-This paper presents a new off-line recognition system for Arabic handwritten words. The proposed system uses scale-invariant descriptor namely SIFT, and based on an image matching technique for achieving classification. The recognition process was done through a Keypoints matching procedure, using a nearest-neighbor distance-ratio. The paper presents also a new large Arabic handwritten word database. This database provides a new framework for benchmarking and gives a new freely available Arabic handwritten word dataset. Several tests have been performed using our new database and the well known IFN/ENIT database for comparison purposes. A high correct recognition rate was reported. Index Terms-Arabic handwriting recognition, Features extraction, SIFT descriptors, Keypoints matching, New Arabic database.
Automatic recognition of handwritten scripts is an area of pattern recognition that is extremely useful in numerous fields, including documentation analysis, mailing address interpretation, bank check processing and more recently the reconstruction and recognition of historical manuscripts.
Recognition of Arabic handwriting remains one of the most challenging problems in the pattern recognition domain. Arabic is written by more than 240 million people, in over 20 different countries. The standard Arabic script contains 28 letters. Each letter has either two or four different shapes, depending on it position within a word.
One of the most challenging aspects of off-line handwriting
recognition is finding a good database that well represents
the variety of handwriting styles. Comparing with the great
number of existing databases for English script, IFN/ENIT
database [
In this research we present a new fast and robust Arabic handwriting recognition system based on SIFT descriptor and a recognizing procedure that use keypoints matching. Contrary to the majority of handwritten characters recognition systems, the proposed method operates without any preprocessing steps, since the used features are invariant regarding images’ transformations and are highly distinctive in a large database. We also introduce a new large database of Arabic handwritten words which provides a comparison tool for research works in characters recognition domain.
The remainder of this paper is divided into six sections. The next section resumes several works done in handwritten Arabic recognition field. Section 3 detail the feature extraction method and section 4 describes our new Arabic handwritten words database. Experimental results including keypoints detection and matching are reported in section 5, where a comparative analysis of the experimental results is also discussed. Finally, some concluding remarks end the paper.
The main idea of scale invariant feature descriptor (SIFT)
[
Diem and Sablatnig [
De Campos [
Zhang et al. [
In [
Character type classification on a document image problem
was addressed in [
Ramana et al. [
Mao et al. [
For Arabic handwriting recognition, we found only one
work which uses SIFT as descriptor introduced by Rothacker
et al. [
Aouadi and Kacem Echi [
Rabi et al. [
II. SIFT DESCRIPTOR
SIFT was developed by David Lowe in 2004 [
Each of these stages are executed in a descending order (cascade approach) and on every stage a filtering process is applied so that only the keypoints that are robust enough are allowed to pass to the next stage. According to Lowe, this will reduce significantly the cost of detecting the features. The descriptor is formed from a vector containing the values of all the orientation histogram entries.
For image matching and recognition, SIFT features are first extracted from a set of learning images and stored in a database. A new image is matched by individually comparing features extracted from it to those previously stocked in the database and finding candidate matching based on Euclidean distances calculated from their feature vectors. The Euclidean distance between the SIFT feature descriptors is considered as a cost measure.
The experiments conducted in this paper use a 4x4x8 = 128 elements in each feature vector of a keypoint. Regarding the image matching procedure, the local descriptors from several images are matched. A complete comparison is performed by computing the Euclidean distance between all potential matching pairs. A nearest-neighbor distance-ratio matching criterion is then used to reduce mismatches.
III. THE NEW ARABIC HANDWRITING WORD DATABASE In order to make the databases as much representative as possible, we have focused on most aspects responsible of variations of handwriting styles like the age, the sex, the educational level, the profession, the residence town, etc.
Data collection was conducted using 2100 forms. Each writer was asked to fill one form comprising 11 Algerian village names, each word is written twice. Also, there is a field for writer’s personal informations including; his name, his age, his residence town, and his profession. Each form possesses 15 exemplars. An example of a filled form represented in a grayscale level is shown in figure 2.
All the extracted images have been archived in two different
formats: grayscale and binary formats in TIFF file format
at 300 dpi resolutions. The Arabic handwritten data were
sorted and saved into four sets. Figure 3 shows some statistics
database was used as a comparison tool to evaluate
researchers’ works during the three competitions of the ICDAR
(International Conference on Document Analysis and
Recognition) organized in 2005, 2007 and 2009 [
In our study, we are not interested by the matching of two distinct images representing the same scene (or parts of the same scene) taken from two different views; our aim is to compare two images of two handwritten words whose similar contents will be in the same area, for all images representing a given word class.
The suggested method divides vertically the word images to be recognized into five frames of equal size. The objective here is to compare the detected keypoints in a given frame with its corresponding in another image representing the same word class. Figure 4 shows an example.
The number of frames was selected through different tests of several scenarios and their impact on the recorded recognition rate (table 1).
For each word class, we build a model of keypoints using 25 images as training samples. Each class model contains a given number of keypoints divided into five subsets, representing the different frames composing the word images. The construction process of each class model is detailed in the flow chart presented in figure 5. This process allows us to filter and improve the robustness of keypoints extracted from the training images of a given class.
The number of training images used to build each class model was also fixed through several experiments. We noticed that using more than 25 images during the learning process will increase the number of detected keypoints without bringing a significant improvement to the recognition rate (figure 6). concerning the number of words, sub-words and characters in each set.
A set of 128 features are extracted for each keypoint, since a keypoint descriptor consists of eight 4x4 orientation histograms. Figure 7 presents the keypoints detection process
IV. A NOVEL RECOGNITION SYSTEM using SIFT descriptors for the five frames representing a word In order to show the efficiency of the proposed system, image taken from our database. experimental tests were achieved on both databases; the Several tests were conducted in order to determine the IFN/ENIT and our new database. IFN/ENIT was produced matching ratio; this parameter fixes the matched keypoints’ by the Institute for Communications Technology at Technical number which affects the recognition rate. Tests show that the University of Braunschweig (Institut fu¨ r Nachrichtentechnik, number of keypoints and the matching ratio are rising at the IFN) and the l’Ecole Nationale d’Ine´gnieurs de Tunis. This same time (figure 8), but the discriminating capacity of these keypoints decreased. Figure 9 shows that Keypoints matching becomes more efficient when the matching ratio is fixed to 0.9 even if the number of keypoints is reduced. Worse still, the recognition rate tends to decrease when the ratio gets higher values.
The number of keypoints representing each model of the 200 used classes, with which the system registered the highest recognition rate, is given in figure 10. from features vectors by comparing the Euclidean distance of the closest neighbor to that of the second closest neighbor.
Keypoints matching of the five frames representing an image pair is illustrated in figure 11.
Once the keypoints were detected in two images, they should be paired. The best candidate match for each keypoint in the first image is found by identifying its nearest neighbor in the second one. In this work, matching keypoints are calculated
In the recognition process, each image of test must be firstly divided into five frames, then the keypoints are calculated for each frame. The matching process is then performed as follows:
Repeat the following steps for each class model and each test image: 1) Each frame representing a part of a test image is compared with its correspondent part of a class model. 2) The matched keypoints rate (MKR) is then calculated for each frame as follows:
M KR =
matched keypoints’ number detected keypoints’ number from a test image + model keypoints’ number 3) An average matching rate (AMR) is then established:
Finally, the model recording the highest average matching rate will be considered as the target class. Figure 12 shows an example summarizing these stages.
The keypoint descriptors are highly distinctive, which allows a single feature to find its correct match with good probability in a large database of features.
Tests conducted on both databases (IFN/ENIT and our new database) are listed in table 2, where we can observe that the system registered high performances with scalability, since a slight loss of approximatively 8% of the accuracy was registered when the number of classes that have to be recognized increased from 40 to 200. We also noticed that a small improvement of the recognition rate was reported during tests done on our new database compared to the IFN/ENIT database.
In order to prove the efficiency of the proposed method, we compare the obtained results with some pertinent works done on handwritten Arabic words recognition. However, only the systems tested on IFN/ENIT database have been mentioned. The reported results (table 3) show that our proposed system
The contribution of this paper is twofold. Firstly, a new large and free database for Arabic handwriting words is presented. Secondly, an effective and robust off-line handwritten Arabic words recognition system is presented and evaluated on this new database.
The developed sytem use a new type of features, namely SIFT descriptors and an efficient recognition method based on (1) (2) KNN HMM
HMM Matching based on Euclidean distance an image matching procedure. A heigh recognition rate was recorded through several experiments conducted on IFN/ENIT and our new database.