Carlos A. B. Mello

A SVM based off-line handwritten digit recognizer

This paper presents an efficient method for handwritten digit recognition. The proposed method makes use of Support Vector Machines (SVM), benefitting from its generalization power. The method presents improved recognition rates when compared to Multi-Layer Perceptron (MLP) classifiers, other SVM classifiers and hybrid classifiers. Experiments and comparisons were done using a digit set extracted from the NIST SD19 digit database. The proposed SVM method achieved higher recognition rates and it outperformed other methods. It is also shown that although using solely SVMs for the task, the new method does not suffer when considering processing time.

A new thresholding algorithm for document images based on the perception of objects by distance

In this work a new method to enhance and binarize document images with several kind of degradation is proposed. The method is based on the idea that by the absolute difference between a document image and its background it is possible to effectively emphasize the text and attenuate degraded regions. To generate the background of a document our work was inspired on the human visual system and on the perception of objects by distance. Snellens visual acuity notation was used to define how far an image must be from an observer so that the details of the characters are not perceived anymore, remaining just the background. A scheme that combines k-means clustering algorithm and Otsus thresholding method is also used to perform binarization. The proposed method has been tested on two different datasets of document images DIBCO 2011 and a real historical document image dataset with very satisfactory results.

Fast and robust skew estimation of scanned documents through background area information

Skew correction of scanned documents is a crucial step for document recognition systems. Due to the problem of high computational costs of the state-of-the-art methods, we present herein a variation of a parallelograms covering algorithm. This variation strongly reduces the computational time and works over noisy documents and documents containing non-textual elements, like: stamps, handwritten components and vertical bars. Experimental studies with different databases show that this variation overcomes well-known techniques, achieving better results over synthetic rotated documents and real scanned documents.

Segmentation of connected handwritten digits using Self-Organizing Maps

Segmentation is an important issue in document image processing systems as it can break a sequence of characters into its components. Its application over digits is common in bank checks, mail and historical document processing, among others. This paper presents an algorithm for segmentation of connected handwritten digits based on the selection of feature points, through a skeletonization process, and the clustering of the touching region via Self-Organizing Maps. The segmentation points are then found, leading to the final segmentation. The method can deal with several types of connection between the digits, having also the ability to map multiple touching. The proposed algorithm achieved encouraging results, both relating to other state-of-the-art algorithms and to possible improvements.

Automatic image segmentation of old topographic maps and floor plans

There are several kinds of information that can be achieved in ancient documents. In general, image processing research on this subject works with images of letters or documents. Topographic maps and floor plans are also an important source of information about history. In this paper, we introduce a new algorithm for image segmentation of ancient maps and floor plans. It aims to remove most part of non textual elements leaving just the text. This allows further automatic identification of the map or plan through automatic character recognition techniques. The proposed method uses a new edge detection algorithm, thresholding and connected component analysis. The results are analyzed both qualitatively and quantitatively by comparison with other technique.

Automatic line and word segmentation applied to densely line-skewed historical handwritten document images

There exists a high interest in the digitization of handwriting historical documents, in the quest to preserve the cultural heritage of nations. In general, these manuscript images present new segmentation difficulties with respect to non-historical documents. The problems come from features such as paper aging, faded ink, back-to-front ink superposition or variable line skew, among others. This paper presents a methodology for detecting and extracting the text lines of images from complex handwritten historical documents. The proposed line segmentation algorithm is based on computing a binary transition map of the document and then extracting and refining the corresponding line regions through skeletonization. To improve the accuracy of line segmentation, a new graph-based splitting method to separate the touching lines is introduced. Once text lines have been segmented, we propose an algorithm based on mathematical morphology operators and position heuristics, to extract the component words on each text line. The robustness and accuracy of our approach was tested on digitalized pages of two complex historical document datasets: the correspondence of Nabuco and the family papers of Graham Bell. We have also successfully compared our algorithms to other general line and word segmentation algorithms presented at the ICDAR 2007 Handwriting Segmentation Contest.

Segmentation of images of stained papers based on distance perception

A new approach to segment images of historical documents that are in stained paper is presented herein. Due to their characteristics, these images are very difficult to segment, especially in documents with high illumination variance along the image, non-uniform degradation and the presence of smudges or smears. We propose herein the decreasing of the foreground information (the ink) by simulating the information we perceive when we go far from the document image. As we stand back, the text tends to disappear remaining just the main colors from the background. The method is very efficient in documents with several types of degradation although it is not suitable for small noises.

Parameter tuning for document image binarization using a racing algorithm

It is investigated the use of I/F-Race to tune document image binarization methods.The method combines visual perception with the minimization of an energy function.Our experiments show that I/F-Race suggests promising parametric configurations.The binarization algorithm configured by I/F-Race outperforms other recent methods. Binarization of images of old documents is considered a challenging task due to the wide diversity of degradation effects that can be found. To deal with this, many algorithms whose performance depends on an appropriate choice of their parameters have been proposed. In this work, it is investigated the application of a racing procedure based on a statistical approach, named I/F-Race, to suggest the parameters for two binarization algorithms reasoned (i) on the perception of objects by distance (POD) and (ii) on the POD combined with a Laplacian energy-based technique. Our experiments show that both algorithms had their performance statistically improved outperforming other recent binarization techniques. The second proposal presented herein ranked first in H-DIBCO (Handwritten Document Image Binarization Contest) 2014.

Speech and phoneme segmentation under noisy environment through spectrogram image analysis

This paper presents a new algorithm for speech segmentation based on image analysis of the spectrogram of the signal. The algorithm works in two loops: the first segments the sound in search for the speech signal. The segmented speech returns to the algorithm for phoneme segmentation. For evaluation, the algorithm was applied to TIMIT speech signals with correct speech segmentation of every tested signal, including signals under real-world noise.

A Complete System for Vehicle License Plate Recognition

This work proposes a system for automatic recognition of vehicle license plate. Three phases are involved in the process: image acquisition, plate recognition and postprocessing. New algorithms for the second phase are described herein. The system first locates the plate in the image, and then it segments the characters of the plate and recognizes them. An algorithm based on vertical edge detection and color alternation in grayscale image, is proposed to locate the plate. To segment the characters, a thresholding algorithm based on Fuzzy logic and a region growth algorithm are used. Two Multi-Layer Perceptron neural networks are used to recognize the characters. The results achieved by each of these steps were over 85% of correctness, what shows the robustness of the proposed algorithms, making them suitable for real world applications.