Abstract

An enormous amount of digital images are being generated and stored every day. Understanding text in these images is an important challenge with large impacts for academic, industrial and domestic applications. Recent studies address the difficulty of separating text targets from noise and background, all of which vary greatly in natural scenes. To tackle this problem, we develop a text detection system to analyze and utilize visual information in a data driven, automatic and intelligent way.

The proposed method incorporates features learned from data, including patch-based coarse-to-fine detection (Text-Conv), connected component extraction using region growing, and graph-based word segmentation (Word-Graph). Text-Conv is a sliding window-based detector, with convolution masks learned using the Convolutional k-means algorithm (Coates et. al, 2011). Unlike convolutional neural networks (CNNs), a single vector/layer of convolution mask responses are used to classify patches. An initial coarse detection considers both local and neighboring patch responses, followed by refinement using varying aspect ratios and rotations for a smaller local detection window. Different levels of visual detail from ground truth are utilized in each step, first using constraints on bounding box intersections, and then a combination of bounding box and pixel intersections. Combining masks from different Convolutional k-means initializations, e.g., seeded using random vectors and then support vectors improves performance. The Word-Graph algorithm uses contextual information to improve word segmentation and prune false character detections based on visual features and spatial context. Our system obtains pixel, character, and word detection f-measures of 93.14%, 90.26%, and 86.77% respectively for the ICDAR 2015 Robust Reading Focused Scene Text dataset, out-performing state-of-the-art systems, and producing highly accurate text detection masks at the pixel level.

To investigate the utility of our feature learning approach for other image types, we perform tests on 8- bit greyscale USPTO patent drawing diagram images. An ensemble of Ada-Boost classifiers with different convolutional features (MetaBoost) is used to classify patches as text or background. The Tesseract OCR system is used to recognize characters in detected labels and enhance performance. With appropriate pre-processing and post-processing, f-measures of 82% for part label location, and 73% for valid part label locations and strings are obtained, which are the best obtained to-date for the USPTO patent diagram data set used in our experiments.

To sum up, an intelligent refinement of convolutional k-means-based feature learning and novel automatic classification methods are proposed for text detection, which obtain state-of-the-art results without the need for strong prior knowledge. Different ground truth representations along with features including edges, color, shape and spatial relationships are used coherently to improve accuracy. Different variations of feature learning are explored, e.g. support vector-seeded clustering and MetaBoost, with results suggesting that increased diversity in learned features benefit convolution-based text detectors.

Library of Congress Subject Headings

Optical pattern recognition; Computer vision; Image analysis

Publication Date

5-12-2016

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Richard Zanibbi

Advisor/Committee Member

Bo Yuan

Advisor/Committee Member

Carl Salvaggio

Comments

Physical copy available from RIT's Wallace Library at TA1650 .Z48 2016

Campus

RIT – Main Campus

Download

Share

COinS