Double-Row License Plate Recognition Using Unified Neural Decoding

The patent merges the identification of upper and lower row characters into a single unified process. The neural network model processes the entire license plate image simultaneously, extracting features from both rows in parallel rather than sequentially. This combining approach eliminates the time-consuming separate marking and identification steps while maintaining high identification accuracy through the model's ability to handle complex spatial relationships between characters.

The patent replaces the traditional mechanical OCR identification system with a deep neural network model. Instead of using rule-based character localization and separate recognition processes, the system employs a learned model that automatically extracts features and identifies characters from both rows simultaneously. This substitution enables parallel processing of multiple character regions, dramatically improving identification speed while maintaining or enhancing accuracy.

The patent segments the neural network into distinct functional modules: convolutional layers for feature extraction, a deconvolutional layer for spatial rearrangement, and an output layer for character identification. This segmentation allows each module to perform a specific function efficiently, reducing the overall computational complexity while maintaining the ability to process both rows simultaneously. The modular structure also enables optimized training and deployment.

The patent introduces a spatial rearrangement dimension through the deconvolutional layer. Instead of directly processing the complex 2D license plate image with both rows, the model transforms the feature map into a rearranged spatial configuration that separates and organizes character features from different rows. This dimensional transformation simplifies the subsequent identification process while maintaining the simultaneous processing capability.