PaperlayerView exact revision on arXiv

Source-linked AI summary

Color Independent Word Segmentation From Transcribed Bangla Passages

Faias Satter, Noor Masrur, Sk. Md. Masudul Ahsan

arXiv:2610.01191v1cs.CV

TL;DR

Handwritten Bangla word segmentation is an important first step for OCR, but the authors describe reliable Bangla OCR as lacking. They develop an image-processing system and bespoke dataset for smartphone-captured handwritten documents; they report better performance than some existing works, with results on CMATERdb1.1.1 close to Sarkar et al.’s count of correctly detected words.

  • Problem

    The authors identify a lack of reliable OCR for Bangla and target word segmentation, the first step in an OCR system.

    Source: §I · p. 1

  • Method

    The paper uses image-processing techniques and a bespoke dataset to segment words in handwritten Bangla paragraphs, including smartphone-captured images across paper and ink colors.

    Source: §I · p. 1

  • Results

    The authors report better performance than some existing works; on CMATERdb1.1.1, their method correctly segmented 13,552 of 15,730 words versus Sarkar et al.’s 13,530.

    Source: C. Experimental Results of Word… · p. 5, §V · p. 6

  • Takeaways & Limitations

    The system is presented for smartphone-captured images across paper and ink colors, and nested operations on boxes containing multiple words could further improve accuracy.

    Source: §V · p. 6

    Segmentation depends on maintaining inter-word, inter-line, and inter-character gaps; very small gaps between words and lines can lead to over-segmentation.

    Source: §V · p. 6

Abstract

from arXiv · show

An optical character recognition(OCR) system can scan paper and extract text, making people's jobs easier. While numerous OCR systems are accessible in the software sector, finding a dependable equivalent solution for Bangla is tough. When it comes to handwritten texts, the case is even more rare. The first fundamental step to any OCR is to segment words from text images. If this stage fails, the total OCR's performance will be poor no matter how promising the later stages perform. This research aims to segment words in a handwritten Bangla text image. This research can be implemented on any smartphone-captured image, irrespective of the color and type of paper and ink. Furthermore, as smartphone-captured images can create shadow interferences, the custom dataset built for this research is created in such a way that every possible obstacle that can be faced is included. For 7374 words, a total of 7278 bounding boxes are generated, which have recall of 90.60 %, precision of 91.80 %, and F1-score of 91.20 %. The system can be further improved with nested operations on bounding boxes containing several words or by adjusting the adaptive thresholding and dilation filter sizes to a more precise level.

I. Introduction

The paper targets handwritten Bangla word segmentation, a first OCR step for a language the authors describe as lacking reliable OCR. It also reviews prior segmentation work and its limitations.

I. Introduction

  • The authors seek to detect handwritten Bangla words in document images, addressing a language for which they report no reliable OCR.Word segmentation is the first step for OCR systems.Source: §I · p. 1
  • Bangla handwriting complicates segmentation because words can nearly touch, Matra connects characters within words, and vowel pieces can be absorbed by neighboring words.Some characters lack Matra, so gaps may also be mistaken for word boundaries.Source: §I · p. 1
  • The proposed scope is word segmentation in handwritten Bangla text images, using image processing and a bespoke dataset to optimize word detection.The authors aim to support smartphone-captured images irrespective of paper or ink color and shadow interference.Source: §I · p. 1

II. Related Works

  • Prior Bangla segmentation work cited here includes methods limited to binary images or small image sets, while reported average accuracies include 81.47% and 84.55%.The cited review also reports that some approaches have specific line-skew or line-extent constraints.Source: §II · p. 2

III. Proposed Methodology

The working procedure is divided into consecutive parts, some prerequisite and others fundamental. The image is scanned in grayscale to retain character information that black-and-white scanning may lose.

A. Working Procedure

  • The procedure comprises consecutive parts, with requisite sections on the flowchart’s left and fundamental sections on its right.Not every section is directly implemented to produce the output.Source: A. Working Procedure · p. 2
  • The input is scanned in grayscale because black-and-white scanning may lose character information.Figure 2 shows the input image and grayscale transformation.Source: 1) Pre-processing · p. 2, Figure 2 · p. 2

1) Pre-processing:

Pre-processing accounts for variation in image resolution and handwriting size, then reduces noise with kernel sizes adapted to contour size and paper color. The method also chooses between Otsu and adaptive thresholding based on image characteristics.

1) Pre-processing:

  • Different capture resolutions and handwriting sizes motivate determining kernel sizes for noise reduction.The method first finds the average contour size (ACS) to determine kernel sizes for different images.Source: 2) Adaptive Noise Reduction · p. 2

2) Adaptive Noise Reduction:

  • Non-local Means denoising replaces a pixel value with an average of values from surrounding pixels.Source: 2) Adaptive Noise Reduction · p. 2
  • ACS guides the denoising kernel size through an empirical mapping; darker paper prompts an additional increase of two or four.The mapping is summarized in Table I, and the method ignores contours with area below 50 as salt noise after thresholding.Source: 2) Adaptive Noise Reduction · p. 2, Table I · p. 3
  • Otsu or adaptive thresholding is selected according to image characteristics.Source: 3) Binarization · p. 3

3) Binarization:

The method checks for shadow interference before selecting a thresholding approach. Adaptive thresholding can fragment text, while Otsu leaves text unaltered; pixel inversion is applied where needed afterward.

3) Binarization:

  • The largest contour area after Otsu thresholding is used to check for illumination or shadow interference; an area above 100 indicates shadow.Source: 3) Binarization · p. 3
  • Adaptive thresholding applies across image types, but can fragment text and exclude character parts from contours; Otsu leaves text unaltered and works best when one threshold suffices.Source: 3) Binarization · p. 3
  • After thresholding, pixel values are inverted where needed to create a uniform scenario.Source: 3) Binarization · p. 3

4) Skew Correction:

Skew correction applies Canny edge detection followed by a Hough transform to identify co-linear lines. The supplied passages also describe thresholding outputs and dilation, but do not explain their role in skew correction.

4) Skew Correction:

  • Table II selects Otsu or adaptive thresholding based on specified conditions.Source: Table II · p. 3
  • Fig. 3 shows the result after applying both thresholding methods to Fig. 2c.Source: Figure 3 · p. 3
  • Canny edge detection followed by a Hough transform identifies co-linear lines as intersecting points used to correct skew.Source: 4) Skew Correction · p. 3
  • Dilation is applied using a squared structured element with white foreground pixels.Source: 5) Word Detection · p. 3

5) Word Detection:

Word detection sets a minimum contour area, applies dilation, and checks contour counts while adjusting the dilation settings. The process stops at the maximum word count before words collide, after which bounding boxes mark the segmented words.

5) Word Detection:

  • A contour is accepted when its connected-pixel area exceeds the minimum area threshold, which determines the dilation process.Source: 5) Word Detection · p. 3
  • For Otsu thresholding, the minimum area threshold is calculated from average contour size; for adaptive thresholding, empirically decided mappings use writing density and then ACS and WD.Source: 5) Word Detection · p. 3
  • A dilation kernel that is too small risks under-segmentation, while one that is too large risks over-segmentation; Table V maps ACS to a starting kernel size.Source: 5) Word Detection · p. 3
  • After an initial dilation, the method recounts contours larger than MAT as dilation iterations or kernel size increase.Fig. 4 presents results for different dilation kernel sizes and iteration numbers.Source: 5) Word Detection · p. 4, Figure 4 · p. 4
  • The search stops at the maximum word count just before two words collide during dilation, and the resulting words are segmented with bounding boxes.For Fig. 4b, ACS is 641 and the starting dilation kernel is 3×3; Fig. 4c shows most words detected at the optimum.Source: 5) Word Detection · p. 4, 6) Word Segmentation Using Bounding… · p. 4

6) Word Segmentation Using Bounding Boxes:

The section identifies empirical mappings used to set MAT and the starting dilation kernel size. The supplied table captions name the inputs and outputs of these mappings.

6) Word Segmentation Using Bounding Boxes:

  • Table III maps writing density (WD) to the minimum area threshold (MAT).Source: Table III · p. 4
  • Table IV further maps ACS and WD to MAT.Source: Table IV · p. 4
  • Table V maps average contour size (ACS) to the starting dilation kernel size.Source: Table V · p. 4

IV. Results

The custom dataset comprises 80 handwritten Bangla documents designed to vary handwriting, paper and ink colors, shadows, and image sizes and resolutions.

IV. Results

  • The custom dataset contains 80 handwritten Bangla documents with varied handwriting styles.Source: A. Datasets · p. 4
  • The documents use papers in white, red, blue, green, and black, with glitter, gel, and ballpoint pens in multiple colors.Source: A. Datasets · p. 4
  • The dataset includes shadow interference and images captured on different devices to vary image size and resolution.Source: A. Datasets · p. 4

B. Evaluation Metrics

The section describes recall, precision, and F1-score as measures of word-detection accuracy, and includes visuals of detected words and customized-dataset samples.

B. Evaluation Metrics

  • Recall, precision, and F1-score are used to measure the system’s word-detection accuracy.F1-score is described as the weighted harmonic mean of precision and recall.Source: B. Evaluation Metrics · pp. 4–5
  • Fig. 5 shows words detected and enclosed by bounding boxes after the stages illustrated in Fig. 2a.Source: Figure 5 · p. 4
  • Fig. 6 presents sample images from the customized dataset.Source: Figure 6 · p. 4

C. Experimental Results of Word Segmentation

The experiments report word-segmentation counts across three datasets and compare results with prior work on two of them. The analysis also identifies segmentation errors and handwriting-dependent trade-offs in dilation and area-threshold choices.

C. Experimental Results of Word Segmentation

  • Of 7,374 words in 80 document images, 6,681 were correctly segmented, 527 had segmentation errors, and 166 were undetected; 7,278 bounding boxes were generated.Table VI is identified as reporting recall, precision, and F1-score, but its values are not stated in the supplied text.Source: C. Experimental Results of Word… · p. 5, Table VI · p. 5
  • On CMATERdb1.1.1, the method correctly segmented 13,552 of 15,730 words, versus Sarkar et al.’s 13,530; Sarkar et al. reported an 86.01% success rate on segmented text lines.Source: C. Experimental Results of Word… · p. 5, Table VI · p. 5
  • On the Bangla Writing dataset, 6,380 of 7,400 words were correctly segmented, while Das et al. reported 88.39% line, 95.65% word, and 84.55% average accuracy.The supplied passage identifies these as Das et al.’s results on a portion of the dataset.Source: C. Experimental Results of Word… · p. 5, Table VI · p. 5
  • The analysis distinguishes under-segmentation, over-segmentation, defect-segmentation, and overlooking errors; punctuation may also receive its own box.Under-segmentation splits one word across boxes, over-segmentation groups multiple words in one box, and defect-segmentation excludes part of a word.Source: C. Experimental Results of Word… · p. 5, Figure 7 · p. 5
  • Tighter dilation or MAT settings can reduce over-segmentation in dense writing but increase under-segmentation or omitted vowels in other handwriting styles.A larger MAT can include distant vowels but also raise over-segmentation and overlooking risks; the selected setting avoids taking any error to an extreme.Source: C. Experimental Results of Word… · pp. 5–6, Figure 8 · p. 6

V. Conclusion

The conclusion presents a word-segmentation system for smartphone-captured handwritten Bangla documents, with performance dependent on visible gaps and background contrast. It identifies nested processing and more precise filter sizing as possible improvements.

V. Conclusion

  • The system segments handwritten Bangla document paragraphs in smartphone-captured images and can handle paper and ink of any color.Source: §V · p. 6
  • The method works best when writers maintain inter-word, line, and character gaps; near-character-valued backgrounds may be detected as words.The conclusion says cropping can eliminate such background areas.Source: §V · p. 6
  • The authors suggest nested processing for boxes containing multiple words and more precise adaptive-thresholding and dilation filter sizes as improvements.Source: §V · p. 6
Loading 2610.01191v1…