MACHINE VISION AND IMAGE PROCESSING
Year:2017 | Volume:3 | Issue:2 |Papers Count:5

Human hand movements are used in the non-verbal interaction between human and computers. Intuitive and natural hand movements is the most important factor motivating researchers to use the hands to improve the human-computerinteraction. In this paper, for hand gesture recognition, hand as the only moving object in the video is detectedusing the difference between frames. After that, hand movement feature vector is extracted. This vector is used to detect hand gesture using artificial neural network. Two methods are proposed for feature-vector extraction. The first method codes themotion trajectory of the final hand pixel in the frames. The second method uses two angle histograms. Identifying six different gestures with recognition rate of 95. 54 percent using the first method and 91. 53 percent using the second method, shows the efficiency of the proposed system. Also, the comparison between the proposed feature vectors with a conventional method shows the superiority of theproposed methods in terms of accuracy, the number of features, and classifier training time.

In order to enhance image resolution and overcome the physical limitations of imaging systems, image super-resolution (SR) is proposed. In SR technique, a high resolution (HR) image is produced by fusing a sequence of low-resolution (LR) images. Since many recent approaches seeking an SR algorithm suppresses noise whilepreserving edges, in this paper, based-upon the statistical techniques, we have proposed an adaptive algorithmwhich is robust in the presenceof Gaussian noise-which has a high detrimental effect on image quality-and moreover, shows a good performance comparing to the other existing methods. In this adaptive algorithm we introducea set of weighting coefficients, which control the contribution between the dataerror and regularization terms in each of the estimated HR pixels. These coefficients are determined according to the neighbors information of the estimated pixel. Experimental results from both synthetic and real images confirm that theproposed algorithm outperforms the other methods.

With a largeamount of multimedia content in the web, storage and retrieval of them by classical learning methods dealt with some major challenges like memory restriction. These limitations in some of learning algorithms like SVM and ANN is so serious that these algorithms cannot be employed in large-scale learning context. Kernel Extreme Learning Machine (KELM) algorithm is one of the powerful methods in machine learning. Learning phase of this method is based on constructing kernel matrix of labeled instances and calculating inverse of it. So, employing this method in large scale learning context with a lot of labeled instances is not feasible. In this research to overcome limitation of employing the KELM in large-scale multi-label learning, a new approach is proposed. The proposed approach is based on prototype selection in neighborhood of each training instance. By using the proposed approach, the size of training set is reduced. So, classical learning methods can be applied on reduced training set. Since multimedia contents are basically multi-label, the proposed prototype selection approach is based on multi-label domains like automatic image annotation. Experimental results on NUS-WIDE large-scale multi-label image set and three other versions include Object, Scene and Lite indicated the effectiveness of the proposed approach in solving the limitation of employing KELM method in large-scale multi-label learning.

A computational approach is proposed which is aimed at quantitatively assessing human recognition abilities when objects undergo different variations in the image. During the recent decades, as a result of its high accuracy and speed, human visual system has been considered an idolfora variety ofcomputational object recognition algorithms in machine vision. Therefore, quantification of its behavior in different situations can lead tobetter modeling algorithms. In this study, human ability is evaluated in an object recognition task in which object images underwent different levels of variation in lighting conditions, pose, size and position. To do this, a variation-controlled object image dataset is generated and presented to humans as well as to a computational model of visual cortex. The computational model is used to measure the effect of each variation on object recognition. Human behavioral results show a decline in recognition performance when objects underwent pose variation. The performance suppression is shown to be the result of disability of untangling object representations in highlevels of pose variation. Quantitatively speaking, images which underwent variations in lighting, pose, size and position, experienced respectively 0. 57, 0. 33, 0. 55 and 0. 73 of representational enhancement travelling from pixel to visual cortex space.

There are many methods for feature extraction from texture images. Local Binary Pattern (LBP) is one of the most important of these methods. It is a simple method for implementation and can extracttexture features efficiently. LBP can be combined with local variance (VAR) to provide higher classification rate. In this paper, a new method is proposedwhich is named Local Entropy Pattern (LEP). The equation of this method is similar to Entropy literally, butit is differ from Entropy in some issues. The proposed method is more robust to noise than LBP and VAR. In addition, by combiningit's features with LBP features the classification rate increases significantly and it provides higher accuracy than LBP/VAR. Local Entropy Pattern shows dissimilarity of a local neighborhood. This approach has all positive points of LBP and some state-of-art similar methods. It is not only rotation and grayscale invariant but also noise robust.