In this paper, a novel multi-head multi-layer perceptron (MLP) structure is presented for implicit neural representation (INR). Since conventional rectified linear unit (ReLU) networks are shown to exhibit spectral bias towards learning low-frequency features of the signal, we aim at mitigating this defect by taking advantage of local structure of the signals. To be more specific, an MLP is used to capture the global features of the underlying generator function of the desired signal. Then, several heads are utilized to reconstruct disjoint local features of the signal, and to reduce the computational complexity, sparse layers are deployed for attaching heads to the body. Through various experiments, we show that the proposed model does not suffer from the special bias of conventional ReLU networks and has superior generalization capabilities. Finally, simulation results confirm that the proposed multi-head structure outperforms existing INR methods with considerably less computational cost. The source code is available at https://github.com/AlirezaMorsali/MH-RELU-INR

Detecting emotions directly from a speech signal plays an important role in effective human-computer interactions. Existing speech emotion recognition models require massive computational and storage resources, making them hard to implement concurrently with other machine-interactive tasks in embedded systems. In this paper, we propose an efficient and lightweight fully convolutional neural network for speech emotion recognition in systems with limited hard-ware resources. In the proposed FCNN model, various feature maps are extracted via three parallel paths with different filter sizes. This helps deep convolution blocks to extract high-level features, while ensuring sufficient separability. The extracted features are used to classify the emotion of the input speech segment. While our model has a smaller size than that of the state-of-the-art models, it achieves a higher performance on the IEMOCAP and EMO-DB datasets. The source code is available https://github.com/AryaAftab/LIGHT-SERNET

Transformer networks, originally developed in the community of machine translation to eliminate sequential nature of recurrent neural networks, have shown impressive results in other natural language processing and machine vision tasks. Self-attention is the core module behind visual transformers which globally mixes the image information. This module drastically reduces the intrinsic inductive bias imposed by CNNs, such as locality, while encountering insufficient robustness against some adversarial attacks. In this paper we introduce K-sparse attention to preserve low inductive bias, while robustifying transformers against adversarial attacks. We show that standard transformers attend values with dense set of weights, while the sparse attention, automatically selected by an optimization algorithm, can preserve generalization performance of the transformer and, at the same time, improve its robustness.

Compressed Sensing is targeted at reconstructing a signal from a small set of measurements, if the signal is sparse in some domains. In this respect, a low coherent measurement matrix plays an important role. In this letter, an efficient approximation of ℓ∞-norm based on the soft maximum is introduced to design a low coherent measurement matrix with a controllable spectral norm. The proposed approximation, called Logarithm of Sum of Exponential Absolute values (LSEAp), is convex (similar to ℓ∞-norm) and almost smooth. We design a low coherent measurement matrix with a small spectral norm via minimization of the ℓ∞-norm of the Gram matrix. The resulting problem is not convex but our simulations show that the LSEAp leads to an improved design of the measurement matrix, as compared to current methods.

While pitch detection has been the focal subject of numerous research efforts for several decades, it is still a challenging task in noisy conditions. In this article, we propose a method to improve the pitch detection accuracy of conventional pitch detection methods. The proposed pitch detection process starts with using the pitch value estimated by a conventional pitch detection method. Then, it extracts pitch candidates according to the most probable types of errors in the initial estimation of high-pitch and low-pitch frames classified by a Deep Convolutional Neural Network (DCNN). Next, a restrained selection procedure is run to find the true pitch value from the set of pitch candidates. In this procedure, we employ two features (harmonic summation and Euclidian deviation), the soft decision of the DCNN, the pitch smoothness feature in successive frames, and the effect of the initial estimation in a cost function. The pitch value which leads to the lowest cost value is chosen as the estimated pitch value. The simulations on CSTR and KEELE databases, in noisy environments with twelve types of noise, were performed. The results show the superiority of the proposed method over the state-of-the-art methods under different SNR conditions.

Deep neural networks have presented superlative performance in many machine learning based perception and recognition tasks, where they have even outperformed human precision in some applications. However, it has been found that human perception system is much more robust to adversarial perturbation, as compared to these artificial networks. It has been shown that a deep architecture with a lower Lipschitz constant can generalize better and tolerate higher level of adversarial perturbation. Smooth regularization has been proposed to control the Lipschitz constant of a deep architecture and in this work, we show how a deep convolutional neural network (CNN), based on non-smooth regularization of convolution and fully connected layers, can present enhanced generalization and robustness to adversarial perturbation, simultaneously. We propose two non-smooth regularizers that present specific features for adversarial samples with different levels of signal-to-noise ratios. The regularizers build direct interconnections for the weight matrices in each layer, through which they control the Lipschitz constant of architecture and improve the consistency of input-output mapping of the network. This leads to more reliable and interpretable network mapping and reduces abrupt changes in the networks output. We develop an efficient algorithm to solve the non-smooth learning problems, which presents a gradual complexity addition property. Our simulation results over three benchmark datasets signify the superiority of the proposed formulations over previously reported methods for improving the robustness of deep architecture, towards human robustness to adversarial samples.

Classification of video events based on frame-level descriptors is a common approach to video recognition. In the meanwhile, proper encoding of the frame-level descriptors is vital to the whole event classification procedure. While there are some pretty efficient video descriptor encoding methods, temporal ordering of the descriptors is often ignored in these encoding algorithms. In this paper, we show that by taking into account the temporal inter-frame dependencies and tracking the chronological order of video sub-events, accuracy of event recognition is further improved. First, the frame-level descriptors are extracted using convolutional neural networks (CNNs) pre-trained on ImageNet, which are fine-tuned on a portion of training video frames. Then, a spatio-temporal encoding is applied to the derived descriptors. The proposed spatio-temporal encoding, as the main contribution of this work, is inspired from the well-known vector of locally aggregated descriptors (VLAD) encoding in spatial domain and from total variation de-noising (TVD) in temporal domain. The proposed unified spatio-temporal encoding is then shown to be in the form of a convex optimization problem which is solved efficiently with alternating direction method of multipliers (ADMM) algorithm. The experimental results show superiority of the proposed encoding method in terms of recognition accuracy over both frame-level video encoding approaches and spatio-temporal video representations. As compared to the state-of-the-art approaches, our encoding method improves the mean average precision (mAP) over both Columbia consumer video (CCV), unstructured social activity attribute (USAA), YouTube-8M, and Kinetics datasets and is very competitive on FCVID dataset.

Median filtering has received extensive attention from forensics analyzers, as a common content-preserving, smoothing, and denoising manipulation. We propose a detection scheme for median filtering of video sequences in compressed domain based on the singular value decomposition of the process matrix, which approximates the median filtering. Projection over some of the eigenspaces of the process matrix gives a set of features of small dimension, even as small as three, making the proposed scheme a fast and suitable detector for video median filtering. The experimental evaluations show that the proposed method outperforms the state-of-the-art detectors of median filtering, and its edge becomes significant for localized detection over small blocks within frames. The new method can potentially be used as a detector of other editing processes.

This paper is focused on video event recognition based on frame level convolutional neural network (CNN) descriptors. Using transfer learning, the image trained descriptors are applied to the video domain to make event recognition feasible in scenarios with limited computational resources. After fine-tuning of the existing CNN concept score extractors, pretrained on ImageNet, the output descriptors of the different fully connected layers are employed as frame descriptors. The resulting descriptors are hierarchically postprocessed and combined with novel and efficient pooling and normalization methods. As major contributions of this paper to the video event recognition, we present a postprocessing scheme in which the hierarchy and the relative shortest distance of concepts in WordNet concept tree is taken into account to alleviate uncertainty of the resulting concept scores at the output of the CNN. Besides, we propose a concept-wise power law normalization method that outperforms the widely used power law normalization. The integration of these approaches results in a high performance average (max) pooling-based video event recognition. Compared to the average (max) pooling combined with the state-of-the-art normalization methods and fine-tuned support vector machine classification, the proposed processing scheme improves the event recognition accuracy in terms of mean average precision over the Columbia consumer video and unstructured social activity attribute datasets, where achieves a pretty comparable result on UCF101 and ActivityNet datasets.

The authors address the use of sparse representation to securely hide a message within non‐overlapping blocks of a given colour image in the wavelet domain. All four sub‐images of the two‐dimensional wavelet transform of two colour bands are used for data embedding without affecting the image perceptibility. Bit error rate of hidden data extraction is reduced to zero by introducing a novel refinement procedure in the proposed algorithm. The refinement procedure introduced solves the hidden bit extraction errors caused by the rounding process, the overflows and the nature of approximation in sparse decomposition. Capacity of the proposed method is calculated using necessary conditions for uniqueness of the sparse representations and validated by means of experimental results. The authors experimental results show that the embedded data are invisible perceptually. The average peak signal‐to‐noise ratio (PSNR) value of the introduced method is about 40 and 11 dB higher than the average PSNR values of the MPSteg‐colour method and the wavelet domain 2‐LSB embedding method, respectively. The security of the proposed method is also evaluated comparatively, using five well‐known steganalysers, which confirms superior performance of the new method comparing with the MPSteg‐colour method and the wavelet domain 2‐LSB embedding method.

Steganalysis of least significant bit (LSB) embedded images in spatial domain has been investigated extensively over the past decade and most well‐known LSB steganography methods have been shown to be detectable. However, according to the latest findings in the area, two major issues of very low‐rate (VLR) embedding and content‐adaptive steganography have remained hard to resolve. The problem of VLR embedding is indeed a generic problem to any steganalyser, while the issue of adaptive embedding specifically depends on the hiding algorithm employed. The latter challenge has recently been brought up again to the area of LSB steganalysis by highly undetectable stego image steganography that offers a content‐adaptive embedding scheme for grey‐scale images. The authors new image steganalysis method suggests analysis of the relative norm of the image Clouds manipulated in an LSB embedding system. The method is a self‐dependent image analysis and is capable of operating on low‐resolution images. The proposed algorithm is applied to the image in spatial domain through image Clouding, relative auto‐decorrelation features extraction and quadratic rate estimation, as the main steps of the proposed analysis procedure. The authors then introduce and use new statistical features, Clouds‐Min‐Sum and Local‐Entropies‐Sum, which improve both the detection accuracy and the embedding rate estimation. They analytically verify the functionality of the scheme. Their simulation results show that the proposed approach outperforms some well known, powerful LSB steganalysis schemes, in terms of true and false detection rates and mean squared error.

In this paper, we present a novel speech steganography method using discrete wavelet transform and sparse decomposition to address the undetectability concern in speech steganography. The proposed speech steganography method exploits the sparse representation to embed secret messages into higher semantic levels of the cover signal, resulting in increased undetectability. The proposed method also yields improvements on both stego signal quality and embedding capacity, which are the two major requirements of a steganography algorithm. Our experimental results illustrate that the stego signals generated by the proposed method are perceptually indistinguishable from the original cover signals, quantified by both SNR and PESQ quality measures. When compared with two well-known steganography methods, the proposed method is shown to be superior on addressing major requirements of a steganography algorithm, imperceptibility, undetectability, and capacity.

This letter proposes a novel normalized subband adaptive filter (NSAF) algorithm, which applies variable step sizes to subband filters to improve the convergence performance of the conventional NSAF and update only a subset of the subbands per iteration to reduce its computational complexity. The selection process for each subband is based on the amount of improvement it makes to the mean square deviation at every iteration. Simulation results show significant reduction in computational complexity, faster convergence rate, and lower misadjustment error achieved using the proposed scheme.

Multi-dimensional spatial analysis of image pixels have not been much investigated for the steganalysis of the LSB Steganographic methods. Pixel distribution based steganalysis methods could be thwarted by intelligently compensating statistical characteristics of image pixels, as reported in several papers. Simple LSB replacement methods have been improved by introducing smarter LSB embedding approaches, e.g. LSB matching and LSB+ methods, but they are basically the same in the sense of the LSB alteration. A new analytical method to detect LSB stego images is proposed in this paper. Our approach is based on the relative locations of image pixels that are essentially changed in an LSB embedding system. Furthermore, we introduce some new statistical features including “local entropies sum” and “clouds min sum” to achieve a higher performance. Simulation results show that our proposed approach outperforms some well-known LSB steganalysis methods, in terms of detection accuracy and the embedding rate estimation.

This paper addresses a new approach to the keyframe extraction problem employing generalized Gaussian density (GGD) parameters of wavelet transform subbands along with Kullback-Leibler distance (KLD) measurement. Shot and cluster boundaries are selected using KLDs between GGD feature vectors, and then keyframes are located based on similarity and dissimilarity criteria. Objective and subjective evaluations show the high accuracy of this new approach compared with traditional methods.

A new adaptive steganography method based on image contrast to improve the embedding capacity and imperceptibility of the stego images is presented. The method exploits the average difference between the gray level values of the pixels in 2×2 blocks of non-overlapping spatially and their mean gray level in order to select valid blocks for embedding. The method was tested on different gray scale images. Results show that our proposed approach provides larger embedding capacity, while being less detectable by steganalysis methods such as ¿2 attack and machine learning steganalysis systems, as compared to some well-known adaptive and non-adaptive steganography algorithms.

In this study, a new feature-based steganalytic method is presented and four classification methods: Fisher linear discriminant, Gaussian naive Bayes, multilayer perceptron, and k nearest neighbor, are compared for steganalysis of suspicious images. The method exploits statistics of the histogram, wavelet statistics, amplitudes of local extrema from the ID and 2D adjacency histograms, center of mass of the histogram characteristic function and co-occurrence matrices for feature extraction process. In order to reduce the proposed features dimension and select the best subset, genetic algorithm is used and the results are compared through principle component analysis and linear discriminant analysis. The results show that the proposed method achieves higher accuracy in discriminating between innocent and stego images, as compared to one of wellknown image steganalysis schemes.

A new high capacity method for transform domain image steganography is introduced in this paper. The proposed steganography algorithm works on the wavelet transform coefficients of the original image to embed the secret data. As compared to current transform domain data hiding methods, this scheme can provide a larger capacity for data hiding without sacrificing the cover image quality. This is achieved through retaining integrity of the wavelet coefficients at high capacity embedding. This improvement to capacity-quality trading-off interrelation is analyzed in detailed and experimentally illustrated in the paper.

This paper presents a new scheme for steganalysis of random LSB embedding, capable of applying to any kind of digital signal in both spatial and transform domains. The proposed scheme is based on defining a space whose elements relate to higher-order statistical properties of the signal and looking for special subsets, which we call closure of sets (CoS) in this space. We use this scheme for steganalysis of the LSB steganography in grayscale images, employing a vector of five accurate and monotone features. Experimental results show significantly higher accuracy of the proposed scheme, as compared to those reported in the literature, especially in low embedding rates applications.

We propose a method for steganalysis of still, grayscale images using a novel set of features that are extracted from images. This feature set employs the Gabor filter coefficients to train a multi-layer perceptron neural network and a support vector machine classifier. We show that incorporation of the Gabor filter coefficients to the feature sets of images could have a significant role in discrimination between clean and altered images. Experimental results show that the proposed method outperforms previous methods, introduced for steganalysis of LSB-matching image steganography, in terms of both discrimination accuracy and feature set dimensionality.

This paper introduces an efficient scheme to image steganography by introducing the hidden message (payload) insertion in spatial domain and transforming the stego-image to compressed domain. We apply a recently-proposed LSB method in order to obtain better statistical behavior of the stego-message and subsequently, the obtained stego-image is transformed and quantized in order to enhance the security of hiding. Performance analysis comparisons confirm a higher efficiency for our proposed method. Compared to recently-proposed approaches, our method offers the advantage that it combines an efficient LSB method with transform domain security.

In this paper, a novel quantization based watermarking method is proposed. For blind detection, a set of nonlinear convex functions based on geometric mean are investigated. In order to achieve minimum distortion, the optimum function set is found. The algorithm is implemented on the approximation coefficients of wavelet transform for natural images. In order to make the algorithm more robust and imperceptible, a new transform domain called Point to Point Graph (PPG), which converts a 1-D signal to a 2-D one, has been used. The error probability of the proposed scheme is analytically investigated. Simulation results show that this algorithm has great robustness against common attacks such as AWGN, JPEG and rotation in comparison with recent methods presented so far.

In this paper, we propose a method for steganalysis of grayscale images using both spatial and Gabor features. The basis of our work is to use Gabor filter coefficients and statistics of the graylevel co-occurrence matrix of images to train a support vector machine. We show that this feature set works well in steganalysis of grayscale images steganographied by LSB matching and S-tools.

This paper introduces a new method for the least significant bit (LSB) steganography in spatial domain. The new method, called one-third LSB embedding, reduces the probability of change per pixel to one-third without sacrificing the embedding capacity, as compared to previous LSB steganography techniques. This improvement results in a better imperceptibility and higher robustness against well known LSB detectors. Each bit of the message is carried through a function of three adjacent cover pixels and it is shown that no significant improvement is achieved by increasing the length of the pixel sequence employed.

This paper introduces a novel scalable speech coding scheme based on embedded matrix quantization of LSF parameters in an LPC model. In the proposed quantizer, codewords are organized based on a tree structure through a cell-merging process, which leads to a fine-grain scalable coder at rates below 900 bps. Near natural sounding is achieved at very low rates by employing an efficient adaptive dual-band scheme to approximate the LPC excitation signals. Evaluation results, obtained from both overall quality measurement and intelligibility assessment, show that the proposed coder could be a reasonable choice for improving the bottom-line speech quality in low bit rates.

In this paper, a set of features is presented and evaluated based on sinusoidal modeling of audio signals. Amplitude, frequency, and phase parameters of the sinusoidal model are used and compared as input features into an audio classifier system. The performance of sinusoidal model features is evaluated for classification of audio into speech and music classes using both the Gaussian and the GMM (Gaussian mixture model) classifiers. Experimental results show superiority of the amplitude parameters of the sinusoidal model, which could be used for the first time for such an audio classification, as compared to the popular cepstral features. By using a set of 40 sinusoidal features, we achieved 95.06% accuracy in the audio classification at frame level, as compared to 92.26% accuracy obtained with the MFCC coefficients, as tested over the same audio corpus.

This paper addresses a new approach to segmentation and classification of audio through analysis of a smaller set of selective frames, which are identified by temporal decomposition (TD). These frames are located at the most steady instants, or event centroids, within a given block of the signal, which yield the maximal diversity over the set of selected features. Based on this selection scheme, the number of frames used in the analysis is reduced by at least 40%, while the temporal resolution is doubled as compared to that in typical audio classifiers. By constructing a classification system to segment audio into speech, music, speech-music, and others, it is shown that the proposed method outperforms the typical classifiers in most cases. In addition, by using hierarchical TD for frame selection, it is made possible to adapt the audio classifier with other segmentation schemes, e.g., visual classification based on motion picture analysis, for accurate audio-visual segmentation of multimedia data.

A new method for two-band approximation of excitation signals in an LPC model, to improve speech naturalness in very low rate coding, is proposed. Based on a simplified model of multi-band excitation, the method accurately determines the degree of periodicity, using the concept of instantaneous frequency (IF) estimation in the frequency domain. The harmonic structure in the spectrum of LPC residual, within individual bands, is identified based on flatness of the IF as a criterion for pitch and voicing detection. On this basis, the excitation is modelled by combining a predefined periodic signal in the lower band and a random signal in the higher band. It is shown that this improves considerably the naturalness of reconstructed speech in very low rate coding in comparison with that obtained using traditional binary excitation. The performance of the technique is also given in temporal decomposition (TD) based coding at 800 b/s.

Temporal decomposition (TD) is an effective technique to compress the spectral information of speech through orthogonalization of the matrix of spectral parameters leading to an efficient rate reduction in speech coding applications. The performance of TD is a function of the parameters used. Although "decomposition suitability" of a parameter set is typically defined on the basis of "phonetic relevance" criterion, it can not be directly used in speech coding. Instead, quality evaluation of reconstructed speech is more appropriate. In this paper, we extend our earlier work in this area and attempt to assess several "popular" spectral parameter sets from the viewpoint of decomposition suitability in very low bit-rate speech coding using parametric, perceptually-based spectral, and energy distance measures.

Temporal decomposition (TD) is a method to reduce the correlation between speech spectral parameter sets using orthogonalization. The resulting parameters (so-called weightings or target vectors) correspond, mostly, to least-dependent phonetic states, which produce the desired sounds. Other phonetic states can be interpreted as transition states which force the event functions to overlap. On the basis of this assumption, we propose to approximate the event functions by sample functions to focus on the phonetically important states, and to use the resulting parameters for constructing a very low-rate speech coder. The objective and subjective evaluation of synthesized speech, mostly from the intelligibility view point, confirm the assumptions made on the events.