in order to localize numbers and characters in the side image of the vehicle.