Text Detection via Neural Network Feature Matrix Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional target object detection methods are ineffective for text detection due to the dynamic nature of text boundaries and spaces between characters, leading to inaccurate and robustness issues in text positioning.
Innovation Solution
A text detection method involving a neural network that generates a target feature matrix, maps it to predicted subregions using a fully connected layer, and employs a text clustering algorithm to connect these subregions into predicted text lines, determining the text area based on text feature information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional target object detection algorithm is used for text detection, then detection process is simple, but text positioning accuracy is low and detection robustness is poor
Solution Approach 1:
The patent divides text detection into multiple stages: first detecting text lines using conventional algorithms, then segmenting characters within each text line, and finally reconstructing complete text. This segmentation approach allows each stage to focus on specific tasks, improving overall accuracy while managing complexity through modular processing.
Solution Approach 2:
The patent introduces text lines as an intermediary structure between image input and final text output. By first identifying text lines and then processing characters within these lines, the system creates a hierarchical intermediate representation that improves detection robustness against spaces and boundary variations.
2Reliability
If conventional target object detection algorithm is used, then processing speed is fast, but detection robustness is low due to spaces between characters
Solution Approach 1:
The patent segments text detection into line-level and character-level processing. By first detecting text lines as complete units and then processing characters within each line separately, the system maintains robustness against inter-character spaces while avoiding the need to process the entire image at full resolution, thus preserving processing speed.
Solution Approach 2:
The patent introduces a hierarchical dimension to text detection by organizing detection results into multiple levels: image level (text lines), line level (character sequences), and character level (individual characters). This dimensional organization allows the system to handle spaces and boundaries more robustly while maintaining efficient processing through progressive refinement.
Data Source
AI summary
This application provides a text detection method, including: obtaining, by a computer device, an image; inputting the image into a neural network, and outputting a target feature matrix; inputting the target feature matrix into a fully connected layer, the fully connected layer mapping each element of the target feature matrix to a predicated subregion corresponding to the image according to a preset anchor; and obtaining text feature information of the predicated subregion, connecting the predicated subregion into a corresponding predicted text line according to the text feature information of the predicated subregion by using a text clustering algorithm, and determining a text area corresponding to the image.