Adaptive Image Sampling for Text Detection in Natural Scenes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional OCR systems face challenges in accurately detecting text from images with poor quality, such as variations in lighting, shadows, contrast, and blur, due to the need for high-quality images and the inefficiency in processing non-text regions in natural scenes.
Innovation Solution
An adaptive sampling method is employed to identify connected pixel regions indicative of text by sampling images horizontally and vertically, using Maximally Stable Extremal Regions (MSERs) to generate text masks, which are then combined and sent to a character recognition engine for processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional OCR engines process entire images uniformly, then text detection coverage is comprehensive, but processing time and computational resources increase significantly
Solution Approach 1:
The patent divides the image processing task into multiple sampling rates (e.g., coarse sampling at 1/8th resolution, fine sampling at original resolution). Different regions are processed at different resolutions based on their text likelihood, allowing comprehensive coverage while reducing overall computational load by processing only necessary regions at high resolution.
Solution Approach 2:
The patent applies adaptive sampling where only portions of the image requiring detailed analysis are processed at full resolution. Regions identified as containing text through preliminary coarse sampling undergo further fine sampling, while non-text regions are discarded early, avoiding excessive processing of entire images uniformly.
2Measurement precision
If high-quality images are used for OCR, then text recognition accuracy improves, but image capture requirements and processing complexity increase
Solution Approach 1:
The patent applies different processing quality levels to different image regions. High-resolution fine sampling and detailed MSER analysis are applied only to regions identified as containing text, while non-text regions undergo only coarse sampling or are discarded. This maintains text recognition accuracy for relevant regions while reducing overall processing complexity.
Solution Approach 2:
The patent performs preliminary coarse sampling and MSER detection at reduced resolution before committing to full-resolution processing. This preliminary action identifies text-containing regions early, allowing the system to focus computational resources on regions where high-quality processing will actually improve recognition accuracy, rather than uniformly processing the entire image at high quality.
3Productivity
If adaptive sampling is used to eliminate non-text regions early, then processing efficiency improves, but text detection precision may be reduced in low-quality images
Solution Approach 1:
The patent implements a dynamic multi-stage sampling approach where the sampling rate adapts based on region characteristics. Coarse sampling with higher discard thresholds is applied to regions clearly identified as non-text, while fine sampling with lower thresholds is applied to ambiguous or text-like regions. This dynamic adaptation maintains text detection precision even in low-quality images where text boundaries may be less distinct.
Solution Approach 2:
The patent uses feedback from MSER detection results to adjust subsequent processing. Regions identified as containing MSERs (potential text) trigger fine sampling for further analysis, while regions without MSERs undergo coarse sampling or are discarded. This feedback mechanism ensures that potentially ambiguous regions in low-quality images receive appropriate attention, maintaining detection precision while preserving processing efficiency through selective refinement.
Data Source
AI summary
Various embodiments enable a device to perform an adaptive sampling method for locating text regions in images of natural scenes. Since only a fraction of pixels in an input image correspond to text regions, it is desirable to eliminate non-text regions early in the text detection process. Therefore, in at least one embodiment, an image is sampled horizontally to identify connected pixel regions indicative of text. In this example, each sampled row of the image is tested for containing Maximally Stable Extremal Regions (MSERs) in order to identify regions containing text. If a respective row contains a connected pixel region indicative of text, the neighboring rows are recursively sampled at finer levels in order to fully contain the text and sampling is terminated for regions which are unlikely to contain text. This sampling process can also be performed for the vertical dimension for regions which are determined to contain text.