Handwriting Removal from Document Images Computer Vision Lab

Handwriting Removal from Document Images Computer Vision Lab Hyeonseob Nam

Overview • Introduction –Motivation & Goal –Related Works • Proposed Method –Image binarization –Character segmentation –Classification of machine-printed/handwritten text –Touching-character separation –Text Line segmentation • Results

Motivation • Scanned document images usually contain annoying handwritten stuff • Harmful in various fields: document recognition, search and retrieval, historic document analysis, etc.

Goal • Remove handwritings from a document image and restore the original machine-printed image

Framework Image Binarization Connected Components Extraction Dimensionality Reduction Local Threshold Learning Testing Touching-character Separation Classification Text Line Segmentation

Image Binarization

Image Binarization • Maximize between-class variance • Minimize within-class variance Gray-level histogram

Block Segmentation • Connected-Component(CC) Analysis

Classification • Classify the CCs into Machine-printed or Hand-written

Classification - dimensionality reduction • Project CCs into a hyper-dimensional character space

Classification • Local distance threshold learning training

Classification - training • Local distance threshold learning Find the optimal radius for each template, such that - maximize the cdf for machine-printed samples - minimize the cdf for handwritten samples

Classification - testing • Determine whether given CC is in the local threshold

Results

Touching-character Problem • Segmentation failure in low-resolution images

Touching-character Separation Vertical Projecton: v(x) Peak-to-Valley function: Pv(x) = {v(x-1)-2*v(x)+v(x+1)}/v(x)

Touching-character Separation • Recursive Touching-character Separation & Classification

Results Original Initial classification Final result

Results Original Initial classification Final result

Results Original Initial classification Final result

Handling Overlapped Cases

Handling Overlapped Cases • Text Line Segmentation –Projection-based Method

Text Line Segmentation

Text Line Segmentation

Text Line Segmentation • Before

Text Line Segmentation • After

Results

Results

Results

Results

Limitations • Cannot recover pure machine-printed text • Cannot handle an isolated character

Thank you