Search

Neural network pruning offers great prospects for facilitating the deployment of deep neural networks on computational resource limited devices. Neural architecture search (NAS) provides an efficient way to automatically seek appropriate neural architecture design for compressed model. It is observed that, for existing NAS-based pruning methods, there is usually a lack of layer information when searching the optimal neural architecture. In this paper, we propose a new NAS approach, namely, differentiable channel pruning method guided via attention mechanism (DCP-A), where the adopted attention mechanism is able to provide layer information to guide the optimization of the pruning policy. The training process is differentiable with Gumbel-softmax sampling, while parameters are optimized under a two-stage training procedure. The neural network block with the shortcut is dedicatedly designed, which is of help to prune the network not only on its width but also on its depth. Extensive experiments are performed to verify the applicability and superiority of the proposed method. Detailed analysis with visualization of the pruned model architecture shows that our proposed DCP-A learns explainable pruning policies. [ABSTRACT FROM AUTHOR]

As an important supporting material for calligraphy and painting relics in collections, the silk substrate is generally in acidification. And owing to an extremely high light response, it is very prone to different forms of lighting damage, including color and mechanical aspects. Considering that the Light Emitting Diode (LED) with flexible spectral compositions is widely used in the lighting environment of collections, the spectral responsivity of acidified silk substrates is the core issue concerning the long-term conservation of calligraphy and painting relics. In this study, based on the accelerated aging experiment of nine narrowband LEDs with different wavelengths (λ) in the visible range on silk samples with different acidification degrees (pH values), the infrared spectrum was measured non periodically. Processed with the Principal Component Analysis (PCA), the obtained ΔPC1 was used to evaluate the lighting damage of silk samples. Based on the damage change curve, its spectral responsivity function, that is, the mathematical relationship between lighting damage and λ and pH values, was further fitted. Combined with the previous research on the spectral responsivity of colorants, an evaluation method for lighting damage was proposed, which is the key to the conservation of calligraphy and painting relics. [ABSTRACT FROM AUTHOR]

State-of-the-art deep neural network plays an increasingly important role in artificial intelligence, while the huge number of parameters in networks brings high memory cost and computational complexity. To solve this problem, filter pruning is widely used for neural network compression and acceleration. However, existing algorithms focus mainly on pruning single model, and few results are available to multi-task pruning that is capable of pruning multi-model and promoting the learning performance. By utilizing the filter sharing technique, this paper aimed to establish a multi-task pruning framework for simultaneously pruning and merging filters in multi-task networks. An optimization problem of selecting the important filters is solved by developing a many-objective optimization algorithm where three criteria are adopted as objectives for the many-objective optimization problem. With the purpose of keeping the network structure, an index matrix is introduced to regulate the information sharing during multi-task training. The proposed multi-task pruning algorithm is quite flexible that can be performed with either adaptive or pre-specified pruning rates. Extensive experiments are performed to verify the applicability and superiority of the proposed method on both single-task and multi-task pruning. [ABSTRACT FROM AUTHOR]

As one of the most important structural and functional materials, rolled-product plays an irreplaceable role in national economy, people's lives and national development. Metallurgy industry is moving from traditional semi-automation to knowledge automation, process intelligence, and manufacturing information. Rolling exerts an essential impact on material properties and product quality as an important part of the production in steel industry. Rolling process is multi-scale, multi-variable, nonlinear and unbalanced with strong coupling and non-steady state. With an increasing rolling speed, more difficulties like process information monitoring, behavior characteristics modeling, and controlling of high speed operating are manifested in high-speed continuous rolling mills. The existing control system of rolling process is difficult to cope with the condition changes of high-speed rolling and the specification changes of complex products. The main reason lies in the fact that the prediction of force parameters is based on traditional mathematical models, and the procedure parameter setting depends on static optimization methods. In order to achieve precise control of large-scale and high-speed rolling, the analysis of rolling process rules based on industrial big data should be considered to establish the dynamic process model, and multi-objective real-time computational method of rolling schedules should also be introduced. Through a summary of steel industry and a review of the history of rolling optimization, the purpose is to explore the relationship between the optimization objectives of the rolling schedule and the process parameters of the rolling process, reveal the rules of how rolling conditions affecting rolling process in high speed rolling and provide theoretical basis and technical support to the production of steel industry. [ABSTRACT FROM AUTHOR]

In this paper, we propose a cavity-coupled perfect absorber (CCPA) consisting of a bottom gold layer, a SiO2 layer, and multilayer MIM (metal–insulator-metal) nanowires. By coupling the Fabry-Pérot (F-P) cavity mode with the magnetic plasmons (MPs) resonance mode in multilayer MIM structure, the MPs resonance in the multilayer MIM structure is enhanced and leads to increased absorption of the structure. By designing and optimizing the parameters, near-perfect absorption in the mid-infrared wavelengths can be achieved. The positions of the two absorption peaks can be adjusted by changing the width of the nanowires and the thickness of the SiO2 layer, resulting in the formant excitation at any position in the mid-infrared. The full-width at half-maximum (FWHM) of the short-wave peak is only 0.076 µm, and the FWHM of the long-wave peak is only 0.46 µm. The long-wave absorption peak maintains high efficiency and stability even at large incident angles, exhibiting “omnidirectional” characteristics. The tunable mid-infrared emission peak can be matched to the characteristic spectra of gases, making it suitable for infrared radiation sources in gas detection. [ABSTRACT FROM AUTHOR]

Accurate registration of chest radiographs plays an increasingly important role in medical applications. However, most current intensity-based registration methods rely on the assumption of intensity conservation that is not suitable for alignment of chest radiographs. In this study, we propose a novel algorithm to match chest radiographs, for which the conventional residual complexity (RC) is modified as the similarity measure and the cubic B-spline transformation is adopted for displacement estimation. The modified similarity measure is allowed to incorporate the neighborhood influence into variation of intensity in a justified manner of the weight, while the transformation is implemented with a registration framework of pyramid structure. The results show that the proposed algorithm is more accurate in registration of chest radiographs, compared with some widely used methods such as the sum-of-squared-differences (SSD), correlation coefficient (CC) and mutual information (MI) algorithms, as well as the conventional RC approaches. [ABSTRACT FROM AUTHOR]

Hyperspectral unmixing is essential for efficient hyperspectral image processing. Nonnegative matrix factorization based on minimum volume constraint (MVC-NMF) is one of the most widely used methods for unsupervised unmixing for hyperspectral image without the pure-pixel assumption. But the model of MVC-NMF is unstable, and the traditional solution based on projected gradient algorithm (PG-MVC-NMF) converges slowly with low accuracy. In this paper, a novel parallel method is proposed for minimum volume constrained hyperspectral image unmixing on CPU-GPU Heterogeneous Platform. First, a optimized unmixing model of minimum logarithmic volume regularized NMF is introduced and solved based on the second-order approximation of function and alternating direction method of multipliers (SO-MVC-NMF). Then, the parallel algorithm for optimized MVC-NMF (PO-MVC-NMF) is proposed based on the CPU-GPU heterogeneous platform, taking advantage of the parallel processing capabilities of GPUs and logic control abilities of CPUs. Experimental results based on both simulated and real hyperspectral images indicate that the proposed algorithm is more accurate and robust than the traditional PG-MVC-NMF, and the total speedup of PO-MVC-NMF compared to PG-MVC-NMF is over 50 times. [ABSTRACT FROM AUTHOR]

Spatial rich model (SRM) steganalysis feature is formed by high-order statistics collected from image noise residuals. These statistics are simply rescaled before machine learning. It is noted that SRM features of different cover images are very different. In this paper, we propose a feature normalization method based on random feature-subsets (NRS) for SRM. We randomly draw feature-subsets from SRM feature. Then these feature-subsets are normalized by using per-sample rescaling method to make the feature-subsets of different images have the same 1-norm (the sum of all elements). The proposed NRS method can adjust the feature distribution and increase the feature diversity. The normalized feature-subsets can achieve better detection performance and also can be used as a useful complement for the existing steganalysis features. Experimental results show that: (1) a small amount of normalized feature-subset supplement can obviously improve the detection performance of SRM feature; (2) under the same dimensionality, the proposed NRS version feature can achieve a better detection accuracy than that of original feature; (3) the proposed NRS method is applicable to projections of spatial rich model feature; (4) compared with steganalysis feature extraction, the computational time of NRS can be negligible. [ABSTRACT FROM AUTHOR]

An enhancement-based Poisson denoising method for photon-limited images is presented. The noisy image is firstly pre-processed for enhancing incomplete object information, and then it is denoised while preserving the restored structural details. A variational regularization model based on Euler's elastica (EE) is proposed for image enhancement pre-processing. A nonlocal total variation (NLTV) regularization model is then employed in the second stage of image denoising. The above two optimization problems are solved by the alternating direction method of multipliers (ADMM). For Poissonian images with low image peak values, experiments demonstrate the validity and efficiency of the proposed method for both restoring geometric structure and removing noise. [ABSTRACT FROM AUTHOR]

Spatial rich model (SRM) is a classic steganalysis method, which collects high-order co-occurrences from truncated noise residuals as feature to capture the local-range dependencies of an image. Increasing the truncation threshold and the co-occurrence order will lead to a higher-dimensional feature, which can exploit more statistical bins and capture dependencies across larger-range neighborhood, but this will suffer from the curse of dimensionality. In this paper, we propose a fast projection method to increase the statistical robustness of the higher-dimensional SRM feature while decreasing its dimensionality. The proposed projection method is applicable to co-occurrence-based steganalysis features. The detection performance and the computational complexity of the proposed method are investigated on three content-adaptive steganographic algorithms in spatial domain. [ABSTRACT FROM AUTHOR]

Blind deblurring, typically underdetermined or ill-posed problem, has attracted numerous research studies over the recent years. Various priors of either the image or the blur kernel are proposed to establish various regularization models to estimate the blur kernel. And sharp edges are often employed as an important clue to recover the blur kernel. However, due to the harmful effects caused by textures and various artifacts, sharp edges are not always beneficial to the kernel estimation. To address this problem, this paper presents a step-edge based blind image deblurring algorithm using steerable gradients. The proposed algorithm adopts a coarse-to-fine multiscale framework with step-edge restoration, kernel estimation and latent image estimation. In each scale, the step-edges are detected and refined through fast image decomposition and thresholding on steerable gradients, while the kernel and latent image are estimated by minimizing the quadratic energy functionals with steerable gradients. Because each of the minimizing functional has a closed-form solution, and can be implemented by using FFTs, our algorithm is also very fast. Experimental results on both synthetic and real data demonstrate that our method outperforms most existing single image blind deblurring methods. [ABSTRACT FROM AUTHOR]

The SRM (Spatial Rich Model) is a very effective steganalysis method. It uses statistics of neighboring noise residual samples as features to capture the dependency changes caused by embedding. Because the noise residuals are the high-frequency components of image and closely tied to image content, the residuals of different types of image regions have different statistical properties and effectiveness for steganalysis. In this paper, the effectiveness of the residuals is investigated. Then the effectiveness of the statistics collected from different types of neighboring residual samples is investigated from the FLD (Fisher Linear Discriminant) viewpoint, and ineffective, effective and high-effective neighboring residual samples are defined. The ineffective neighboring residual samples are not likely to change during embedding, and if they are counted in statistics, they may mix the features with noise and make the features impure. Pure SRM features are extracted based on neighboring noise residual sample selection strategy. Furthermore, multi-order statistical features are proposed to increase the statistical diversity. Steganalysis performances of the statistical features collected from different types of neighboring residual samples are investigated on three content adaptive steganographic algorithms. Experimental results demonstrate that the proposed method can achieve a more accurate detection than SRM. [ABSTRACT FROM AUTHOR]

An efficient two-stage non-blind deblurring framework is proposed for recovering blurred images progressively. To this date, most approaches commonly solve a single variational regularization problem incorporated with chosen priors, limiting the attained restoration quality. To address this, two different priors are adopted in separated stages to restore an image in a coarse-to-fine manner and each stage follows a variational regularization scheme. In the first stage, salient edges and large scale textures are produced by minimizing the e norm of gradient. The intermediate result is then refined by non-local auto regression model in the next stage. Finally, experimental results demonstrate that the proposed methodology is efficient and achieves nice performance. [ABSTRACT FROM AUTHOR]

Recently an infeasible interior-point algorithm for linear programming (LP) was presented by Liu and Sun. By using similar predictor steps, we give a (feasible) predictor-corrector algorithm for convex quadratic programming (QP). We introduce a (scaled) proximity measure and a dynamical forcing factor (centering parameter). The latter is used to force the duality gap to decrease. The algorithm can decrease the duality gap monotonically. Polynomial complexity can be proved and the result coincides with the best one for LP, namely, $O(\sqrt{n}\log n\mu^{0}/\varepsilon)$. [ABSTRACT FROM AUTHOR]

The total variation model proposed by Rudin, Osher, and Fatemi performs very well for removing noise while preserving edges. However, it favors a piecewise constant solution in BV space which often leads to the staircase effect, and small details such as textures are often filtered out with noise in the process of denoising. In this paper, we propose a fractional-order multi-scale variational model which can better preserve the textural information and eliminate the staircase effect. This is accomplished by replacing the first-order derivative with the fractional-order derivative in the regularization term, and substituting a kind of multi-scale norm in negative Sobolev space for the L norm in the fidelity term of the ROF model. To improve the results, we propose an adaptive parameter selection method for the proposed model by using the local variance measures and the wavelet based estimation of the singularity. Using the operator splitting technique, we develop a simple alternating projection algorithm to solve the new model. Numerical results show that our method can not only remove noise and eliminate the staircase effect efficiently in the non-textured region, but also preserve the small details such as textures well in the textured region. It is for this reason that our adaptive method can improve the result both visually and in terms of the peak signal to noise ratio efficiently. [ABSTRACT FROM AUTHOR]