Region Proposal Networks for Text Segmentation in OCR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Optical character recognition (OCR) technologies face challenges in effectively processing images with poor quality due to background interference, clutter, and complex orientations, leading to inefficient and inaccurate text extraction from physical documents.
Innovation Solution
The implementation of region proposal networks (RPNs) for pre-processing images to identify and crop text regions, combined with neural networks for corner point identification and perspective transformation, enhances image quality and aligns text for improved OCR results, even in low-contrast or skewed documents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional OCR is applied directly to images with background interference and clutter, then text extraction can be performed, but accuracy deteriorates due to noise and orientation issues
Solution Approach 1:
The patent segments the image processing task into distinct stages: first identifying candidate text regions through region proposal networks, then refining these regions through corner point detection and perspective transformation, and finally performing OCR on the transformed regions. This segmentation allows each stage to focus on specific aspects of the problem, improving overall accuracy while handling background interference systematically.
Solution Approach 2:
The patent performs preliminary actions before OCR by using region proposal networks to identify candidate text regions, detecting corner points to determine document orientation, and applying perspective transformation to correct skew. These preliminary steps prepare the image by removing background clutter and aligning text horizontally, thereby improving OCR accuracy on poor quality images.
2Productivity
If the entire image is processed by OCR to find text, then text extraction is comprehensive, but processing efficiency deteriorates due to analyzing unnecessary background areas
Solution Approach 1:
The patent divides the image into multiple candidate text regions using region proposal networks before applying OCR. This segmentation strategy allows OCR to be applied only to relevant regions containing text, rather than the entire image, thereby improving processing efficiency while maintaining completeness of text extraction through systematic coverage of all candidate regions.
Solution Approach 2:
The patent applies different processing qualities to different regions of the image. Candidate regions identified by the region proposal network receive focused processing including corner point detection and perspective transformation, while background areas are effectively ignored. This local quality approach improves efficiency by concentrating computational resources on regions likely to contain text.
3Measurement precision
If region proposal networks are used to identify text regions first, then OCR accuracy improves, but device complexity increases due to additional processing stages
Solution Approach 1:
The region proposal network serves multiple functions within the processing pipeline: it identifies candidate text regions, provides initial localization for corner point detection, and enables efficient routing of OCR processing to relevant areas. This multi-functionality justifies the added complexity by providing substantial benefits in accuracy and efficiency across multiple processing stages.
Solution Approach 2:
The region proposal network performs preliminary region identification before the main OCR processing stage. This preliminary action simplifies the subsequent OCR task by providing pre-processed candidate regions with associated metadata (such as bounding boxes and confidence scores), thereby improving overall system accuracy despite the added complexity of the preliminary stage.
4Measurement precision
If corner point identification and perspective transformation are applied, then text alignment improves, but processing time increases due to additional transformation steps
Solution Approach 1:
The patent applies corner point identification and perspective transformation only to candidate text regions identified by the region proposal network, rather than to the entire image. This local application of transformation operations improves text alignment accuracy for regions that need it while minimizing processing time by avoiding unnecessary transformations in background areas.
Solution Approach 2:
The patent applies perspective transformation to all candidate regions identified by the region proposal network, even though not all of these regions will ultimately contain text. This partial action approach ensures that text alignment accuracy is maximized for all potential text regions, accepting some additional processing time as a trade-off for improved precision in the critical text extraction stages.
Data Source
AI summary
Arbitrary image data may be transformed into data suitable for optical character recognition (OCR) processing. A processor may generate a plurality of intermediate feature layers of an image using convolutional neural network (CNN) processing. For each intermediate feature layer, the processor may generate at least one text proposal using a region proposal network (RPN). The at least one text proposal may comprise a portion of the intermediate feature layer that is predicted to contain text. The processor may merge the text proposals with one another to form a patch of the image that is predicted to contain text. The processor may determine outer coordinates of the patch. The outer coordinates may comprise at least leftmost, rightmost, topmost, and bottommost coordinates. The processor may generate a quadrilateral of the image that is a smallest quadrilateral including the leftmost, rightmost, topmost, and bottommost coordinates.