State estimation in time-varying systems requires adaptive or recursive methods, such as the Kalman filter (KF), to accurately track dynamic states in noisy environment. The KF employs the minimum mean square error (MMSE) criterion to optimize its performance in presence of Gaussian noise. However, when the state of a system is adversely affected by the non-Gaussian noise, the performance of the KF degrades significantly. To overcome this, the article proposes a hyperbolic tangent Geman-McClure (HTGM) cost function based robust variant of KF. The proposed HTGM based Kalman Filter (HTGMKF) effectively handles non-Gaussian noise, such as Gaussian-mixture and Laplacian noise in time-varying systems, and provides lower root mean square error (RMSE) and average RMSE. The proposed HTGMKF outperforms existing state-of-the-art filters such as maximum correntropy criterion-based KF, minimum error entropy-based KF, robust MEEKF, q-Rényi kernel function-based KF, and multi-kernel maximum correntropy-based KF under diverse noise conditions. This study also provides the lower and uniform upper bounds of the error covariance matrix (ECM) for HTGMKF. The simulation results show that the ECM is bounded and asymptotically stable.

Multimodal frameworks have appeared as a potential solution to achieve breakthrough results in the field of sign language and hand gesture recognition. In this paper, we propose a classifier combination based multimodal framework for continuous sign language recognition. In this work, we propose to use transfer learning to perform classification task on individual modalities. Further, we apply majority voting scheme to combine all the individual classification performances to obtain the final classification accuracy. For transfer learning, pretrained deep neural networks like GoogleNet, MobileNet-v2 and EfficientNet-b0 are used independently from one another. We applied these networks independently on individual modalities of IPN Hand dataset and combined the classification results obtained on these modalities to obtain the final classification result. IPN Hand dataset is one of the most challenging and dynamic continuous hand gesture datasets. Among the used pre-trained networks, EfficientNet-b0 appeared as the best network in terms of accuracy whereas GoogleNet took the least computational time during classification. On this dataset, our proposed approach performs exceptionally well with individual modalities. After combining the classification results of individual modalities according to our employed algorithm, it is observed that our proposed approach performs better than the earlier reported results. Our method surpasses the reported benchmark performance as well as performs superior to many state-of-the-art results.

In this paper, we present a multimodal sentence level sign language synthesis system. Sentence level sign language synthesis requires synthesis of signs as well as synthesis of transition segments between the signs. In this paper, our focus is to develop. an efficient transition segment between two signs. For this, we propose to use edge features and trajectory features obtained from the transition segment of the query sign sentences. Edge features are extracted using the morphological operations on the hand images, whereas trajectory features are obtained from centroid detection of consecutive hand images. From trajectory, we obtain the direction of motion and shape and co-ordinates of the trajectory. Further, with the help of interpolation techniques we synthesize the hand gestures between the two signs. The correctness of the synthesis is analyzed and modified with the help of edge features and trajectory features. Our proposed approach is tested on some Indian Sign Language phrases and sentences. The proposed method is evaluated based on the mean opinion scores obtained from several users. Evaluation was based on three criteria, namely clarity in understanding, smoothness of the generated videos and similarity to the original sign videos. We obtained decent mean scores and encouraging feedbacks from the users.

Transformer models have demonstrated remarkable success in many domains such as natural language processing (NLP) and computer vision. With the growing interest in transformer-based architectures, they are now utilized for gesture recognition. So, we also explore and devise a novel ConvMixFormer architecture for dynamic hand gestures. The transformers use quadratic scaling of the attention features with the sequential data, due to which these models are computationally complex and heavy. We have considered this drawback of the transformer and designed a resource-efficient model that replaces the self-attention in the transformer with the simple convolutional layer-based token mixer. The computational cost and the parameters used for the convolution-based mixer are comparatively less than the quadratic self-attention. Convolution-mixer helps the model capture the local spatial features that self-attention struggles to capture due to their sequential processing nature. Further, an efficient gate mechanism is employed instead of a conventional feed-forward network in the transformer to help the model control the flow of features within different stages of the proposed model. This design uses fewer learnable parameters which is nearly half the vanilla transformer that helps in fast and efficient training. The proposed method is evaluated on NVidia Dynamic Hand Gesture and Briareo datasets and our model has achieved state-of-the-art results on single and multimodal inputs. We have also shown the parameter efficiency of the proposed ConvMixFormer model compared to other methods. The source code is available at https://github.com/mallikagarg/ConvMixFormer.

Dynamic gesture recognition is one of the challenging research areas due to variations in pose, size, and shape of the signer's hand. In this letter, Multiscaled Multi-Head Attention Video Transformer Network (MsMHA-VTN) for dynamic hand gesture recognition is proposed. A pyramidal hierarchy of multiscale features is extracted using the transformer multiscaled head attention model. The proposed model employs different attention dimensions for each head of the transformer which enables it to provide attention at the multiscale level. Further, in addition to single modality, recognition performance using multiple modalities is examined. Extensive experiments demonstrate the superior performance of the proposed MsMHA-VTN with an overall accuracy of 88.22% and 99.10% on NVGesture and Briareo datasets, respectively.

In this paper, the problem of movement epenthesis in continuous sign language sentences is considered. Movement epenthesis caused due to unwanted but unavoidable hand movement in between two sign gestures in continuous signing has emerged as one of the most challenging problems in automatic sign language recognition. To handle this problem, a novel method based on unsupervised learning approach has been proposed in this paper to separate out video frames corresponding to meaningful sign gestures from meaningless movement epenthesis segments in a continuous signing gesture video clip. Our proposed method is based on K-means clustering of the norm values of the absolute difference between current frames and the reference frame to detect the movement epenthesis frame and then classify the frames of the video as movement epenthesis frames and sign frames. Exhaustive experimentation on the publicly available standard ChaLearn LAP ConGD gesture video dataset was carried out to test our algorithm. Experimental results demonstrate that the proposed method is good enough to detect and separate out the movement epenthesis frames in sentence level sign language videos. Our proposed approach performs the movement epenthesis detection task accurately in 91% of the videos of the ChaLearn LAP ConGD dataset.

Compressed sensing (CS) is extensively used to reduce magnetic resonance imaging (MRI) acquisition time. State-of-the-art deep learning-based methods have proven effective in obtaining fast, high-quality reconstruction of CS-MR images. However, they treat the inherently complex-valued MRI data as real-valued entities by extracting the magnitude content or concatenating the complex-valued data as two real-valued channels for processing. In both cases, the phase content is discarded. To address the fundamental problem of real-valued deep networks, i.e. their inability to process complex-valued data, we propose a complex-valued generative adversarial network (Co-VeGAN) framework, which is the first-of-its-kind generative model exploring the use of complex-valued weights and operations. Further, since real-valued activation functions do not generalize well to the complex-valued space, we propose a novel complex-valued activation function that is sensitive to the input phase and has a learnable profile. Extensive evaluation of the proposed approach1on different datasets demonstrates that it significantly outperforms the existing CS-MRI reconstruction techniques.

The Fourier transform of a N point time series is a N point complex series, while the S‐transform (ST) of the same time series is a 2D time–frequency complex matrix. The computation and storage of additional points are a major drag on the usage of ST. In this study the compact S‐transform (cST) is presented, with efficiencies brought about through computation of only selected voices (frequencies). The cST spectrum has uncomputed voice gaps that increase in width towards the higher frequencies. Plot of the cST magnitude spectrum is virtually indistinguishable from the ST magnitude plot. Local spectrum at any spot on the cST can be quickly examined in detail through interpolation. The cST requires the computation of approximately voices compared to for the ST. The proportion of computed voices decrease for larger N. For , ∼20% of the voices in the time‐frequency spectrum is computed; for only 14% of the voices is computed. For applications, such as audio and speech signal processing where segments of one million samples are not uncommon, <1% of the voices are computed, thereby reducing the computation time by ∼99%.

In-sensor energy-efficient deep learning accelerators have the potential to enable the use of deep neural networks in embedded vision applications. However, their negative impact on accuracy has been severely underestimated. The inference pipeline used in prior in-sensor deep learning accelerators bypasses the image signal processor (ISP), thereby disrupting the conventional vision pipeline and undermining accuracy of machine learning algorithms trained on conventional, post-ISP datasets. For example, the detection accuracy of an off-the-shelf Faster RCNN algorithm in a vehicle detection scenario reduces by 60%. To make in-sensor accelerators practical, we describe energy-efficient operations that yield most of the benefits of an ISP and reduce covariate shift between the training (ISP processed images) and target (RAW images) distributions. For the vehicle detection problem, our approach improves accuracy by 25-60%. Relative to the conventional ISP pipeline, energy consumption and response time improve by 30% and 34%, respectively.

Time-frequency (TF) analysis through well-known TF tool namely S-transform (ST) has been extensively used for QRS detection in Electrocardiogram (ECG) signals. However, Gaussian window-based conventional ST suffers from poor TF resolution due to the fixed scaling criterion and the long taper of the Gaussian window. Many variants of ST using different scaling criteria have been reported in literature for improving the accuracy in the detection of QRS complexes. This paper presents the usefulness of zero-order prolate spheroidal wave function (PSWF) as a window kernel in ST. PSWF has ability to concentrate maximum energy in narrow and finite time and frequency intervals, and provides more flexibility in changing window characteristics. Synchrosqueezing transform is a post processing method that improves the energy concentration in a TFR remarkably. This paper proposes a PSWF-based synchrosqueezing ST for detection of R peaks in ECG signals. The results show that the proposed method accurately detects R peaks with a sensitivity, positive predictivity and accuracy of 99.96 %, 99. 96% and 99. 92% respectively. It also improves upon on existing techniques in terms of the aforementioned metrics and the search back range.

For numerous applications in the field of signal processing, it is desired to design a compact window that can simultaneously concentrate maximum energy in finite time interval and frequency band. Although zero‐order discrete prolate spheroidal sequence (DPSS) meets this requirement, it is of infinite support. Limiting this sequence to finite support no longer guarantees the optimality property. This study aims at designing a discrete finite length window that can maximise the energy simultaneously in narrow time interval and frequency band. A multi‐objective optimisation approach is adopted to obtain the upper bound of maximum achievable time and frequency domain energy concentrations for finite length sequences. The optimal sequence thus obtained is termed as the optimal window with finite support (OWFS), and its various associated properties are discussed. It is shown analytically that as the support of OWFS approaches infinity, it converges to zero‐order DPSS. In order to illustrate its optimality, the compactness of the proposed OWFS is compared with those of various window functions. Extending the proposed OWFS, this study also discusses the formulation and associated properties of the zero‐order periodic DPSS. Further, the closed form expression for the upper bound of achievable time and frequency domain energy concentrations is also derived.

The S-transform (ST) is a commonly used tool for time-frequency (TF) localization. It uses scalable Gaussian kernel for multiresolution analysis. It is widely used for the analysis of seismic signals. However, the long taper of the Gaussian window results in the degradation of the time resolution at lower frequencies and, hence, lowers its effectiveness in reliable detection of onset of events. This letter proposes an asymmetrical Kaiser window-based ST that provides better time resolution in the forward direction which facilitates sharp detection of events in seismic signals. The inherent optimal energy concentration characteristics of the Kaiser window provide better TF localization compared to existing approaches. The window parameters are also generalized to provide frequency-dependent asymmetry while avoiding unwanted degradation of frequency resolution at higher frequencies. The efficacy of the proposed approach is illustrated by testing it with a synthetic seismic trace and a real noisy seismograph. Simulation results illustrate the improvement in time resolution with the proposed approach as compared to other asymmetrical window-based STs for the detection of event's onset.

S-transform (ST) is a well-known tool for time-frequency analysis. However, the long taper of the Gaussian window (GW) and the scaling criterion lead to poor time resolution at lower frequencies and poor frequency resolution at higher frequencies. This letter proposes a novel optimally concentrated discrete window (OCDW) based on a constraint optimization problem of maximization of the product of time- and frequency-domain energy concentrations in given time and frequency intervals. Furthermore, it is extended to design an OCDW-based ST for multiresolution analysis. A new scaling criterion is also presented for the OCDW-based ST, which prevents unnecessary deterioration in frequency resolution at higher frequencies and time resolution at lower frequencies. The efficacy of the proposed approach is examined using a synthetic signal having multiple power quality disturbances. Simulation results show that the proposed OCDW-based ST provides higher energy concentration in time-frequency distribution (TFD) than that of the GW-based ST. The simulation results also illustrate the superiority of the proposed scaling criterion in terms of the tradeoff between time and frequency resolution in TFD.

Estimation of time varying downlink channel for Long Term Evolution (LTE) systems in the high-mobility environment is a challenging task. In literature, the state-of-the-art techniques used for LTE channel estimation are extended Kalman filter (EKF), 2D interpolation using least square (2DILS), etc. Channel estimation has been performed by inserting pilot symbols in the frame. For fast time varying channels, Kalman filter based channel estimation is a fast and less complex technique compared to other conventional filters. The time correlation is modeled as a first order auto-regressive (AR) random process. Linearization of the state transition function is carried out using Taylor approximation. However, the higher bit error rate in fast fading channels discourages the use of aforementioned channel estimation techniques. This paper proposes a modified extended Kalman filter (MEKF) for joint estimation of the channel response and AR model coefficients. The bit error vs SNR performance of the proposed estimation technique has been demonstrated for different fast time varying LTE channels such as pedestrian user, vehicular user at different velocities such as 50 km/h and 70 km/h. The simulation results show that the proposed MEKF based approach leads to lower bit error rate than its two state-of-the-art counterparts based on EKF and 2DILS.

Due to ever increasing demand of services in the field of mobile communication, the fourth generation based systems has reached a saturation stage. The world today is looking towards fifth generation (5G), which would open doors for a variety of new services. Generalized Frequency Division Multiplexing (GFDM) is a potential candidate for forming the physical layer of the 5G due to its appealing properties such as low out of band (OOB) radiation and optimal use of time-frequency resources. Various filters such as root raised cosine (RRC), Xia pulse, etc. are proposed in the literature for pulse shaping of GFDM waveform. These filters affect the amount of OOB radiations and symbol error rate (SER) performance of the system in different ways. There is always a trade-off between the two performance indices, especially in low latency scenarios. Thus, it is important to design a filter impulse response, which can provide an optimal solution so as to get the best out of the GFDM system. This paper proposes a novel ramp-based finite impulse response filter for pulse shaping the GFDM waveform in the frame structure of tactile internet. The proposed filter makes an effort to address the limitation of the RRC and Xia filters. For the three filters, the power spectral density is calculated with respect to different sub-carriers. The SER v/s signal to noise ratio performance is tested in additive white Gaussian noise (AWGN) channel. Simulation results show that a system with proposed ramp-based filter offers significantly low OOB radiation as compared to RRC filter and Xia pulse. Furthermore, the SER performance of the proposed ramp-based filter is superior to Xia pulse in AWGN channel model.

The discrete two‐dimensional (2D) S‐transform (ST‐2D) is a space‐frequency representation that provides the local spectrum at each pixel in an image. The inherent localisation properties of the ST‐2D make it a preferred candidate for space‐frequency analysis. However, the ST‐2D is sensitive to the image rotation. This limits its use in many image processing applications such as texture analysis. This study describes the rotational sensitivity of the ST‐2D, and explains how this limits its application. To overcome these problems, a novel rotationally invariant S‐transform (RIST) is formulated, and its rotational invariance is investigated. The RIST can provide approximate rotational invariance for magnitudes and complex values over a wide range of rotation angles, but with a trade‐off of higher computation time. The RIST may be useful for many medical, scientific and industrial applications.

S ‐transform (ST) is a time–frequency representation (TFR) that is popularly used in a variety of applications, but prohibitive in both storage and computation for large time series. This study proposes a one‐dimensional fast time–frequency transform (FTFT‐1D) which generates a highly compressed form of the ST directly, without the need to compute the complete ST matrix initially and compress it subsequently. The compression technique used in the FTFT‐1D helps in storage, transmission and visualisation of the ST. The FTFT‐1D encodes the TFR information uniformly and finds the local spectrum instantaneously and accurately. It is useful for real‐time applications such as monitoring, control and filtering. In addition, the FTFT‐1D is memory efficient, robust and adaptive in nature.