Text Detection Model Using Transformer Sequence Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current text detection models face challenges in accurately detecting text in complex natural scenes with varying font sizes and severe interference, often requiring complex post-processing and consuming significant computing resources, and struggle with overlapping bounding boxes.
Innovation Solution
A text detection model comprising a text feature extraction sub-model, a text encoding sub-model, a decoding sub-model, and an output sub-model is trained using a method that inputs sample images with labeled position information to obtain predicted position and category information, improving feature modeling and reducing false detections by using a Transformer architecture for sequence encoding and decoding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional text detection models are used, then the detection process is simple, but the detection accuracy is poor in complex natural scenes with varying font sizes and severe interference
Solution Approach 1:
The model is divided into four distinct sub-models: text feature extraction sub-model, text encoding sub-model, decoding sub-model, and output sub-model. Each sub-model handles a specific aspect of text detection, allowing the system to process complex features systematically while maintaining modularity and manageability.
Solution Approach 2:
The patent introduces a text sequence dimension by encoding text features into sequence vectors through the text encoding and decoding sub-models. This transforms the detection problem from traditional 2D spatial analysis to include temporal sequence processing, enabling better handling of text with varying fonts, sizes, and orientations.
2Measurement precision
If complex post-processing is applied to improve detection accuracy, then the accuracy improves, but the computing resources consumed increase significantly
Solution Approach 1:
The text feature extraction sub-model performs preliminary processing by extracting and encoding text features before the main detection process. The text encoding sub-model pre-encodes text vectors into sequence representations, preparing the data in advance to reduce the computational burden during detection and eliminate the need for complex post-processing.
3Productivity
If traditional detection methods are used, then the processing speed is fast, but the ability to handle overlapping bounding boxes is poor
Solution Approach 1:
The decoding sub-model generates sequence vectors that provide feedback about text characteristics and relationships. This feedback mechanism allows the model to iteratively refine its understanding of overlapping text regions, improving the accuracy of detecting multiple texts in overlapping bounding boxes while maintaining efficient processing through the streamlined architecture.
Data Source
AI summary
A method training a text detection model and a method of detecting a text. The training method includes: inputting a sample image into a text feature extraction sub-model of a text detection model to obtain a text feature of a text in the sample image, the sample image having a label indicating an actual position information and an actual category; inputting a predetermined text vector into a text encoding sub-model of the text detection model to obtain a text reference feature; inputting the text feature and the text reference feature into a decoding sub-model of the text detection model to obtain a text sequence vector; inputting the text sequence vector into an output sub-model of the text detection model to obtain a predicted position information and a predicted category; and training the text detection model based on the predicted and actual categories, the predicted and actual position information.