Majid Komeili

A novel hybrid algorithm for binarization of badly illuminated document images

In this paper, we present a novel hybrid algorithm for binarization of badly illuminated document images. This algorithm locally enhances the document image and makes the gray levels of text and background pixels separable. Afterward a simple global binarization algorithm binarizes the enhanced image. The enhancement process is a novel method that uses a separate transformation function to map the gray level of each pixel into a new domain. For each pixel, the transformation function is determined using its neighboring pixels gray level. The proposed binarization algorithm is robust for wide variety of degraded document images. Evaluation over a set of degraded document images illustrates the effectiveness of our proposed binarization algorithm.

Local Feature Selection for Data Classification

Typical feature selection methods choose an optimal global feature subset that is applied over all regions of the sample space. In contrast, in this paper we propose a novel localized feature selection (LFS) approach whereby each region of the sample space is associated with its own distinct optimized feature set, which may vary both in membership and size across the sample space. This allows the feature set to optimally adapt to local variations in the sample space. An associated method for measuring the similarities of a query datum to each of the respective classes is also proposed. The proposed method makes no assumptions about the underlying structure of the samples; hence the method is insensitive to the distribution of the data over the sample space. The method is efficiently formulated as a linear programming optimization problem. Furthermore, we demonstrate the method is robust against the over-fitting problem. Experimental results on eleven synthetic and real-world data sets demonstrate the viability of the formulation and the effectiveness of the proposed algorithm. In addition we show several examples where localized feature selection produces better results than a global feature selection method.

TED: A texture-edge descriptor for pedestrian detection in video sequences

This paper presents a novel descriptor, TED, for pedestrian detection in video sequences. TED describes texture and edge information simultaneously. TED is a local descriptor because it is defined over a neighborhood. The size of the TED, independent of the neighborhood size defined over it, is 8 bits. TED is based on intensity difference, and so it is robust against illumination changes. We demonstrate TED performance in a block-based framework for pedestrian detection. Experimental results show the effectiveness of the proposed descriptor when applied in different outdoor and indoor environments.

Degraded document image binarization based on combination of two complementary algorithms

In this paper we combine two binarization algorithms that are complementary to each other. The main idea is to select the better algorithm in each part of document image. There are algorithms that properly distinguish the text from the background in the regions close to the text, but get wrong in the regions far from the text and introduce some part of background as text. We propose a new binarization algorithm that effectively eliminates background and reliably extracts some parts of each character. Then according to the distance of each pixel form the text, the appropriate algorithm is selected to binarize that pixel. Proposed method is applicable for various types of degraded document images. After extensive experiment, the proposed binarization algorithm demonstrate superior performance against four well-know binarization algorithms on a set of degraded document images captured with camera.

A non-parametric pixel-based background modeling for dynamic scenes

Background modeling is one of the most important parts of visual surveillance systems. However, making the pixels background model can be problematic, especially in high resolution video sequences of dynamic scenes. We proposed an efficient non-parametric method for modeling which consumes little computational resources and memory. In our method, the history of each pixel is modeled by a 2×Kp matrix, where Kp changes adaptively with complexity of that pixel. Experimental results demonstrate the effectiveness of the algorithm when applied in complex real-world sequences.

TED: A Texture-Edge Descriptor based on LBP for pedestrian detection

This paper presents a novel texture-edge descriptor, TED, for background modeling and pedestrian detection in video sequences which models texture and edge information of each image block simultaneously. Each image block is modeled as a group of adaptive TED histograms that are calculated for pixels of block over a rectangular neighborhood. TED is an 8-bit binary number which is independent of the neighborhood size. Experimental results over real-world sequences from PETS database clearly show the better performance of our method.

Effective hierarchical background modeling and foreground detection in surveillance systems

Background modeling is one of the most important parts of visual surveillance systems. Most background models are pixel-based which extract detailed shape of moving objects, but they are so sensitive to non-stationary scenes. In many applications there is no need to detect the detailed shape of moving objects. So some researchers use block-based methods instead of pixel-based which are more insensitive to local movements. These two methods are complementary to each other. We propose an efficient hierarchical method by which the block level information is utilized intelligently to improve the efficiency and robustness of pixel level. Experimental results demonstrate the effectiveness of the algorithm when applied in different outdoor and indoor environments.

A fuzzy approach for multi-feature pedestrian tracking with particle filter

Particle filter is one of the best methods of object tracking in video sequences. Particle filter usually is used with only one feature. In this paper, we propose a novel method for multi-feature object tracking in a particle filter framework. A fuzzy inference system by which reliability of features can be measured has been designed. This is done based on observations diversity and spatial scattering of particles. The features are combined in proportion to their reliabilities. Efficiency of our algorithm is demonstrated using color, edge and texture features. Experimental results over a set of real-world sequences show that our methodpsilas performance is better than some other solutions proposed for feature weighting.

Vigilance lapse identification using sparse EEG electrode arrays

Mental vigilance monitoring is useful in helping people improve their prowess in situations such as sports, musical performance or occupational demands. Electroencephalogram (EEG) signals provide appropriate information for identification of vigilance lapses; however, almost all of the previous EEG-based vigilance-monitoring studies require high-density EEG montages, with the added inconvenience of conductive gels and the requirement for accurate placement of electrodes. In this paper we propose a practical machine learning approach for identification of vigilance lapses using EEG signals recorded from a very sparse electrode configuration with only 4 electrodes: 2 electrodes at the forehead and 2 electrodes behind the ears. This is a challenging problem since these four electrodes are easily contaminated by eye blinks and muscle artifacts. The performance of the proposed machine-learning based algorithm is demonstrated in a real world scenario where vigilance lapses are identified with about 95% accuracy.

Document image binarization by using texture-edge descriptor

In this paper we propose a new approach for text region extraction in camera-captured document images. Texture-Edge Descriptor, TED, is utilized for text region extraction. TED is an 8-bit binary number which its bits are structural. This structural bits and special text region characteristics in document images make TED an appropriate descriptor for text region extraction. Applying well-known water flow method to the text regions extracted by TED, results in fast and good quality document image binarization. Experimental results demonstrate the effectiveness of our method for text region extraction and document image binarization.