Redundant Representation Research Articles

Co-Saliency Detection With Co-Attention Fully Convolutional Network

Co-saliency detection aims to detect common salient objects from a group of relevant images. Some attempts have been made with the Fully Convolutional Network (FCN) framework and achieve satisfactory detection results. However, due to stacking convolution layers and pooling operation, the boundary details tend to be lost. In addition, existing models often utilize the extracted features without discrimination, leading to redundancy in representation since actually not all features are helpful to the final prediction and some even bring distraction. In this paper, we propose a co-attention module embedded FCN framework, called as Co-Attention FCN (CA-FCN). Specifically, the co-attention module is plugged into the high-level convolution layers of FCN, which can assign larger attention weights on the common salient objects and smaller ones on the background and uncommon distractors to boost final detection performance. Extensive experiments on three popular co-saliency benchmark datasets demonstrate the superiority of the proposed CA-FCN, which outperforms state-of-the-arts in most cases. Besides, the effectiveness of our new co-attention module is also validated with ablation studies.

Intelligent Real-world Visual Auxiliary System

In this paper, the object recognition module of a visual auxiliary system, InVision, called IR-VP, is presented. (1) Deep multimodal neural network (DMNN) is presented to enhance CNNs’ sampling abilitiy and feature resolutions. (2) Deep distance learning (DDL) is presented to find similarity and reduce representation redundancy. (3) The light weight MobileNet is used to accelerate. Extensive experiments demonstrate that the proposed approach significantly outperforms state-of-the-arts.

VLSI Implementation of Digital Filter using Novel RTSD Adder and Booth Multiplier

Finite Impulse Response (FIR) filters are most important element in signal processing and communication. Area and speed optimization are the essential necessities of FIR filter design. This work looks at the design of Finite Impulse Response (FIR) filters from an arithmetic perspective. Since the fundamental arithmetic operations in the convolution equations are addition and multiplication, they are the objectives of the design analysis. For multiplication, Booth encoding is utilized in order to lessen the quantity of partial products. Consequently, considering carry-propagation free addition strategies should improve the addition operation of the filter. The redundant ternary signed-digit (RTSD) number framework is utilized to speedup addition in the filter. The redundant ternary representation utilizes more bits than required to denote the single binary digit because of which most numbers have several representations. This special behavior of RTSD allows the addition along with the absence of typical carry propagation. Xilinx ISE design suite 14.5 is used for the design and validation of proposed method. From the implementation result, the proposed design of FIR filter is compared with other conventional techniques to show the better performance by means of power, area and delay.

Tensorized Multi-view Subspace Representation Learning

Self-representation based subspace learning has shown its effectiveness in many applications. In this paper, we promote the traditional subspace representation learning by simultaneously taking advantages of multiple views and prior constraint. Accordingly, we establish a novel algorithm termed as Tensorized Multi-view Subspace Representation Learning. To exploit different views, the subspace representation matrices of different views are regarded as a low-rank tensor, which effectively models the high-order correlations of multi-view data. To incorporate prior information, a constraint matrix is devised to guide the subspace representation learning within a unified framework. The subspace representation tensor equipped with a low-rank constraint models elegantly the complementary information among different views, reduces redundancy of subspace representations, and then improves the accuracy of subsequent tasks. We formulate the model with a tensor nuclear norm minimization problem constrained with $$\ell _{2,1}$$ -norm and linear equalities. The minimization problem is efficiently solved by using an Augmented Lagrangian Alternating Direction Minimization method. Extensive experimental results on diverse multi-view datasets demonstrate the effectiveness of our algorithm.

Semi-Supervised Multi-View Deep Discriminant Representation Learning.

Learning an expressive representation from multi-view data is a key step in various real-world applications. In this paper, we propose a semi-supervised multi-view deep discriminant representation learning (SMDDRL) approach. Unlike existing joint or alignment multi-view representation learning methods that cannot simultaneously utilize the consensus and complementary properties of multi-view data to learn inter-view shared and intra-view specific representations, SMDDRL comprehensively exploits the consensus and complementary properties as well as learns both shared and specific representations by employing the shared and specific representation learning network. Unlike existing shared and specific multi-view representation learning methods that ignore the redundancy problem in representation learning, SMDDRL incorporates the orthogonality and adversarial similarity constraints to reduce the redundancy of learned representations. Moreover, to exploit the information contained in unlabeled data, we design a semi-supervised learning framework by combining deep metric learning and density clustering. Experimental results on three typical multi-view learning tasks, i.e., webpage classification, image classification, and document classification demonstrate the effectiveness of the proposed approach.

Penetrating the influence of regularizations on neural network based on information bottleneck theory

Regularization is a very effective algorithm to solve overfitting problem in neural network, which improves the generalization ability of the model. However, their working mechanisms and the impact on the model performance have not been fully explored. In this paper, we study and analyze them using information bottleneck theory and one theory from human brain sensory system. We propose a metric to characterise the encoding length of hidden layers, named as AEntry value. Then, we implement extensive experiments on MNIST and FashionMNIST datasets with several commonly used regularization algorithms, and calculate the corresponding AEntry values. We analyze these results and obtain three conclusions. (1) The introduction of regularization influences the encoding of relative features with prediction task in neural network. The early stopping technique avoids introducing unrelated information with the task into the model by stopping the training process as an appropriate iterations. Laplace, Gaussian and Sparse Response regularizations compress the related representation and improve the performance of neural network by introducing the prior information into the model. In contrast, Dropout, Batch Normalization, and Layer Normalization increase the encoding length of features by adopting redundant representation to improve the performance. (2) The encoding of neural network does not satisfy the data processing inequality of information theory, which is mainly caused by redundant coding of extracted features. (3) The overfitting is caused by introducing irrelative information with the target. These results can give us insight into building more efficient regularization algorithm to improve the performance of neural network model.

Neural Networks with a Redundant Representation: Detecting the Undetectable.

We consider a three-layer Sejnowski machine and show that features learnt via contrastive divergence have a dual representation as patterns in a dense associative memory of order P=4. The latter is known to be able to Hebbian store an amount of patterns scaling as N^{P-1}, where N denotes the number of constituting binary neurons interacting P wisely. We also prove that, by keeping the dense associative network far from the saturation regime (namely, allowing for a number of patterns scaling only linearly with N, while P>2) such a system is able to perform pattern recognition far below the standard signal-to-noise threshold. In particular, a network with P=4 is able to retrieve information whose intensity is O(1) even in the presence of a noise O(sqrt[N]) in the large N limit. This striking skill stems from a redundancy representation of patterns-which is afforded given the (relatively) low-load information storage-and it contributes to explain the impressive abilities in pattern recognition exhibited by new-generation neural networks. The whole theory is developed rigorously, at the replica symmetric level of approximation, and corroborated by signal-to-noise analysis and MonteCarlo simulations.

A View Transformation Model Based on Sparse and Redundant Representation for Human Gait Recognition.

Background:Human gait as an effective behavioral biometric identifier has received much attention in recent years. However, there are challenges which reduce its performance. In this work we aim at improving performance of gait systems under variations in view angles, which present one of the major challenges to gait algorithms.Methods:We propose employment of a view transformation model based on sparse and redundant (SR) representation. More specifically, our proposed method trains a set of corresponding dictionaries for each viewing angle, which are then used in identification of a probe. In particular, the view transformation is performed by first obtaining the SR representation of the input image using the appropriate dictionary, then multiplying this representation by the dictionary of destination angle to obtain a corresponding image in the intended angle.Results:Experiments performed using CASIA Gait Database, Dataset B, support the satisfactory performance of our method. It is observed that in most tests, the proposed method outperforms the other methods in comparison. This is especially the case for large changes in the view angle, as well as the average recognition rate.Conclusion:A comparison with state-of-the-art methods in the literature showcases the superior performance of the proposed method, especially in the case of large variations in view angle.

Brain status modeling with non-negative projective dictionary learning

Accurate prediction of individuals’ brain age is critical to establish a baseline for normal brain development. This study proposes to model brain development with a novel non-negative projective dictionary learning (NPDL) approach, which learns a discriminative representation of multi-modal neuroimaging data for predicting brain age. Our approach encodes the variability of subjects in different age groups using separate dictionaries, projecting features into a low-dimensional manifold such that information is preserved only for the corresponding age group. The proposed framework improves upon previous discriminative dictionary learning methods by incorporating orthogonality and non-negativity constraints, which remove representation redundancy and perform implicit feature selection. We study brain development on multi-modal brain imaging data from the PING dataset (N = 841, age = 3−21 years). The proposed analysis uses our NDPL framework to predict the age of subjects based on cortical measures from T1-weighted MRI and connectome from diffusion weighted imaging (DWI). We also investigate the association between age prediction and cognition, and study the influence of gender on prediction accuracy. Experimental results demonstrate the usefulness of NDPL for modeling brain development.

MEMS gyroscope wavelet de-noising method based on redundancy and sparse representation

Since the wavelet analysis does not need exact noise model, wavelet decomposition is suitable for analyzing the signal with complex noise characteristics. In this paper, a real-time wavelet de-noising method is used for the error compensation of MEMS gyroscope. In order to avoid the shortcomings that wavelet threshold processing ignores the wavelet coefficient integrity, the sparse and redundant representation methods are used to optimize the wavelet coefficients. The reverse use of compressed sensing algorithm represents the low-frequency wavelet coefficients redundantly, and optimizes the sparse coefficients, and then reconstructs the low-frequency wavelet coefficients. The sparse representation method is used to sparsely represent the high-frequency wavelet coefficients, and the high-frequency wavelet coefficients are optimized. Then the optimized high and low frequency wavelet coefficients are used to reconstruct the original signal and the lag correction algorithm is used to reduce the boundary effect of wavelet decomposition. Finally, the feasibility of the algorithm is verified by simulation and experiments, which provides a new idea for wavelet de-noising.

On visual BMI analysis from facial images

Automatically assessing body mass index (BMI) from facial images is an interesting and challenging problem in computer vision. Facial feature extraction is an important step for visual BMI estimation. This work studies the visual BMI estimation problem based on the characteristics and performance of different facial representations, which has not been well studied yet. Various facial representations, including geometry based representations and deep learning based, are comprehensively evaluated and analyzed from three perspectives: the overall performance on visual BMI prediction, the redundancy in facial representations and the sensitivity to head pose changes. The experiments are conducted on two databases: a new dataset we collected, called the FIW-BMI and an existing large dataset Morph II. Our studies provide some deep insights into the facial representations for visual BMI analysis: 1) The deep model based methods perform better than geometry based methods. Among them, the VGG-Face and Arcface show more robustness than others in most cases; 2) Removing the redundancy in VGG-Face representation can increase the accuracy and efficiency in BMI estimation; 3) Large head poses lead to low performance for BMI estimation. The Arcface, VGG-Face and PIGF are more robust than the others to head pose variations.

Analytics of Unconstrained Convex Problems in Sparse and Redundant Representations: A Case Study on Image Processing Applications

Analytics of Unconstrained Convex Problems in Sparse and Redundant Representations: A Case Study on Image Processing Applications

Interferometric SAR Phase Denoising Using Proximity-Based K-SVD Technique.

This paper addresses the problem of interferometric noise reduction in Synthetic Aperture Radar (SAR) interferometry based on sparse and redundant representations over a trained dictionary. The idea is to use a Proximity-based K-SVD (ProK-SVD) algorithm on interferometric data for obtaining a suitable dictionary, in order to extract the phase image content effectively. We implemented this strategy on both simulated as well as real interferometric data for the validation of our approach. For synthetic data, three different training dictionaries have been compared, namely, a dictionary extracted from the data, a dictionary obtained by a uniform random distribution in , and a dictionary built from discrete cosine transform. Further, a similar strategy plan has been applied to real interferograms. We used interferometric data of various SAR sensors, including low resolution C-band ERS/ENVISAT, medium L-band ALOS, and high resolution X-band COSMO-SkyMed, all over an area of Mt. Etna, Italy. Both on simulated and real interferometric phase images, the proposed approach shows significant noise reduction within the fringe pattern, without any considerable loss of useful information.

RobOMP: Robust variants of Orthogonal Matching Pursuit for sparse representations.

Sparse coding aims to find a parsimonious representation of an example given an observation matrix or dictionary. In this regard, Orthogonal Matching Pursuit (OMP) provides an intuitive, simple and fast approximation of the optimal solution. However, its main building block is anchored on the minimization of the Mean Squared Error cost function (MSE). This approach is only optimal if the errors are distributed according to a Gaussian distribution without samples that strongly deviate from the main mode, i.e. outliers. If such assumption is violated, the sparse code will likely be biased and performance will degrade accordingly. In this paper, we introduce five robust variants of OMP (RobOMP) fully based on the theory of M-Estimators under a linear model. The proposed framework exploits efficient Iteratively Reweighted Least Squares (IRLS) techniques to mitigate the effect of outliers and emphasize the samples corresponding to the main mode of the data. This is done adaptively via a learned weight vector that models the distribution of the data in a robust manner. Experiments on synthetic data under several noise distributions and image recognition under different combinations of occlusion and missing pixels thoroughly detail the superiority of RobOMP over MSE-based approaches and similar robust alternatives. We also introduce a denoising framework based on robust, sparse and redundant representations that open the door to potential further applications of the proposed techniques. The five different variants of RobOMP do not require parameter tuning from the user and, hence, constitute principled alternatives to OMP.

Nonlinearity of perturbations in -symmetric quantum mechanics

In the so called crypto-Hermitian formulation of quantum theory (incorporating, in particular, the -symmetric quantum mechanics as its special case) the unitary evolution of a system is known to be described via an apparently redundant representation of the states in a triplet of Hilbert spaces. Two of them are unitarily equivalent while the auxiliary, zeroth one is unphysical but exceptionally user-friendly. The dynamical evolution equations are, naturally, solved in the friendliest space . The evaluation of experimental predictions then requires a Hermitization of the observables. This yields, as its byproduct, the correct physical Hilbert space . The formalism offers the conventional probabilistic interpretation due to the Dyson-proposed unitary equivalence between and a certain conventional but prohibitively complicated textbook space . The key merit of the innovation lies in its enhanced flexibility opening new ways towards nonlocal or complex-interaction unitary models. The price to pay is that the ad hoc inner product in the relevant physical Hilbert space is, by construction, Hamiltonian-dependent. We show that this implies that the Rayleigh-Schrödinger perturbation theory must be used with due care. The warning is supported by an elementary sample of quantum system near its exceptional-point phase transition. In contrast to a naive expectation, the model proves stable with respect to the admissible, self-consistently specified (i.e., physical) small perturbations.

Ultrasound image restoration based on a learned dictionary and a higher-order MRF

The restoration of images degraded by blur and multiplicative noise is a critical preprocessing step in medical ultrasound images which exhibit clinical diagnostic features of interest. This paper proposes a novel non-smooth non-convex variational model for ultrasound images denoising and deblurring motivated by the successes of sparse representation of images and FoE based approaches. Dictionaries are well adapted to textures and extended to arbitrary image sizes by defining a global image prior, while FoE image prior explicitly characterizes the statistics properties of natural image. Following these ideas, the new model is composed of the data-fidelity term, the sparse and redundant representations via learned dictionaries, and the FoE image prior model. The iPiano algorithm can efficiently deal with this optimization problem. The new proposed model is applied to several simulated images and real ultrasound images. The experimental results of denoising and deblurring show that proposed method gives a better visual effect by efficiently removing noise and preserving details well compared with two state-of-the-art methods.

A new compressive sensing based image denoising method using block-matching and sparse representations over learned dictionaries

Suppressing noise and preserving detail information such as edges and textures are two key challenges in image denoising. In this paper, a new method for eliminating noise from images is presented which is based on not only compressive sensing but also sparse and redundant representations over trained dictionaries. The objective function of the proposed technique consists of two terms. The first term processes the noisy image by the hard thresholding operator in the bandelet domain to provide the noise-free image as well as guaranteeing the similarity between the denoised image and the noisy image, while the second term ensures that the image admits a sparse decomposition in a dictionary. In addition, the proposed method takes advantage of the block-matching technique for representing the dictionary elements such that the noisy image is firstly grouped by the block-matching technique, and then an identical sparse vector is used for all patches in a group. Simulations using images contaminated by additive white Gaussian noise demonstrate that the performance of the proposed method considerably surpasses that of state-of-the-art methods, both visually and in terms of quantitative criteria, namely peak signal to noise ratio and structural similarity.

Changes in Neuronal Representations of Consonants in the Ascending Auditory System and Their Role in Speech Recognition.

A fundamental task of the ascending auditory system is to produce representations that facilitate the recognition of complex sounds. This is particularly challenging in the context of acoustic variability, such as that between different talkers producing the same phoneme. These representations are transformed as information is propagated throughout the ascending auditory system from the inner ear to the auditory cortex (AI). Investigating these transformations and their role in speech recognition is key to understanding hearing impairment and the development of future clinical interventions. Here, we obtained neural responses to an extensive set of natural vowel-consonant-vowel phoneme sequences, each produced by multiple talkers, in three stages of the auditory processing pathway. Auditory nerve (AN) representations were simulated using a model of the peripheral auditory system and extracellular neuronal activity was recorded in the inferior colliculus (IC) and primary auditory cortex (AI) of anaesthetized guinea pigs. A classifier was developed to examine the efficacy of these representations for recognizing the speech sounds. Individual neurons convey progressively less information from AN to AI. Nonetheless, at the population level, representations are sufficiently rich to facilitate recognition of consonants with a high degree of accuracy at all stages indicating a progression from a dense, redundant representation to a sparse, distributed one. We examined the timescale of the neural code for consonant recognition and found that optimal timescales increase throughout the ascending auditory system from a few milliseconds in the periphery to several tens of milliseconds in the cortex. Despite these longer timescales, we found little evidence to suggest that representations up to the level of AI become increasingly invariant to across-talker differences. Instead, our results support the idea that the role of the subcortical auditory system is one of dimensionality expansion, which could provide a basis for flexible classification of arbitrary speech sounds.

An efficient redundant binary adder with revised computational rules

Redundant binary representation (RBR) offers a carry-free addition of two redundant binary (RB) numbers. The computational rules of the conventional RB adder (CRBA) generate intermediate sum and carry vectors in RBR, leads to area overhead and pre-hardware elements for reverse conversion (RC). We have considered that the intermixing of inverted encoding of negabits (IEN) representation and conventional binary bits or posibits can be realized using standard hardware blocks. This paper provides a new computational rule for RB adder generating the intermediate sum and intermediate carry in posibit and IEN representations replacing the redundant digits. Thus, the proposed RB adder provides a single stage RB adder omitting the requirement of intermediate RB digits. For circuit synthesis of the proposed designs, we have considered Encounter® RTL Compiler and Xilinx Synthesis Technology in ASIC and FPGA platforms respectively. The comparative study of proposed NRBA offers improved design parameters as compared to CRBA.

Linear Independent Component Analysis Over Finite Fields: Algorithms and Bounds

Independent Component Analysis (ICA) is a statistical tool that decomposes an observed random vector into components that are as statistically independent as possible. ICA over finite fields is a special case of ICA, in which both the observations and the decomposed components take values over a finite alphabet. This problem is also known as minimal redundancy representation or factorial coding. In this work we focus on linear methods for ICA over finite fields. We introduce a basic lower bound which provides a fundamental limit to the ability of any linear solution to solve this problem. Based on this bound, we present a greedy algorithm that outperforms all currently known methods. Importantly, we show that the overhead of our suggested algorithm (compared with the lower bound) typically decreases, as the scale of the problem grows. In addition, we provide a sub-optimal variant of our suggested method that significantly reduces the computational complexity at a relatively small cost in performance. Finally, we discuss the universal abilities of linear transformations in decomposing random vectors, compared with existing non-linear solutions.