Handwritten Chinese Text Recognition by Integrating Multiple Contexts

ABSTRACT:
This paper presents an effective approach for the offline recognition of unconstrained handwritten Chinese texts. Under the general integrated segmentation-and-recognition framework with character over segmentation, we investigate three important issues: candidate path evaluation, path search, and parameter estimation. For path evaluation, we combine multiple contexts (character recognition scores, geometric and linguistic contexts) from the Bayesian decision view, and convert the classifier outputs to posterior probabilities via confidence transformation. In path search, we use a refined beam search algorithm to improve the search efficiency and, meanwhile, use a candidate character augmentation strategy to improve the recognition accuracy. The combining weights of the path evaluation function are optimized by supervised learning using a Maximum Character Accuracy criterion. We evaluated the recognition performance on a Chinese handwriting database CASIA-HWDB, which contains nearly four million character samples of 7,356 classes and 5,091 pages of unconstrained handwritten texts. The experimental results show that confidence transformation and combining multiple contexts improve the text line recognition performance significantly. On a test set of 1,015 handwritten pages, the proposed approach achieved character-level accurate rate of 90.75 percent and correct rate of 91.39 percent, which are superior by far to the best results reported in the literature.


ARCHITECTURE:

EXISTING SYSTEM:
ANDWRITTEN Chinese character recognition has long been considered a challenging problem.

Handwritten Chinese text recognition (HCTR) is a challenging problem due to the large character set, the diversity of writing styles, the character segmentation difficulty, and the unconstrained language domain.

Both isolated character recognition and character string recognition have been studied intensively but are not solved yet.
In isolated Chinese character recognition, most methods were evaluated on data sets of constrained writing styles though very high accuracies (say, over 99 percent on Japanese Kanji characters and over 98 percent on Chinese characters) have been reported [1]. The accuracy on unconstrained handwritten samples, however, is much lower In Chinese character string recognition, most works aimed at the recognition of text lines or phrases in rather constrained application domains, such as legal amount recognition in bank checks [4] and address phrase recognition for postal mails [5], [6], [7], [8], where the number of character classes is very small or there are very strong lexical constraints.

DISADVANTAGES OF EXISTING SYSTEM:
The large set of Chinese characters (tens of thousands of classes) brings difficulties to efficient and effective recognition. The divergence of writing styles among different writers and in different geographic areas aggravates the confusion between different classes. Handwritten text recognition is particularly difficult because the characters cannot be reliably segmented prior to character recognition. The difficulties of character segmentation originate from the variability of character size and position, character touching and overlapping.

A text line of Chinese handwriting must be recognized as a whole because it cannot be trivially segmented into words (there is no more extra space between words than between characters).

Last, handwritten text recognition is more difficult than bank check recognition and mail address reading because the lexical constraint is very weak: Under grammatical and semantic constraints, the number of sentence classes is infinite.

PROPOSED SYSTEM:
In this study, we investigate three key issues of integrated segmentation-and-recognition for HCTR: candidate path evaluation, path search, and parameter estimation. By elaborating the techniques for these issues, we achieved significant improvements on unconstrained handwritten Chinese texts. In path evaluation, we integrate character recognition scores, geometric context, and linguistic context from the Bayesian decision view, and convert the classifier outputs to posterior probabilities via confidence transformation (CT).
ADVANTAGES OF PROPOSED SYSTEM:
In path search, a refined beam search algorithm is used to improve the search efficiency and, meanwhile, a candidate character augmentation (CCA) strategy is applied to benefit the recognition accuracy.

To balance the multiple contexts in path evaluation function, we optimize the combining weights on a data set of training text lines using a Maximum Character Accuracy (MCA) criterion.

We evaluated the recognition performance on a large database CASIA-HWDB [21] of unconstrained Chinese handwritten characters and texts, and demonstrated superior performance by the proposed methods.



MODULES:
1) Draw Panel Creation Module
2) Over-Segmentation Module
3) Character recognition
4) Ranked List
5) Result String definition Module

MODULES DESCRIPTION:

1) Draw Panel Creation Module
Draw Panel Creation module is the first module, where we create the interface for the users to provide the input texts. The user can able to draw the texts in the panel, using the panel area provided. In case if the text is given wrongly then the options are provided to clear the panel and so that the user can able to input a new text to recognition.



2) Over-Segmentation Module
First, the input text line image is oversegmented into a sequence of primitive segments using the connected component-based method. Consecutive primitive segments are combined to generate candidate character patterns, forming a segmentation candidate lattice After that, each candidate pattern is classified to assign a number of candidate character classes, and all the candidate patterns in a candidate segmentation path generate a character candidate lattice. If a word level language model is used, each sequence of candidate characters is matched with a word lexicon to segment into candidate words, forming a word candidate lattice. All of these character (or word) candidate lattices are merged to construct the segmentation-recognition lattice of text line image. Each path in this lattice is constructed by a character sequence paired with a candidate pattern sequence, and this path is called a candidate segmentation recognition path. Finally, the task of string recognition is to find the optimal path in this segmentation-recognition lattice. Considering that the text lines are segmented from text pages, we utilize the linguistic dependency between consecutive lines to improve the recognition accuracy by concatenating multiple top-rank recognition results of the previous line to the current line for recognition.


3) Character recognition
Considering that Chinese texts mix with alphanumeric characters and punctuation marks and different characters show distinct outline features (e.g., size, position, aspect ratio, and within-character gap), we design two classdependent geometric models, namely, single-character geometry (unary geometric model) and between-character geometry (binary geometric model), respectively. In addition, two class-independent geometric models are designed to indicate whether a candidate pattern is a valid character or not, and whether a gap is a between-character gap or not, respectively. The four geometric models (unary and binary class-dependent, unary and binary class-independent) are abbreviated as “ucg,” “bcg,” “uig,” and “big,” respectively, and have been used successfully in transcript mapping of handwritten Chinese documents
4) Ranked List
We evaluated the effects of different techniques. First, we compared the effects of different path evaluation functions. Second, the effects of different confidence transformation methods, combinations of geometric models and language models were evaluated. Last, we show the results of different numbers of candidate character classes, beam widths, and candidate character augmentation methods in path search.

5) Result String definition Module
This module presented an approach for handwritten Chinese text recognition under the character over segmentation and candidate path search framework. We evaluate the paths from the Bayesian decision view by combining multiple contexts, including the character classification scores, geometric and linguistic contexts. The combining weights of path evaluation function are optimized by a string

The experimental results justify the benefits of confidence transformation of classifier outputs, geometric context models, and language models. Nevertheless, the effect of candidate character augmentation is limited. We also evaluated performance. The objective of over segmentation is to improve the tradeoff between the number of splitting points (affecting the complexity of search space) and the accuracy of separating characters at their boundaries. The objective of character classification is to improve the classification accuracy and the tradeoff between the number of candidate classes and the probability of including the true class.
SYSTEM REQUIREMENTS:
HARDWARE REQUIREMENTS:
PROCESSOR        :    PENTIUM IV 2.6 GHz
RAM                    :         512 MB
MONITOR           :         15”
HARD DISK         :       20 GB
CDDRIVE            :         52X
KEYBOARD         :       STANDARD 102 KEYS
MOUSE                :         3 BUTTONS

SOFTWARE REQUIREMENTS:
FRONT END                  :    JAVA, SWING, APPLET
TOOLS USED                :    JFRAME BUILDER
OPERATING SYSTEM:    WINDOWS XP
REFERENCE:

Qiu-Feng Wang, Fei Yin, and Cheng-Lin Liu, Senior Member, IEEE, “Handwritten Chinese Text Recognition by Integrating Multiple Contexts”, IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 34, NO. 8, AUGUST 2012.