978-1-4799-5148-2/14/$31.00 ©2014 IEEE 638
2014 11th International Conference on Fuzzy Systems and Knowledge Discovery
Address Block Localization for Chinese Postal
Envelopes with Clutter Background
Meiling Cheng, Jinhua Xu
Department of Computer Science and Technology, East China normal university 200241
Abstract—In this paper we propose a novel supervised model to
localize the address block for Chinese postal envelopes. The
problem is formulated as a binary classification problem. We get
the probability map via joint Conditional Random Field (CRF)
training and dictionary learning. Histograms of Oriented
Gradients (HOG) are used as descriptors. We evaluate our
model on a challenging Chinese postal envelope database with
clutter background. Experiment results demonstrate our model
performs well and is robust to appearance variations in
illumination, rotation, and clutter background.
Keywords-Address block localization; Histogram of oriented
gradient; Conditional random field; Dictionary learning
I. INTRODUCTION
Automated address block Localization (AABL) in
envelopes (especially for plastic envelopes) has been probed
and raised a growing concern because it should be done before
optical character recognition system. However, it is still a
challenging problem due to the appearance variations and
clutter backgrounds between different envelope images. For
example, the address blocks on envelopes are handwritten or
printed, single-line or multi-line. The image size, scale,
rotation and clutter background also increase the difficulty of
the AABL problem, especially for plastic envelopes whose
transparency makes backgrounds more complex and thus
harder to detect.
In recent years, many related research has been done to
tackle different aspects of the AABL problem. The methods
can be divided into two classes: selecting the best one from
candidate blocks and directly extracting address block. The
method proposed by Jeong et al. [1] extracts connected
components from a threshold mail image, merging them into
lines and grouping the text lines into clusters. The destination
address block is determined by selecting the best one from
some clusters. Eiterer et al. [2] presented a more efficient
approach based on fractal dimension. They used k-means
technique to label pixels as background, noise or semantic
objects (e.g., address block, stamp, postmark). Gaceb et al.
[3] proposed an approach for AABL based on the hierarchical
graph coloring and the pyramidal organization of data. This
architecture ensured a good coherence between the various
modules in AABL problem. Dong et al. [4] proposed a simple
segmentation algorithm for image binaryzation and used a
morphology method to remove uninteresting objects with
address block remaining. Menoti et al. [5] mainly used wavelet
decomposition to transform image into fundamental building
blocks and identified salient points from stamp, address block,
etc. The model achieved accurate and robust results.
In this paper, we propose a novel method to automatically
locate and extract address block. The main idea is to use HOG
as feature representation. Via using a joint Conditional
Random Field (CRF) and dictionary learning algorithm, we get
the probability map of address block. We evaluate our model
on a challenging Chinese postal envelope database with clutter
backgrounds. Experiment results were promising reaching a
high accurate rate over 88% .
The rest of the paper is organized as follows. In Section II
we briefly introduce some background material. We describe
our model in detail which is mainly based on HOG descriptors
and CRF model in Section III. We evaluate our model on a
Chinese postal envelope database with clutter backgrounds and
show the results in Section IV. More detailed conclusions and
future works are presented in Section V.
II. B
ACKGROUND MATERIAL
In this Section we first introduce the HOG descriptor and
explain its advantage in our model, then briefly describe
dictionary and sparse coding methods that work as feature
representation in the proposed model. Finally, we present the
CRF model.
A. Histogram of Oriented Gradient
AABL problem is a challenging task owing to their
variable appearance and clutter backgrounds. Therefore, the
critical step is to choose a discriminative feature set which can
characterize the envelope images very well. HOG is a robust
feature set due to its fine-scale gradients, fine orientation
binning, relatively coarse spatial binning, and high-quality
local contrast normalization in overlapping descriptor blocks
[10]. The basic idea of HOG is that local object appearance
and shape can often be characterized well by the distribution of
local intensity gradients or edge direction [10]. It outperforms
other existing methods including edge orientation histograms,
sift descriptors and shape context [13] for object detection and
recognition.
The envelope address blocks always have a unified
appearance: a line or multi-line of text, and thus their HOG
descriptors are usually similar, and different form the HOG