[Paper Review] Exploring Geometric Property Thresholds For Filtering Non-Text Regions In A Connected Component Based Text Detection Application

Teresa Nicole Brooks|arXiv (Cornell University)|Sep 11, 2017

Handwritten Text Recognition Techniques3 references3 citations

TL;DR

This paper investigates geometric property thresholds for filtering non-text regions in connected component-based text detection using MSER and stroke width variation. It demonstrates that manually tuned thresholds for aspect ratio, eccentricity, solidity, and stroke width variation effectively detect text in both natural and unnatural scenes, with higher stroke width max thresholds and reduced expansion amounts improving detection in stylized, high-contrast non-natural images.

ABSTRACT

Automated text detection is a difficult computer vision task. In order to accurately detect and identity text in an image or video, two major problems must be addressed. The primary problem is implementing a robust and reliable method for distinguishing text vs non-text regions in images and videos. Part of the difficulty stems from the almost unlimited combinations of fonts, lighting conditions, distortions, and other variations that can be found in images and videos. This paper explores key properties of two popular and proven methods for implementing text detection; maximum stable external regions (MSER) and stroke width variation.

Motivation & Objective

• To evaluate the impact of font style and image characteristics (e.g., blur, contrast, texture) on geometric property thresholds in text detection.
• To determine which geometric properties (e.g., aspect ratio, eccentricity, solidity) are most effective for filtering non-text regions across diverse image types.
• To assess the robustness of MSER and stroke width variation in detecting text under varying conditions, especially in unnatural scenes like product packaging.
• To identify threshold tuning strategies that improve detection performance for highly stylized or custom fonts in non-natural images.
• To lay groundwork for automating threshold selection to enhance scalability and robustness in real-world text detection systems.

Proposed method

• Employed Maximally Stable External Regions (MSER) for initial region detection in grayscale images.
• Extracted geometric properties—aspect ratio, eccentricity, solidity, extent, Euler number—from MSER regions to filter non-text candidates.
• Applied stroke width variation metric: standard deviation of stroke width divided by mean stroke width, used with a max variance threshold.
• Used image contrast enhancement via standard deviation-based stretching as a pre-processing step.
• Implemented bounding box expansion and merging to group individual character detections into word or line candidates.
• Validated results using MATLAB’s Tesseract-based OCR engine for text recognition and visual inspection of detected regions.

Experimental results

Research questions

• RQ1 How do different font styles in natural and unnatural scene images affect the optimal geometric property thresholds for text detection?
• RQ2 Which geometric properties (e.g., aspect ratio, solidity, eccentricity) are most effective at distinguishing text from non-text regions across diverse image types?
• RQ3 How do image-specific characteristics such as high contrast, blur, and texture influence the required threshold values for reliable text detection?
• RQ4 To what extent do custom or highly stylized fonts necessitate adjustments in stroke width variation and bounding box expansion parameters?
• RQ5 Can a manually tuned, lightweight system using MSER and stroke width variation achieve robust text detection across both natural and unnatural scenes?

Key findings

• The application successfully detected primary and secondary text regions in both natural and unnatural images after manual tuning of geometric thresholds.
• Stroke width variation max threshold needed to be doubled (increased by a factor of 2) for non-natural images compared to natural images to detect all characters.
• Reduced expansion amount for neighboring bounding boxes improved detection in non-natural images, particularly for wide or irregularly spaced text.
• High-contrast images degraded performance when contrast was further increased via standard deviation-based stretching, indicating a threshold effect.
• Despite challenges from stylized fonts and high contrast, the combination of MSER and stroke width variation proved robust, with minimal anticipated impact on threshold sensitivity.
• The OCR module using Tesseract successfully recognized standard English text without retraining, though custom fonts required future model training.