Most real-time speech signals are frequently disrupted by noise such as traffic, babbling, and background noises, among other things. The goal of speech denoising is to extract the clean speech signal from as many distorted components as possible. For speech denoising, many researchers worked on Sparse representation and dictionary learning algorithms. These algorithms, however, have many disadvantages, including being overcomplete, computationally expensive, and susceptible to orthogonality restrictions, as well as a lack of arithmetic precision due to the usage of double-precision. We propose a greedy technique for dictionary learning with Sparse representation to overcome these concerns. In this technique, the input signal's singular value decomposition is used to exploit orthogonality, and here the ℓ1-ℓ2 norm is employed to obtain sparsity to learn the dictionary. It improves dictionary learning by overcoming the orthogonality constraint, the three-sigma rule-based number of iterations, and the overcomplete nature. And this technique has resulted in improved performance as well as reduced computing complexity. With a bit-precision of Q7 fixed-point arithmetic, this approach is also used in resource-constrained embedded systems, and the performance is considerably better than other algorithms. The greedy approach outperforms the other two in terms of SNR, Short-Time Objective Intelligibility, and computing time.

With the widespread using Internet in any device and services, several homes and workplace applications have been provided to avoid attacks. Connecting a system or device to an insecure network can create the possibility of being infected by unwanted files. Detecting such files is a vital task in any system. Employing machine learning (ML) is the most efficient method to detect these penetrations. On the other hand, malware programmers try to design malicious files that are hard to detect. A file can hide from detection in a feature view, but concealing in all views would be very difficult. In this paper, inspiring Multi-View learning (MVL), we proposed to incorporate some various features such as Opcodes, Bytecodes, and System-calls to achieve complementary information to identify a file. In this way, we developed a modified version of Sparse Representation based Classifier (SRC) to aggregate the effect of all modalities in a unified classifier. To show the efficiency of the proposed method, we used several real datasets. Experimental results show the high performance of the proposed approach and its ability to cope with the imbalanced conditions.

Joint Photographic Experts Group (JPEG) is one of the most widely used image compression methods, but it causes annoying blocking artifacts at low bit-rates. Sparse representation is an efficient technique which can solve many inverse problems in image processing applications such as denoising and deblocking. In this paper, a post-processing method is proposed for reducing JPEG blocking effects via Sparse representation. In this method, a dictionary is learned via the single input blocky image using KSVD. There is no need for any prior knowledge about the blocking artifacts. Experimental results on various images demonstrate that the proposed post-processing method can efficiently alleviate the blocking effects at low bit-rates and outperform some new well-known image deblocking methods.

Voiced-based age detection and gender recognition are important problems in the telephone speech processing to investigate the identity of an individual. In this paper, a new gender and age recognition system is introduced based on the generative incoherent models learned using Sparse non-negative matrix factorization and the atom correction step as a post-processing method. The proposed classification algorithm includes training step to provide the appropriate trained atoms for each data class and also the test phase to assess the classification performance. Since the classification accuracy depends highly on the selected features, the Mel-frequency cepstral coefficients are employed to train basis for the better representation of the voice structure. These bases are learned over the data of male and female speakers using non-negative matrix factorization with the sparsity constraint. Then, atom correction is carried out using an energy-based algorithm to decrease the coherence between different categories of the trained dictionaries. In the Sparse representation of each data class, the atoms related to other sets with the highest energy are replaced with the lowest energy bases if the reconstruction error does not exceed from a specified limit. The experimental results showed that the proposed algorithm performs better than the earlier methods in this context especially in the presence of background noise.

In this paper, Sparse double selective channel estimation using compressed sensing (CS) theory for OFDM systems is investigated. This theory helps to reduce the required pilot ratio and equivalently increases the spectral efficiency to achieve a constant mean square error. This is of great importance especially for double selective channels in which the required number of unknowns to be estimated and also the required number of pilot symbols are high. To take the advantage of compressed sensing, it is proposed that the sparsity enhancement of the coefficients of basis expansion model (BEM) should be considered in BEM design. It is also proposed to use K-SVD algorithm that is one of the most popular dictionary learning algorithms. Moreover, in this paper clustered pilot symbols are used to avoid inter-carrier interference. It is noteworthy that the channel coefficients representing intercarrier interference are also estimated to be used in equalization. Numerical experiments have shown that the compressed sensing estimator employing the proposed basis, outperforms the one employing DFT-DPSS in terms of NMSE and system BER.

Summary In this paper, the task is to return from a set of multiplicities from a model to obtain an approximation of that model using Sparse approximation. The term ‘ approximation’ indicate the sufficiency of an interpretation that is close enough to the true mode, i. e. reality. In geosciences, the multiplicities are provided by multiple-point statistical (MPS) methods. Realistic modeling of the earth interior demands for more sophisticated geostatistical methods based on true available images, i. e. the training images. Among the available MPS methods, the DisPat algorithm is a distance-based MPS method, which generates appealing realizations for stationary and non-stationary training images by classifying the patterns based on distance functions using kernel methods. Advances in non-stationary image modeling is an advantage of the DisPat method. Realizations generated by the MPS methods form the training set for the Sparse approximation. The Sparse approximation is comprising of two steps, Sparse coding and dictionary update, which are alternately used to optimize the trained dictionary. Model selection algorithms like LARS are used for Sparse coding. LARS optimizes the regression model sequentially by choosing a proper number of variables and adding the best variable to the active set in each iteration. The ILS-DLA dictionary learning algorithm addresses the internal structure of the dictionary by considering the overlapping or non-overlapping blocks and the inversion task according to the internal structure of the trained dictionary. The ILS-DLA is fast in the sense that it inverts smaller blocks constructing the trained dictionary rather than inverting the entire dictionary. The trained dictionary is sequentially updated by alternating between Sparse coding and dictionary training steps. According to the experiments, the compressed sparsity-based image model is superior to 90% of the generated realizations by 90% probability...