Character Recognition Training With Self-Supervised Encoder Pretraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing character recognition models trained using supervised methods can only learn features from a small number of tagged images, leading to low accuracy in character recognition.
Innovation Solution
Perform self-supervised training on an encoder using untagged training samples and supervised training on a decoder using tagged training samples, combining both methods to enable the model to learn image and text prior information simultaneously.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised training method is used, then the model can learn from tagged images, but the recognition accuracy remains low due to limited tagged data
Solution Approach 1:
The patent performs self-supervised pre-training on the encoder using abundant untagged images before supervised fine-tuning. This preliminary action allows the model to learn general image features and text priors from large-scale untagged data, which then improves the model's ability to learn from limited tagged images during supervised training, ultimately enhancing character recognition accuracy
Solution Approach 2:
The patent implements self-supervised learning where the model generates its own training signals from untagged images through data augmentation and reconstruction tasks. The encoder learns to reconstruct original images from augmented versions, creating self-generated supervision signals that enable the model to learn from untagged data without requiring manual annotations, thus expanding the effective training data quantity
2Loss of information
If only supervised training is used, then training process is simple, but the model cannot learn prior information from untagged images
Solution Approach 1:
The patent divides the training process into two distinct phases: self-supervised pre-training phase and supervised fine-tuning phase. The encoder and decoder are trained separately with different objectives - the encoder learns from untagged images through self-supervised reconstruction, while the decoder learns from tagged images through supervised text recognition. This segmentation allows each component to learn specific prior information independently before working together
Solution Approach 2:
The patent introduces an intermediary reconstruction task as a bridge between untagged images and the final recognition task. The encoder-decoder pair is trained to reconstruct input images, serving as an intermediary objective that enables the model to learn rich image and text priors from untagged data. This intermediary task facilitates knowledge transfer to the final recognition task without requiring direct supervision
Data Source
AI summary
The present disclosure provides a character recognition model training method and apparatus, a character recognition method and apparatus, a device and a medium, relating to the technical field of artificial intelligence, and specifically to the technical fields of deep learning, image processing and computer vision, which can be applied to scenarios such as character detection and recognition technology. The specific implementing solution is: partitioning an untagged training sample into at least two sub-sample images; dividing the at least two sub-sample images into a first training set and a second training set; where the first training set includes a first sub-sample image with a visible attribute, and the second training set includes a second sub-sample image with an invisible attribute; performing self-supervised training on a to-be-trained encoder by taking the second training set as a tag of the first training set, to obtain a target encoder.