3D Beam Search Decoding for License Plate Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional electronic devices for license plate number recognition are computationally resource-intensive and prone to errors due to noise from vehicle motion, camera rolling shutter, and ambient conditions, leading to incorrect identification of license plate numbers.
Innovation Solution
An electronic device employing a 3D beam search operation within a connectionist-temporal-classification (CTC) matrix to efficiently recognize license plate numbers by decoding only prominent likelihood values, reducing redundancy and improving accuracy in real-time identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional OCR techniques are applied on different image frames to extract license plate numbers, then the license plate number can be identified from multiple frames, but the computational resource consumption increases and multiple erroneous results are generated due to noise
Solution Approach 1:
The patent segments the license plate recognition process into distinct stages: image frame extraction, CTC matrix generation for each frame, and 3D beam search decoding across multiple frames. This segmentation allows the system to process multiple frames systematically while managing computational resources efficiently through the beam search algorithm that prunes unlikely paths.
Solution Approach 2:
The patent transitions from 2D CTC matrices (individual frames) to a 3D CTC matrix that incorporates the temporal dimension across multiple image frames. This dimensional extension allows the beam search algorithm to leverage information across time, improving recognition accuracy by considering the sequence of frames rather than treating each frame independently.
2Productivity
If conventional OCR techniques are applied on different image frames to extract license plate numbers, then the license plate number can be identified from multiple frames, but multiple erroneous license plate numbers are output leading to confusion
Solution Approach 1:
The 3D beam search algorithm incorporates feedback mechanisms where the recognition results from previous frames influence the decoding of subsequent frames. The algorithm maintains a beam of candidate paths and uses likelihood scores from each frame to update and refine the candidate license plate numbers, providing continuous feedback that improves accuracy over time while maintaining real-time processing capability.
3Reliability
If 3D beam search is applied in CTC matrix to decode prominent likelihood values, then recognition accuracy is improved and redundancy is reduced, but the algorithm complexity increases
Solution Approach 1:
The beam search algorithm applies partial action by maintaining only a limited number of candidate paths (the beam width) rather than exploring all possible character sequences. This selective approach decodes only the most prominent likelihood values above a certain threshold, achieving high accuracy while avoiding the combinatorial explosion that would result from exhaustive decoding of the entire 3D CTC matrix.
Data Source
AI summary
An electronic device for license plate recognition includes control circuitry and a memory that stores a sequence of image frames. The control circuitry extracts a first set of license plate images from the sequence of image frames and estimates a set of two dimensional (2D) connectionist-temporal-classification (CTC) matrices based on application of a pre-trained CTC-based deep neural network (DNN) on the first set of license plate images. The control circuitry further generates a three dimensional (3D) CTC matrix from the estimated set of 2D CTC matrices and decodes a set of alphanumeric characters of a license plate number of a first license plate. The set of alphanumeric characters are decoded based on a 3D beam search through the generated 3D CTC matrix. The decoded set of alphanumeric characters correspond to maximum likelihood values for the decoded set of alphanumeric characters in the 3D beam search.