علی آقا گل زاده

In pansharpening, a low-resolution multispectral image (LRMS) and a high-resolution panchromatic image (PAN) are fused to produce a high-resolution multispectral image (HRMS). Recent studies have shown that convolutional neural networks can be used for sharpening remote sensing images with excellent results. However, there are still two major problems that need to be resolved. Despite the lack of ideal HRMS images for learning, most current methods require more effort to generate simulated data. On the other hand, these methods usually ignore the rich spatial information in panchromatic images. To address these issues, we have proposed an unsupervised fusion framework based on a multiscale dense network called UMP-GAN for pansharpening. This framework, which employs generative adversarial networks, can be trained directly on full-resolution images without requiring any preprocessing step. First, a multiscale dense generator network is proposed to extract features from original input images to generate HRMS images. In the following, two separate discriminator networks are used to protect the spectral information and spatial details of the input images in the final image. Therefore, the proposed method provides the possibility of training two discriminating networks, each of which has a different and complementary task. Finally, new lost functions are proposed to perform training under unsupervised settings. This method can deepen the exchange of gradient information between the generator network and the discriminator networks. The results from WorldView-2, GaoFen-2, and QuickBird satellite images show that the proposed method performs better than other state-of-the-art models in the fusion of remote sensing images.

A challenging aspect of video error concealment (VEC) is reducing blockiness around the missing region. In this paper, we have devised a novel method for calculating the boundary matching distortion based on the edge direction surrounding the lost area to deal with this problem. After detecting the corresponding slice’s erroneous Macroblock (MB), it is divided into four 8×8 pixels sub-blocks. Then, the gradient of outer boundary pixels is derived and determines each side’s smoothness, which is further used for drawing the hypothetical line. Furthermore, this paper proposes a novel optimization-based model that can accurately measure boundary matching error for loss recovery. We observe better results for our proposed technique than other related VEC algorithms in terms of PSNR and SSIM. The proposed algorithm improves the average PSNR by 1.09 dB and increases the average SSIM by 0.0135 at the packet loss rate of 10%. In addition, for a PLR of 20%, the PSNR is increased by 1.28 dB, and the SSIM enhancement is 0.019. This algorithm is slightly more computationally complex than the compared methods, but it is still acceptable. Thus, by adding a few computations to the video decoder, the proposed method maintains the quality of the video, especially in the rough, damaged regions of a decoded frame.

Recently, a method for multi-exposure images fusion based on structural decomposition of images into three parts including signal strength, signal structure and signal mean has been introduced. In this paper, we seek to use this decomposition, for images fusion in other fields, including multimodal medical, multi-focus, and infrared and visible images. To increase the fusion quality, besides the introduction of the proposed weighting factor in the structural decomposition, contourlet transformation and the pyramidal structure have also been used. First, each of the K input images are represented into low frequency and high frequency subbands, by using contourlet transform. Then, all the corresponding subbands (resulting from the same scales and directions) are fused with each other, separately and in an iterative process. In this iterative process, first, a separate pyramid structure (including approximation and detail layers) is created for each of the corresponding K subbands. These layers are obtained by the down-sampling of subbands and structural separation based on the proposed new weighting factor. Then, the fusion is performed in the reverse direction of the pyramidal structure and the fused image of the K corresponding subband is obtained. By repeating this process, the fused image will be obtained for all the corresponding subbands. At the end, the final fused image is obtained by the inverse contourlet transformation on the fused images of the subbands. Several visual and quantitative comparisons, with 7 common methods in this field, have been made. In the visual aspect, the proposed method shows the highest quality.

Nowadays, unique identifiers are used to authenticate individuals. These identifiers are called biometrics with iris being one of the most popular ones. Research performed in this paper consist of designing a new identification system based on improved security in iris database. In this paper, we propose an iris template protection method based on Visual Cryptography (VC) and RSA. For iris template protection, the binary iris template is divided into two shares using VC, where one share is stored in the database and the other is kept with the user on a smart    card. In order to provide additional security, the share that is stored in the database is encrypted with RSA encryption. In the next step the patterns obtained from each image are compared with the patterns in the database and, by defining a threshold, the identification procedure is carried out. In this paper, the CASIA - v1 image bank has been used. The EER value after encryption is approximately 0.01. We then expose the system to different noises such as salt and pepper, white Gaussian noise and mean filt.r. After applying the noises, we haveit is observed that the system has shows been better performance compared to the than other systems which we have investigated.

Local smoothness and nonlocal self-similarity of natural images are two main priors in single image super resolution (SISR) problem. Although local sparsity is efficiently utilized to describe the local smoothness, but ignoring the correlation between the sparse representation coefficients of similar patches can lead to inaccurate sparse coding coefficients. In this paper, we propose the method that enforce the local smoothness and nonlocal self-similarity by sparse representation in a unified framework, called adaptive group-based sparse domain selection (A-GSDS). Nonlocal patches with similar structures are exploited and stacked in the form of matrix as the basic unit of sparse representation called group. These groups are converted into a column vector, each column selects the best fitted PCA sub dictionary which is learned from the training data. The sparse coding process for each column in the domain of group leads to find sparse vectors which can be easily estimated by the selected orthogonal sub dictionaries. To further improve the performance of the group-based sparse representation, we use nonlocal means regularization term.  Extensive experimental results validate the effectiveness of the proposed method comparing with the state-of- the-art algorithms.

Blind recognition of error correction codes parameters from intercepted bit-stream at the receiver side, is highly considered in military and commercial applications. In fact, identification of the encoding scheme used in the transmitter without any prior information, is a challenging task to the adversary. Several methods have been presented for blind code recognition. In this paper, a statistical method for recognition of  the length of the code word and the length of the block of information is presented. This scheme not only is resistant to error, but also its performance sustains in long codes. In this work, the method has been tested using some clustering algorithms such as K-Means and Jenks Natural Breaks. Then, a novel method to extract features of systematic binary linear block codes has been presented. Simulation results in MATLAB show that the proposed method, in addition to having low computational complexity and high performance rate, have an acceptable result in identifying systematic block codes with long lengths and even at high error levels.

Sufficient number of linear and noisy measurements for exact and approximate sparsity pattern/support set recovery in the high dimensional setting is derived. Although this problem as been addressed in the recent literature, there is still considerable gaps between those results and the exact limits of the perfect support set recovery. To reduce this gap, in this paper, the sufficient condition is enhanced. A specific form of a Joint Typicality decoder is used for the support recovery task. Two performance metrics are considered for the recovery validation; one, which considers exact support recovery, and the other which seeks partial support recovery. First, an upper bound is obtained on the error probability of the sparsity pattern recovery. Next, using the mentioned upper bound, sufficient number of measurements for reliable support recovery is derived. It is shown that the sufficient condition for reliable support recovery depends on three key parameters of the problem; the noise variance, the minimum nonzero entry of the unknown sparse vector and the sparsity level. Simulations are performed for different sparsity rate, different noise variances, and different distortion levels. The results show that for all the mentioned cases the proposed methodology increases convergence rate of upper bound of the error probability of support recovery significantly which leads to a lower error probability bound compared with previously proposed bounds.

Visual cryptography is one of the newest techniques is image encryption. Visual cryptography encrypts visual information and generates two or more shares which contain no information separately, but reveal the secret when superposed. The main advantage of this scheme is that the decoding process does not need any knowledge of cryptography and human visual system is able to decrypt the secret message. In this article, a new encryption method based on particle swarm optimization is provided. The basis matrices are obtained using this approach and then used for encryption. In previous work, a Genetic Algorithm based method was proposed. In our work we fix a subtle yet crucial bug in the GA based method which is generating pictures with negative contrast in some cases and propose a simpler alternative. Simulation results show that the proposed method meets all the initial conditions of visual cryptography while decreasing the number of function evaluations and can provide a contrast enhancement. Moreover, the proposed method has the advantage of being generic and can be used in various threshold based visual cryptography.

In this paper, a sparse representation based manifold learning method is proposed. The construction of the graph manifold in high dimensional space is the most important step of the manifold learning methods that is divided into local and gobal groups. The proposed graph manifold extracts local and global features, simultanstly. After construction the sparse representation based graph manifold, two linear and nonlinear methods are proposed to extracte the embedded data. The proposed method is compared with the common manifold learning methods, LLE, LEM, LPP and PCA. The results on two Persian handwritten databases, HODA and IFHCDB,show the better performance of the proposed method and the recognition rates of 91.89 and 93.89 are achieved on HODA and IFHCDB, respectively. Also, a modification of the proposed method is proposed to reduce the computational complexity. The results on HODA demonstrate the good performance of the modified method and decrease the computational complexity around 6 times.

This paper investigates the wireless visual sensor networks energy minimization in the presence of occlusion which prevents the target tracking. An algorithm is proposed for selecting the suitable visual sensor and adjusting its parameters (the focal length and the frame rate) in a way to satisfy the user quality requirements. For modeling the received image quality, a mathematical model has been proposed for the variance value of the received images that it uses for the considered problem formulization. The used concept for the mathematical problem solving is convex optimization which leads to priority function extraction for visual sensor selection algorithm and its parameters adjustment. Combining the priorities with the repetitive algorithm provides the solution. Simulation results reveal that the proposed algorithm represents an acceptable efficiency in compare with the complex and optimum algorithm (exhaustive search) whiles its complexity is suitable for the practical applications.

In this paper, attempting to improve the recognition rate and solve some problems such as pose, lighting variations and partial occlusion in video sequences using Orthogonal Locality Preserving Projection (OLPP). In this research, first of all face in video frames is detected for background removing. Then each set of images is distributed on a nonlinear manifold and clustered using appropriate methods then the center of each cluster is considered as the cluster representative. It is also shown that by using OLPP the key frames are projected to the new space, where the frames in each manifold are better closed also the frames in different manifold are better separated. The recognition is done by projecting the test video sequence to the new space and calculating the distance between manifolds. Comparing the experimental results of the proposed method with other methods demonstrate the effectiveness of the proposed approach.

This paper presents a new approach for modeling the normal behaviors and detecting the abnormal behaviors. The approach consists of several main steps. First, using a detection method, the foreground and background regions are separated. Then, the busy-idle rates are defined as the behavioral features and, based on these features, a behavioral model is extracted for each pixel block. In the following, spectral clustering is used to classify the normal behaviors on the condition that a set of normal data is provided. In the classification process, the pixel blocks with similar behaviors are grouped together. A behavioral model is defined for each group of the blocks with similar behaviors. The behavioral model adopted in this paper is Hidden Markov Model. The results of the obtained classification and normal behaviors are used to detect the abnormal behaviors; i.e., based on the normal-behavior model for each cluster, if the observation sequence probability given by the normalbehavior model is lower than the threshold level, the pixel block is identified as the region in which the abnormal behaviors happened. The experimental results obtained from video data confirm the efficiency, accuracy, and speed of the approach adopted in this paper.

In this paper, a novel image fusion algorithm based on phase congruency is presented. The proposed algorithm has two steps. In the first step, the information content of the input images is evaluated by using phase congruency model. In this step, the main idea of the work is based on the fact that image features like edges are located in the points of image which frequency components have the maximum phase congruency. In the second step, the dominant features are selected and accordingly image fusion is done. Also, for reducing the computational complexity and improving the overall performance, the proper blocks are chosen by determining the appropriate threshold and the other blocks are chosen using K-means clustering algorithm through an iterative process. The experimental results on several images show that the proposed method has significant advantages compared to the other existing methods.

Tracking has considered many years ago and it has been concerned as an unsolved problem. This paper presents a methods using Genetic algorithm. The proposed algorithm uses background subtraction and multi scale LBP to detect moving target region. We have used Chi-square Criterion to calculate the similarity between two regions. The simulation results shows that the proposed method is efficient in accuracy and speed. In each frame the average time for calculating the position of the vehicle is 0.33 seconds while the sequential search time is 02.15 seconds.

Video-based face recognition has received significant attention in recent years. There are some problems in video-based face recognition method such as pose, lighting and expression variations, occlusion, aging and low image resolutions. In this paper, first of all face in video frames is detected for background removing. Then each set of images is distributed on a nonlinear manifold and clustered using appropriate methods. Determination of the number of clusters and linear models is quite intelligent and with implementation of this method, the number of clusters depends on different movements of head. Two methods are proposed for clustering and linear models. The distance between nonlinear manifolds that is composed of the combination of two methods, are used for recognition. The first method uses the distance between the clusters which are considered as the linear space and the second method uses the distance between key frames for recognition. Finally the results are compared with other methods.

Due the importance of the machine vision systems to object tracking, this is essential to use a system that is stable in different conditions of environment. In this paper, a new method by combining tracking algorithms is used to detect and track objects in different scales, and the video frames in real time. Hybrid algorithm mentioned above, contains two algorithms, scale invariant feature transform [1] and the fast symmetry detection. Scale invariant feature transform is a stable way to change the scale, which is able to detect the object during the tracking. Also, the fast symmetry detection algorithm is capable of real-time tracking of the object. The combination of two features covers the shortage of single feature and two models are used in a synergetic manner. The proposed method was evaluated in different environmental conditions, and the results show the speed increase and stability in the difficult environment.

In this study an algorithm, based on artificial neural networks (ANNs), for classifying four different varieties of Iranian pistachio nuts, Kaleghouchi (Ka), Akbari (Ak), Badami (Ba) and Ahmadagaee (Ah) is presented. To develop the ANN models a total of 3200 pistachio sound signals, 800 samples for each variety, were recorded. Features of pistachio nut varieties were extracted from analysis of sound signal in both time and frequency domains by means of fast fourier transform (FFT), power spectral density (PSD) and principal component analysis (PCA) methods. Altogether forty features were selected as input vector to ANN models. Network output vector consisted of four neurons for classification of varieties. In developing the ANN models, several ANN architectures, each having different numbers of neurons in hidden layer, were evaluated. The optimal model was selected after several evaluations based on minimizing the mean square error (MSE), correct separation rate (CSR), and correlation coefficient (r). Selected ANN for classification was of 40-12-4 configuration. CSR of the proposed ANN model for four pistachio varieties, Ka, Ak, Ba, and Ah were 96.97%, 97.64%, 96.36%, and 99.10%, respectively. Net weight average of system accuracy was found to be 97.51 %.