Wheel-Type Meter Character Recognition With Bias-Weighted Labels

The patent changes the annotation parameter from simple class labels to weight values that represent character bias degrees. Instead of annotating dual partial-character images with a single class label, the system calculates weight values for each character based on their visual prominence, transforming the annotation into a quantitative parameter that reflects the actual character distribution and bias in the image.

The patent introduces weight values as an intermediary between the image content and the training labels. These weight values serve as a mediator that captures the complex character bias information in dual partial-character images, allowing the training model to learn the nuanced relationships between different character representations without requiring complex manual annotation processes.

The system implements self-service by automatically calculating weight values for each character in the image based on their visual characteristics and positions. Instead of requiring manual annotation of weight values, the system uses image processing algorithms to automatically determine the bias degree of each character, making the annotation process self-executing and reducing human intervention while maintaining high training sample accuracy.

The patent performs preliminary character detection and weight value calculation during the annotation phase. By pre-calculating the weight values and character regions before training, the system prepares the training data in advance with accurate bias information, which then can be directly used during model training without requiring complex real-time processing during the training execution phase.