Reinforcement Learning Segmentation Agent for License Plate Character Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Automatic License Plate Recognition (ALPR) systems face challenges in accurately segmenting characters from license plate images, particularly in the presence of complex backgrounds and partial obstructions, leading to issues with under-segmentation and over-segmentation due to the limitations of projective segmentation techniques.
Innovation Solution
A reinforcement learning approach is employed to initialize a segmentation agent that learns to identify proper segmentation paths from top to bottom in a cropped license plate image, using rewards for good and bad moves, and combining this with projective segmentation to improve robustness and reduce image capturing unit requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If projective segmentation techniques are used to segment characters in license plate images, then computational resources are minimized and basic character boundaries are obtained, but segmentation accuracy deteriorates in the presence of complex backgrounds and partial obstructions
Solution Approach 1:
A reinforcement learning agent is introduced as an intermediary between the image input and character segmentation output. The agent processes the license plate image through a series of decisions (moves) to identify character boundaries, serving as a mediator that resolves the contradiction by providing high accuracy without requiring complex traditional image processing techniques
Solution Approach 2:
The patent changes the approach from using fixed projection histogram parameters to using a reinforcement learning agent that learns optimal segmentation paths through training. This parameter change enables the system to adapt to complex backgrounds and obstructions while maintaining computational efficiency through the learned policy
2Measurement precision
If segmentation thresholds are adjusted to reduce under-segmentation events, then character boundary detection improves, but over-segmentation events increase
Solution Approach 1:
The reinforcement learning agent uses feedback in the form of rewards and penalties during training to learn optimal segmentation behavior. Good moves (correct boundary detection) receive positive reinforcement while bad moves (over or under-segmentation) receive penalties, enabling the system to achieve stable and accurate segmentation without manual threshold adjustment
Solution Approach 2:
The segmentation approach transitions from static fixed thresholds to a dynamic reinforcement learning agent that adapts its segmentation decisions based on the specific image content. The agent's policy is learned through interaction with the environment, allowing it to dynamically adjust segmentation behavior to balance under-segmentation and over-segmentation risks
3Reliability
If reinforcement learning training is performed to improve segmentation accuracy, then robustness to complex backgrounds improves, but training time and computational resources increase
Solution Approach 1:
The reinforcement learning agent is trained in advance during a training phase before actual deployment. This preliminary action allows the agent to learn optimal segmentation paths and accumulate reward signals that guide its behavior during real-time operation, achieving high robustness without requiring extensive computational resources at runtime
Data Source
AI summary
Methods and systems for achieving accurate segmentation of characters with respect to a license plate image utilizing a reinforcement learning approach. A vehicle image can be captured by an image capturing unit and processed utilizing an ALPR (Automatic License Plate Recognition) unit. The reinforcement learning (RL) approach can be configured to initialize a segmentation agent with a starting location. A proper segmentation path (cuts) from top to bottom and from a darker to lighter area in a cropped license plate image can be identified by the segmentation agent during a training phase. Rewards can be provided based on a number of good and bad moves. The association between a current state and a sensory input with a preferred action can be learned by the segmentation agent at the end of the training phase.