Download paper

Signature Segmentation and Recognition from Scanned Documents

Categories: Plants


In this signature detection paper basically we detect signature using two approach. First one in which signature blocks are segmented from scanned document using wordwise component extraction and classification. In which Some Gradient based features are extracted from each segment at the word level to perform the classification .

Scale-Invariant Feature Transform( SIFT)

While in second approach we use some standard algorithm such as Scale-Invariant Feature Transform( SIFT) as descriptors and Spatial Pyramid Matching (SPM) based approaches are used for signature recognition.

Both approaches Support Vector Machines are employed as the classifier for both levels in this experiment. Basically this experiments are performed on Tobacco-800 and GPDS datasets and the results obtained from the experiments are promising.

Signature is used as unique identity of a person which provide information about a person as they consist of unique properties of human behaviour. So that we can used in verification or authentication purposes. In a document signature used as verification so that we restrict fraud.

Today world large quantities of administrative documents scanned and then merged as image so that space is reduced. Due to all data are stored as a scanned image so we need a tremendous demand for robust ways to access and manipulate the information that these images contain.

The main objective is to obtain the information resources relevant to the query information from such repositories. A document scanned from the dataset is also find by a signatures in which sign used as key information for searching and retrieval of documents.

Top Experts
Bella Hamilton
Verified expert
5 (234)
Professor Harris
Verified expert
4.9 (457)
Verified expert
4.8 (756)
hire verified expert

Thus, the handwritten signature will give as a advantage for document indexing and searching. Hence, segmentation and recognition of signatures from documents is very significant because of its various applications.

During segmentation and recognition of signatures from scanned documents have been found to be a challenging task. We faced a problem in Separation of handwritten annotations from scanned documents .so we discussed and proposed a Signature detection in scanned a multi-scale structural saliency approach to capture the dynamic curvature using a signature production model for signature detection and segmentation.

Signature segmentation techniques

And after further studies we reached another Signature segmentation techniques from machine printed documents that is to segment signatures from bank cheques and other documents based on sliding window to calculate the entropy and finally fit the window to signature block. But there also major issue of this technique is that we assumed priori information of the location of the signature Then after remember these issue proposed a Speeded Up Robust Features (SURF) based approach for signature segmentation from document images.

In which signature document retrieval methods have been discussed and describes a method for document image decomposition and retrieval based on connected component analysis and geometric properties of the labelled regions.In which Scanned documents have a Arabic & Persian signature are taken for the experiment.

Then proposed a method on signature-based retrieval of scanned documents. A model based on Conditional Random Fields is used to label extracted segments of scanned documents as machine-printed, signature and noise. Now then we further classification technique based on Support Vector Machine is used to remove noise and printed text overlapping on the extracted signature images. Finally, a global shape-based feature is computed for each signature image.

Now presented a signature based document retrieval technique from documents with cluttered background. Then we extract some feature from each blob and the K-means clustering algorithm is used to create the codebook of blobs. During retrieval, Generalized Hough Transform is used to detect the query signature and a voting is casted to find possible location of the query signature in a document.

This paper proposes a two-stage approach for signature segmentation and recognition. Gradient-based features and the SVM classifier are used for signature segmentation. SIFT descriptors with an extended version of the Bag-of-Features (BoF) algorithm is employed for signature recognition task.

Methodology :

Signature detection technique have two stage first image segmentation second recognition of a signature.

Signature Segmentation :

During signature segmentation it contain marks or stroke large as compare to the strokes of the printed text. So due to this distinct feature of signature is very important to get the difference of signature from printed strokes and it is used to discriminate signature from printed text.

Then we apply proposed method for a signature segmentation from a machine printed scanned document. Which is also two step procedure and is proposed for signature segmentation. At first stage signature blocks are extracted. Then signature blocks are processed in order to remove non-signature components, such as touching or overlapping printed name. Following are the classification technique Block-wise word extraction, word level feature extraction are apply which are as follow.

The scanned documents are in grayscale so we have to convert image as 0 and 1 so we used Otsu’s threshold selection method which convert document images into two-tone 0 ; 1. Then in order to correct skew of document we used Hough transform-based. Now we have binarized document images and are segmented into words based on the inter-character spacing between words.

Now we apply a morphological dilation operation so that the we can apply connected component labelling method in order to find bounding boxes of the word patches on the dilated document image. So that the compute features feed into svm. So that we easily classify image through a gradient based feature extraction technique and SVM as classifier to classify those segmented words as signature or printed words. Thus gradient based features is obtained by a grey scale local orientation histogram of the component . After classification we estimated result by a Gaussian kernel.

k(x, y) = exp ? |x ? y|2

Finally printed text characters have overlapping of the signature block and are touched with the signature which are removed using touching character analysis in hypothetical zones of printed text characters.

Signature Recognition:

Which is a efficient patch-based SIFT descriptors with Spatial Pyramid Matching (SPM)-based pooling scheme is applied for the proposed signature recognition task. The feature extraction module has three components. A flow diagram of signature recognition system is presented in Fig. 3. First, SIFT descriptors are extracted from the signature and quantised using the K-means clustering algorithm. Next, the SPM-based scheme is applied for the representation of an image. Finally, the SVM is employed for classification. The modules are described in the following sub-sections.

SIFT descriptor:

it is a technique for detecting salient, stable feature points in an image. • For every such point, it also provides a set of features” that “characterize/describe” a small image region around the point. These features are invariant to rotation and scale

The SIFT is a local shape descriptor to characterize local gradient information. Here 128-dimensional vector for each SIFT keypoint is extracted which stores the gradients of 4 × 4 locations around a pixel in a histogram bin of 8 directions. The SIFT descriptor is scale and rotation invariant. The gradients are aligned to the main direction, which makes it a rotation invariant descriptor. Different Gaussian scale spaces are considered for the computation of a vector to make it scale invariant.

The blue circles in Fig. 4(a) represent the 16 × 16 SIFT patches and Fig. 4(b) 1) SIFT descriptor: The SIFT (Scale-Invariant Feature Transform) [19] is a local shape descriptor to characterize local gradient information. Here, 128-dimensional vector for each SIFT keypoint is extracted which stores the gradients of 4 × 4 locations around a pixel in a histogram bin of 8 directions. The SIFT descriptor is scale and rotation invariant. The gradients are aligned to the main direction, which makes it a rotation invariant descriptor. Different Gaussian scale spaces are considered for the computation of a vector to make it scale invariant. The blue circles in Fig. 4(a) represent the 16 × 16 SIFT patches and Fig. 4(b)

Spatial Pyramid Matching (SPM): The SPM is an extended version of Bag-of-Features (BoF) model, which is simple and computationally efficient. As BoF model discards the spatial order of local descriptors, it restricts the descriptive power of the image representation. The limitation of BoF is vanquished by SPM [20] approach, which is successfully applied on image categorization tasks. An image is partitioned into 2 l × 2 l segments where l = 0, 1, 2, …., n; represents different resolutions.

Next, the BoF histograms are computed within each of the 2 l segments, and finally, all the histograms are concatenated to form a vector representation of the image. SPM reduces to BoF, when the value of the scale l = 0. Here, the pyramid matching is performed in two-dimensional image space and use a traditional clustering technique in feature space. The number of matches at level l is given by the histogram intersection function: I(HX, HY ) = ? D i=1 min(HX(i), HY (i)) (1) Finally, the representation of the image for classification is the total number of matches from all the histograms, which is given by the definition of a pyramid match kernel: K?(X, Y ) = ? L i=1 1 2 i (Ii?1 ? Ii) (2) 3)

Feature Extraction and Classification: This section briefly describes the feature extraction method from signature for signature recognition. First, the signature image is divided into 16 × 16 patches. The higher dimensional SIFT descriptors of 16 × 16 pixel patches are computed over a patch. Next, K-means clustering technique is applied on the patches from the training set for the generation of codebook. The typical vocabulary size for our experiments is 1024. Finally, SPM scheme is employed to generate the feature vector, which is then fed to the SVM classifier.

In our experiment, the image is divided into 2 l × 2 l segments in three different scales l = 0, 1, 2. 21 (16+4+1) BoF histograms are computed from these three levels, and all the histograms are concatenated to get the final vector representation of an image. The equation below represent the pyramid match kernel for three scales: K? = I2 + 1 2 (I1 ? I2) + 1 4 (I0 ? I1) (3) The SVM using the one-versus-all is employed for multiclass signature classification. The signature recognition experiment is repeated 6 times with different randomly selected training and test images.

. Experimental Dataset 3080 signatures from GPDS dataset and 7684 English printed words collected mainly from books, newspapers, magazines, journals, etc. are used to train the SVM classifier for signature detection task. For the testing of signature segmentation method, all the signed documents from the dataset of ‘Tobacco-800’ industrial archives [2] is used.

The documents are written in English and the signatures on these documents also contain handwritten English characters. 300 classes of genuine signatures from GPDS dataset and 50 classes of Devnagari genuine signatures are used in the experiment. 24 genuine signatures are available in each class. B. Signature Segmentation The features computed from the patches obtained from morphological dilation of ‘Tobacco-800’ dataset, are fed to classifier and an overall accuracy of 95.58% is achieved for signature block detection.

The errors are mainly due to segmentation problem at block level. Some broken parts of signatures are identified as non-signature and some patches which contain printed words of two consecutive rows are misclassified as signature block. To get a comparative idea, the performance of our proposed method and the performance of an earlier similar work on the same dataset are given in Table I.


Approach Dataset Accuracy (%) Multi-scale structural saliency [15] Tobacco-800 92.80 Conditional Random Field [14] 101 documents 91.20 Proposed Method Tobacco-800 95.58 C. Signature Recognition The signature recognition experiment on GPDS dataset demonstrates the excellent performance of our proposed approach. Table II shows the results when the experiment is repeated for 6 times for both the datasets using Linear SVM as a classifier.

First and second rows show the results on 300 and 100 classes of GPDS signature dataset. The third row shows the accuracy from the experiment on 50 classes of Devnagari signature dataset. Overall 99.95%, 99.98%, 99.60% accuracy have been achieved from 300 classes, 100 classes of GPDS and 50 classes of Devnagari dataset, respectively. The ratio between True Positive Rate (TPR) and False Positive Rate (FPR) (i.e. ROC curve) is presented in Fig. 5. It shows the performance of the signature recognition experiment of English and Devnagari scripts, which is based on the combination of SIFT descriptor, SPM with the SVM classifier. Table II 6-FOLD CROSS VALIDATION


. F1-F6 REPRESENT THE ACCURACY IN PERCENT OF 6-FOLDS Signature Data F1 F2 F3 F4 F5 F6 GPDS (300) 99.97 99.97 99.91 99.96 99.97 99.96 GPDS (100) 99.98 99.99 99.98 99.98 99.98 99.95 Devnagari (50) 99.53 99.53 99.56 99.53 99.71 99.75 The signature recognition experiments have also been performed under two other configurations. The results obtained from the experiments are shown in Table III. A 54.33% accuracy is obtained from a HMM-based classification technique. Geometrical features [21] are computed and are fed to HMM classifier. The gradient-based feature and the SVMbased classification technique, which are used for signature segmentation, have also been employed for the signature recognition task and a 69.80% accuracy is obtained. The proposed method outperformed the results obtained on the
experiments using 300 classes of genuine signatures from the GPDS dataset.


Approach Accuracy (%) Geometrical features [21] and HMM classifier 54.33 Gradient-based feature and SVM classifier 69.80 Proposed Method 99.95 The previously proposed approaches on signature segmentation and recognition have been tested on different publicly available datasets such as “Tobacco-800” and a few experiments have been conducted on the author’s own collected dataset. A recall of 78.4% and 84.2% precision is reported by Srinivasan and Srihari [14] for the signature based document retrieval task. 96.13% accuracy is reported by [13] on Arabic/Persian documents. In [15], 93.20% MAP and 89.5% MRP have been reported for document retrieval based on signatures.


Signature segmentation and recognition is a task of interest for content-based document retrieval based on signature information. In this paper, we propose an approach for efficient segmentation and recognition of signatures from document images. The signature region is detected in machine printed documents using the classification of components at the word level. The gradient-based feature and the SVM classifier are employed for signature detection. The signature recognition task is performed using SIFT descriptors with an SPM scheme. The empirical results of the experiments are encouraging and compare well with other state-of-the-art approaches in the literature.

Cite this page

Signature Segmentation and Recognition from Scanned Documents. (2020, Jun 02). Retrieved from

Are You on a Short Deadline? Let a Professional Expert Help You
Let’s chat?  We're online 24/7