290 likes | 528 Views
A Gradient Vector Flow-Based Method for Video Character Segmentation. Trung Quy Phan , Palaiahnakote Shivakumara , Bolan Su and Chew Lim Tan. Agenda. Introduction Proposed method Experimental results Summary. Agenda. Introduction Proposed method Experimental results Summary.
E N D
A Gradient Vector Flow-Based Method for Video Character Segmentation TrungQuyPhan, PalaiahnakoteShivakumara, Bolan Su and Chew Lim Tan
Agenda • Introduction • Proposed method • Experimental results • Summary
Agenda • Introduction • Proposed method • Experimental results • Summary
Introduction • Text extraction from video frames video search and retrieval • Graphics text:artificially added • Scene text:part of the scene
Introduction • Character segmentation: split text line into individual characters • Important preprocessing step to improve recognition rate
Character Segmentation • Challenges • Low resolution • Complex background • Unconstrained fonts & colors • Touching characters
Character Segmentation • Previous methods • Document analysis methods binary • Projection analysis [Huang09] sensitive Image from [Lienhart02]
Character Segmentation • Previous methods • Document analysis methods binary • Projection analysis [Huang09] sensitive • Neural networks [Saidane08] vertical • Proposed method: gray, curved, less sensitive
Agenda • Introduction • Proposed method • Experimental results • Summary
Steps • Cut candidate identification • “Seed” pixels • Path finding • Least cost path based on seed pixels • Path verification • Eliminate false positives
1. Cut Candidate Identification • Good cut: far from character edges • Gradient vector flow • Propagate gradient information into background • Mainly used for registration and tracking Edge Gap
1. Cut Candidate Identification • Candidate cut pixel GVF field g(x, y) = (u(x, y), v(x, y))
1. Cut Candidate Identification • Gap pixels • “Medial” pixels: middle of character strokes
1. Cut Candidate Identification • Gap pixels • “Medial” pixels: middle of character strokes • Medial pixels lead to transitions between background and characters large intensity variations handled by cost function in the next step
2. Path Finding col x-1 x x+1 • Input image = graph • Nodes: pixels • Edges • Cost function • Penalize paths that go through medial pixels (higher cost) • Small penalty for diagonal moves row y row y+1
2. Path Finding • Cut = least cost path from top row to bottom row • Dynamic programming • Multiple cuts: path finding multiple times • Starting point every h / 4 pixels h
3. Path Verification • Starting points in same gap converge to the same end point end points are more reliable • Backward path finding from each end point • False cuts switch to the sides of the characters
Agenda • Introduction • Proposed method • Experimental results • Summary
Experiments • Text lines from TRECVID videos • Compare with [Kopf05] • Path finding based on intensities • Proposed method: gradient, verification
Results – Horizontal Text Kopf’s method Proposed method
More Results – Proposed Method • Non-horizontal scene text • Logo text
Performance – English Text • Similar recall • Proposed method has much better precision • Criterion for “good” cuts
Performance – Chinese Text • Proposed method • Precision for Chinese: sub-components
Recognition Accuracy • Segmentation helps to improve binarization? • Binarization method: [Su10] • Reported to outperform classical methods, e.g. Otsu’s and Niblack’s Line level ‘TO’ Character level ‘TONIGH §’
Recognition Accuracy • Tesseract OCR engine on English text • With segmentation: 7.5% increase
Agenda • Introduction • Proposed method • Experimental results • Summary
Summary • Segmentation: important preprocessing for recognition • GVF, path finding, verification gray, curved • Future: end-to-end video text recognition
References Huang et al. (2009). A new video text extraction approach. In Proc. ICME. Kopf et al. (2005). Robust Character Recognition in Low-Resolution Images and Videos. Technical report. University of Mannheim. Lienhart et al. (2002). Localizing and Segmenting Text in Images and Videos. CSVT. Saidane et al. (2008). An Automatic Method for Video Character Segmentation. Image Analysis and Recognition. Su et al. (2010). Binarization of historical document images using the local maximum and minimum. In Proc. DAS.