نشریه ماشین بینایی و پردازش تصویر
سال یکم شماره 2 (پاییز و زمستان 1392)

Edge detection is one of the most important and basic problems in image processing and image analysis and machine vision. The importance of this issue is that the output of many image analysis algorithms are dependent on accurately and correctly edge detection algorithm. However a comprehensive and perfect solution that be able to do edge detection with considering all of the edge patterns and factors affecting it, so far not provided. In this paper, firstly, inspired by computational model of retina and primary visual cortex cells in the brain, we propose an appropriate Gabor function as a gradient operation different directions and scales with adaptive parameters. Then, for fusion of rectified gradient responses, we introduce an algorithm based on edge segments contraction from large-scales toward small-scales. This research is trying to satisfy three criteria introduced by Canny properly. With the implementation of multi-scale and multi-directional analysis and contractive fusion of image gradients, better detection and more precise positioning of edges has been obtained. The experimental results indicate that the effective ness of this method on natural and artificial images is better than the conventional methods (e.g. Canny).

In this paper, new visual features for personality identification based on Farsi handwriting images are proposed. The purpose of this study is to establish a computer aided system to automatically extract some attributes that are used by graphologists for the analysis of handwriting. Therefore, these attributes should be presented in the forms that be extractable from the handwriting images. To do this, first the handwriting images are preprocessed and the features such as pen width, spaces between lines, heights of sub-words, etc. are extracted. In the next step, some features including spaces between words, the size of letters and sub-words, the deviation of letters and the height of vertical letters are calculated. These features are employed to recognize the type of handwriting. Eventually, the extracted features from handwriting images are examined on a set of images which have been commented on by graphologist. The experimental results confirm the effectiveness of the proposed features.

In the recent years, due to the growth and development of the Internet, 3D model retrieval has become a wide research field. Several methods have been presented to retrieve 3D models. Each method employs a special feature. In this paper, a geometric based method is proposed to retrieve the objects. In 3D models, the adjacent points on the surface of object are highly correlated. So, the linear prediction analysis has a proper performance to estimate these models. In this paper, the linear prediction coefficients are extracted from cylindrical projections of 3D objects as an appropriate descriptor. At first, the object is projected onto the lateral surface of a bounding cylinder. Then, the linear prediction coefficients are extracted from cylindrical projection. To alleviate the occlusion problem, cylindrical projection is generalized to three cylinders along all principal axes. The principal component analysis is employed to normalize the descriptor against rotation. The performance of the proposed descriptor is evaluated employing the McGill database of 3D models. Experimental results demonstrate that the proposed method discriminates the objects with different structure and remains robust against noise as well.

A semantic video analysis system is a semi- or fully-automated system for investigation of video contents and extraction of concepts and semantics from video. These systems can be classified from different viewpoints. In this article, the semantic gap between low-level features and high-level concepts in the semantic video analysis systems is discussed and the literature is reviewed from viewpoint of semantic hierarchy for video production.To this end, after a brief description of video analysis systems and their general block diagram, two main challenges of these systems are discussed: sensory gap and semantic gap. Then, different approaches for reduction of the semantic gap are studied based on a semantic hierarchy used for video production. According to this hierarchical structure, there are three main steps for reduction of semantic gaps: frame processing; content analysis; and semantic extraction. Finally, open problems in this field of research are presented as the following. Usually, a wide variety of events and concepts may occur in a video. Different concepts may be assigned to an event in different circumstances. Some high-level concepts may occur in a relatively long duration of video; thus, extraction of such concepts requires processing of a relatively long duration of video to construct a semantic network between the concepts extracted from short durations of video. Usually, analysis of multi-modal information may close the semantic gaps that exist in the analysis of single-modal information.

In this paper, we propose a system for recognizing static signs of Persian sign language (PSL). This system, designed based on a novel geometric feature, can automatically recognize static signs of Persian sign language alphabet with high accuracy. In feature extraction stage, we first find the center of gravity and contour of the hand shape in images, and then sample from obtained hand contour points. Then we consider a circle for each point on the contour. For this purpose, we compute the distance between the center of gravity of hand shape and each point along the contour. We select half of the computed distance for each point and the point located in the middle of this distance as radius and center of circle respectively. The resulting circles contain valuable information about the shape of hand that they are organized as feature vector for each sign. To recognize signs with high reliability, we propose a system that combine support vector machine (SVM) with K nearest neighbor (KNN). Experiments are performed on PSL database. The accuracy and reliability of proposed system with the test data are 93.33% and 98.73% respectively. The obtained results show the performance of proposed system is satisfactory.

Studies show that 25% to 47% of elderly will at least once experience falls and this figure is approximately 50% among the elderly living in nursing home. In this paper, based on the Gaussian Mixture Model (GMM) and estimating their parameters by Expectation Maximization (EM) algorithm, a new method has been proposed that firstly, the binary movement of the elderly is segmented from video sequences. Next, the occurrence of falls in older persons is done relying on anatomic body changes and Motion History Images (MHI). Elevation of the system performance was set up on a set of video frames received from the elderly residing in Mother Health Care Center in Sabzevar city and CAVIAR database containing the actual occurrence the of falling. Then, based on the standard deviation and the C-motion coefficient of the walking, suspected incident falls and actual falls are accurately segregated and finally, the sensitivity of 92.68% and the specificity of 96% were obtained which represent a desired capability of the output system. In overall, appropriate simulation of algorithms on the data set due to low error rate in which is less than 6% and meanwhile a careful monitoring of the elderly’s falls will be provided by implementing this system in elderly nursing and residential homes.