Use Of Convolution Research Articles

Learning in Deep Radial Basis Function Networks.

Learning in neural networks with locally-tuned neuron models such as radial Basis Function (RBF) networks is often seen as instable, in particular when multi-layered architectures are used. Furthermore, universal approximation theorems for single-layered RBF networks are very well established; therefore, deeper architectures are theoretically not required. Consequently, RBFs are mostly used in a single-layered manner. However, deep neural networks have proven their effectiveness on many different tasks. In this paper, we show that deeper RBF architectures with multiple radial basis function layers can be designed together with efficient learning schemes. We introduce an initialization scheme for deep RBF networks based on k-means clustering and covariance estimation. We further show how to make use of convolutions to speed up the calculation of the Mahalanobis distance in a partially connected way, which is similar to the convolutional neural networks (CNNs). Finally, we evaluate our approach on image classification as well as speech emotion recognition tasks. Our results show that deep RBF networks perform very well, with comparable results to other deep neural network types, such as CNNs.

Enhanced open biomass burning detection: The BranTNet approach using UAV aerial imagery and deep learning for environmental protection and health preservation

Open biomass burning (OBB) in agriculture presents a significant and well-documented challenge, posing severe consequences for both environmental and human health. OBB releases air pollutants that degrade air quality and contribute to climate change, leading to premature deaths in regions with high concentrations of open crop straw burning (OCSB) emissions. Although policies aimed at prohibiting OBB are in place, the efficacy of these regulations in mitigating OCSB emissions remains ambiguous. Consequently, early prevention and monitoring of open biomass combustion are imperative for environmental preservation. Traditional monitoring techniques, reliant on fixed-position cameras, are constrained by their location and monitoring intensity, making concealed fire recognition a complex problem. To address this limitation and monitor the human living environment more flexibly and accurately, we propose a new method to identify straw fires in UAV Aerial Image Using CNN Branch Reinforce Transformer which named BranTNet, enabling early detection and rapid response to crop straw fires. By integrating computer vision technology and deep learning algorithms, straw fires in UAV-acquired aerial survey images can be detected and categorized. In the realm of artificial intelligence algorithms, we skillfully merge convolution and attention mechanisms, harnessing the full potential of both methodologies. Moreover, we seamlessly incorporate transfer learning, skillfully unifying self-training convolution modules with pre-trained transformer modules. This strategic amalgamation not only minimizes time costs but also ensures optimal experimental outcomes. Regarding data, we meticulously collected a substantial number of authentic samples, ensuring the sufficiency of our experimental dataset. The experimental results demonstrate that our proposed method exhibits exceptional accuracy and robustness in detecting and identifying straw fires in UAV aerial survey images. Our approach outperforms the use of convolution or attention mechanisms alone. By integrating this approach with drone technology, we unlock the potential for developing more versatile and precise monitoring solutions, expanding the application of drones to diverse domains. This progress contributes significantly to the early detection and prevention of crop straw fires, fundamentally reducing environmental pollution, curbing carbon emissions, and advancing the cause of carbon neutrality. This innovative technique for monitoring and preventing OBB holds substantial promise in mitigating the adverse effects of OBB on the environment and human health.

Hierarchical interaction and pooling network for co-salient object detection

Co-salient object detection (CoSOD) aims to detect common and salient objects across the given image group. Due to the particularity of CoSOD, images in the given group are processed synergistically to excavate the relevance between them. Inspired by the tracking methods, previous works tend to utilize pixel-wise correspondences to measure the relevance. However, because of the complexity of the image groups, the obtained feature maps could be easily affected by the common interference. Moreover, current works tend to utilize classification labels to ensure intra-group coherence and inter-group separability, which may cause some overfitting problems. In this paper, we propose the hierarchical interaction and pooling network to alleviate the above problems. We first design a pyramid pooling interaction module and perform convolution with dimension permutation, making full use of convolution and multi-receptive field information. We further propose the coherence confirmation module along with the four-branch architecture. Without the classification labels, the module achieves comparable or even better performance. Extensive experiments demonstrate that the proposed method can detect common and salient objects more accurately and achieves the new state-of-the-art.

Image processing using the quantum quaternion Fourier transform

There is a growing interest in quantum image processing (QIP) that rose from the desire to exploit the properties of quantum computing to improve the performance of classical techniques and their applications. Since the introduction of a quaternion by Hamilton in 1843, quaternions had been used in a lot of applications. We show how the quantum states and quantum operators work in the field of quaternions and how we can use it to extend the quantum complex Fourier transform (QCFT) to the quantum quaternion Fourier transform (QQFT). We propose for the first time the novel algorithm for the QQFT. We explain the impossibility of the use of convolution in quantum image processing; we design filters in the quaternion Fourier spectrum and the space domain an edge detector and a quantum median filter.

Inversion of seismic source parameters from satellite InSAR data based on deep learning

Satellite interferometric synthetic aperture radar (InSAR) has been playing an important role in the earthquake surface deformation observation and source model inversions with its advantage of fine spatial resolution in comparison with seismological data. At present, the two-step inversion approach, including a nonlinear inversion step for determining the fault geometry (e.g., length, width, depth, strike, dip, rake, slip) and a linear inversion step for estimating the slip distribution, is widely used to obtain seismic source parameters from satellite InSAR data. However, the nonlinear inversion step has some weaknesses, such as a prior knowledge requirement, local minima solution, complex and time-consuming. Coseismic differential interferogram is input into back propagation neural network (BPNN) to perform real time inversion of fault geometry in the previous study. But due to the simpleness of network structure, rake and slip parameters is fixed and inversion accuracy is limited. In this paper, we propose a deep learning approach of Earthquake Source Parameters Inversion using ResNet (abbreviated as ESPI-ResNet) from satellite InSAR data. ESPI-ResNet is able to firstly classify the 4 fault types and then invert the 7 fault geometry parameters based on uniform slip model. We train our model by firstly creating a large dataset of simulated interferograms with source parameters ranges confined by seismological and geological data. The accuracy of fault type classification is 99.6% and root mean squared error (RMSE) of inversed fault geometry parameters is low in the simulated test dataset. To find most suitable model for seismic source parameters inversion, we further compare the inversed accuracy of different networks, including BPNN, VGG-16, ResNet-18, ResNet-34, ResNet-50 and DenseNet. We find that moderate increase of network depth and the use of convolution, deep residual learning can improve the model's performance for source parameters inversion, and therefore ResNet-34 is chosen as the backbone network in this study. Finally, real differential interferograms of Yutian earthquake (2020 Mw 6.3), Jiuzhaigou earthquake (2017 Mw 6.5) and Menyuan earthquake (2016 Mw 5.9) are used to validate the proposed method. Real earthquakes validation shows all of them have correct fault type classification, and the inverted results are consistent with the current seismic source parameters data.

Parametric method for evaluation the state of complex organizational objects

This article proposes an algorithm and mathematical model for a comprehensive assessment of the scope of activities of complex organizational objects. The mathematical model is based on informational entropy and involves the use of convolution of quality indicators in an integrated integrated assessment. When implementing this mathematical model, such important qualitative indicators of an integrated assessment of the quality of organizations’ activities as the objectivity of assessment, scalability, simplicity and usability can grow. Reviewed in more detail the problem of parametric evaluation of the activity of local authorities in general and in the areas of activity is considered. A formula is proposed for calculating a comprehensive assessment indicator for areas of activity and local authorities in general. Based on expert assessments, weights of indicators were determined and based on statistical information, the work presents the results of a comprehensive assessment of local authorities. Comprehensive assessments are used to support decision making by machine learning algorithms.

GLOBAL MESSAGE PASSING IN NETWORKS VIA TASK-DRIVEN RANDOM WALKS FOR SEMANTIC SEGMENTATION OF REMOTE SENSING IMAGES

Abstract. The capability of globally modeling and reasoning about relations between image regions is crucial for complex scene understanding tasks such as semantic segmentation. Most current semantic segmentation methods fall back on deep convolutional neural networks (CNNs), while their use of convolutions with local receptive fields is typically inefficient at capturing long-range dependencies. Recent works on self-attention mechanisms and relational reasoning networks seek to address this issue by learning pairwise relations between each two entities and have showcased promising results. But such approaches have heavy computational and memory overheads, which is computationally infeasible for dense prediction tasks, particularly on large size images, i.e., aerial imagery. In this work, we propose an efficient method for global context modeling in which at each position, a sparse set of features, instead of all features, over the spatial domain are adaptively sampled and aggregated. We further devise a highly efficient instantiation of the proposed method, namely learning RANdom walK samplIng aNd feature aGgregation (RANKING). The proposed module is lightweight and general, which can be used in a plug-and-play fashion with the existing fully convolutional neural network (FCN) framework. To evaluate RANKING-equipped networks, we conduct experiments on two aerial scene parsing datasets, and the networks can achieve competitive results at significant low costs in terms of the computational and memory.

Image processing of concentrated and scattered objects

In modern control systems and information processing, the recognition of objects in the image is complicated by the fact that the impact of negative factors introduces uncertainty into this process, leading to blurring of images. In this regard, it is necessary to develop models and algorithms that would reduce the degree of uncertainty in image processing. These models are necessary, for example, when monitoring environmentally hazardous objects, for search and detection of unauthorized burial of household waste, in the field of information security, in the analysis of x-rays and thermograms, in the actions of unmanned aerial vehicles of law enforcement agencies in autonomous mode. This article presents a description of information technology for recognition in the automated mode of objects in images. The basis of this technology is the algorithm of contour analysis of images. The main distinguishing feature of the algorithm is the use of convolution of the image in four directions, as well as the tracing procedure. The aim of the study was to develop algorithms for high-speed automated visualization of external objects. We present the results of the study of the algorithm of contour analysis in the processing of various images in the visible and infrared wavelengths. Recommendations are formulated for the choice of parameters of the contour analysis algorithm, such as the mean square deviation in image blur, minimum and maximum thresholds for filtering. The results of the study can be used in production management systems, life support of the city, technical vision, environmental conditions, monitoring of business processes, as well as in the creation of simulators for training operators of complex systems, etc. In addition, we show the expediency of applying the algorithm we developed in decision support systems.

Automation of recognition of chemicals using electronic sensor technology based on neural network data processing

The composition of the initial substance was determined using an electronic sensor “electronic nose”, consisting of 8 sensors, to which air was supplied with a syringe with alcohol vapor containing various kinds of impurities. The signal from the sensors was recorded with a sampling frequency of 1 s for 120 s. The output of the device was presented in two different interpretations - in the form of curves obtained from each sensor, or the areas under the curves. The purpose of the work is to build a recognition system for 11 impurities and water in the starting material. The composition of the initial substance was determined using an “electronic nose”, which allows one to obtain 120 values from each of 8 sensors in the form of curves or the values of the areas under the curves. A large number of classes (12), the dynamic presentation of the source data information make it advisable to build a pattern recognition system based on a neural network - a multilayer perceptron trained on the basis of the error back propagation algorithm. When training the network, existing samples are used, indicating which class they belong to. The properties of each analyte are represented as a vector of 120 values of 8 attributes defining one of 12 classes. To reduce the dimensionality of the input data of the neural network, the authors proposed the use of convolution of the available information without significant loss of information capacity of signs by constructing 8 polynomial regressions of the 19th order that describe the curves from each of the 8 sensors of the “electronic nose”. The input matrix obtained as a result of convolution consisted of 20 polynomial regression coefficients of each of 8 curves for 12 classes under consideration. A two-layer neural network with 43 neurons and a sigmoidal activation function in the hidden layer and 12 neurons and a linear activation function in the output layer was constructed. As a result of network training, 2 classification errors were obtained, which allows us to use the approach proposed by the authors to build a recognition system based on preliminary convolution of data dynamically obtained from the “electronic nose”

Deep Learning: Current State

Deep learning, a derived from machine learning, has grown into widespread usage with applications as diverse as cancer detection, elephant spotting, and game development. The number of published studies shows an increasing interest by researchers because of its demonstrated ability to achieve high performance in the solution of complex problems, the wide availability of data and computing resources, and the groundbreaking development of effective algorithms. This paper reviews the current state of deep learning. It includes a revision of basic concepts, such as the operations of feed forward and backpropagation, the use of convolution to extract features, the role of the loss function, and the optimization and learning processes; the survey of main stream techniques, in particular convolutional, recurrent, recursive, deep belief, deep generative, generative adversarial, and variational auto-enconder neural networks; the description of an ample array of applications organized by the type of technique employed; and the discussion of some of its most intriguing open problems.

Correlation analysis and prediction of personality traits using graphic data collections

The questions of building mechanisms for identifying patterns and building modern tools for analyzing data from social networks are considered. It is proposed to apply modern methods of web pages’ automatic analysis, testing hypotheses about the presence of correlation links, automatic classification of graphic information using the apparatus of artificial neural networks. The presence of correlation between personality traits of the "Big Five" is investigated. Strong fluctuations in the values of personality traits were revealed depending on various types for groups of people. The problem of predicting the personality traits of the Internet user by the images posted by him is investigated, artificial neural networks are used as a tool. Two series of experiments were carried out, in the first series, a convolutional neural network, trained on the images and results of the NEO-FFI questionnaire, was used to predict personality traits. The sequential use of convolution and subsampling in the convolution network leads to the so-called increase in the level of features: if the first layer extracts local features from the image, then subsequent layers extract common features that are called high-order features. In the second series of experiments, this type of artificial neural network was used to extract high-level features, which were then used to train a direct distribution network that performs forecasting. Thus, the more layers are used, the more features associated with personality traits are extracted from the images. For processing arrays of graphic information, the “Microsoft Cognitive Toolkit library” and the Nvidia Geforce GTX 1080 Ti graphics accelerator were used. The results of the experiments revealed those personality traits that are most correlated with the images posted by Internet users.

Some properties of a class of analytic functions

Making use of convolution, we introduce and investigate a certain class of functions which is analytic in the open unit disk. We obtain interesting properties of starlikeness and convexity for this function class. Special cases and some useful consequences of our main results are also mentioned.

A heat flow approach to Onsager’s conjecture for the Euler equations on manifolds

We give a simple proof of Onsager's conjecture concerning energy conservation for weak solutions to the Euler equations on any compact Riemannian manifold, extending the results of Constantin-E-Titi and Cheskidov-Constantin-Friedlander-Shvydkoy in the flat case. When restricted to $\mathbb{T}^{d}$ or $\mathbb{R}^{d}$, our approach yields an alternative proof of the sharp result of the latter authors. Our method builds on a systematic use of a smoothing operator defined via a geometric heat flow, which was considered by Milgram-Rosenbloom as a means to establish the Hodge theorem. In particular, we present a simple and geometric way to prove the key nonlinear commutator estimate, whose proof previously relied on a delicate use of convolutions.

A spectral nudging method for the ACCESS1.3 atmospheric model

Abstract. A convolution-based method of spectral nudging of atmospheric fields is developed in the Australian Community Climate and Earth Systems Simulator (ACCESS) version 1.3 which uses the UK Met Office Unified Model version 7.3 as its atmospheric component. The use of convolutions allow for flexibility in application to different atmospheric grids. An approximation using one-dimensional convolutions is applied, improving the time taken by the nudging scheme by 10–30 times compared with a version using a two-dimensional convolution, without measurably degrading its performance. Care needs to be taken in the order of the convolutions and the frequency of nudging to obtain the best outcome. The spectral nudging scheme is benchmarked against a Newtonian relaxation method, nudging winds and air temperature towards ERA-Interim reanalyses. We find that the convolution approach can produce results that are competitive with Newtonian relaxation in both the effectiveness and efficiency of the scheme, while giving the added flexibility of choosing which length scales to nudge.

Further Investigations into the Effects of Temperature on a 975 nm Tapered Laser Bar Using Convolution to Ascertain the Dominant Effect of Temperature on a Laser Bar

Temperature continues to be an issue in the reliability of high power laser diodes. Any effort, therefore, made to understand the dynamics of temperature on the performance of laser diodes is important. This is because it serves as a catalyst for the generation of nonradiative recombination centers, which ultimately kills laser diodes. In this paper, the convolution of simulation results was done to compare it with experimental results, which revealed more details that hitherto, could not have been captured by experiments alone. The inability of experiments to capture these details is due to the limited spatial resolution of the charged-coupled device (CCD) camera, which is approximately 30 µm much larger than the thickness of the quantum well active region used in the experiment. The convolution results showed a further smile-shaped profile within the four groups of array emitters, which in the experimental results, were considered as one emitter due to the limited spatial resolution of the CCD camera. The use of convolution to determine more details was investigated, due to the dominant effect of temperature found across the output power distribution of high power semiconductor laser diode bars.

Use of convolution and geotechnical rock properties to analyze free flowing discharge test

Discharge tests are conducted in order to determine aquifers transmissivity (T) and storage coefficient (S). However the interpretation of test data is not unique and the results may vary depending on adopted hypothesis. In this work, the convolution technique is applied for the reconstruction of observed drawdown curves aiming for the reduction of uncertainty. The interference between a flowing well and an observation well is evaluated in order to determine the hydrogeological parameters of the confined Guarani Aquifer System (Araujo et al. 1999). Discharge test data are analyzed according to the Jacob and Lohman (1952) method. The application of convolution enabled the determination of the most reliable solution. Obtained transmissivity (T = 411.0 m2/d) and storage coefficient (S = 2.75x10–4) are close to values estimated by direct evaluation of the geotechnical sandstone properties.

A lower-variance randomized algorithm for approximate string matching

Several randomized algorithms make use of convolution to estimate the score vector of matches between a text string of length N and a pattern string of length M, i.e., the vector obtained when the pattern is slid along the text, and the number of matches is counted for each position. These algorithms run in deterministic time O(kNlogM), and find an unbiased estimator of the scores whose variance is (M−c)2/k where c is the actual score; here k is an adjustable parameter that provides a tradeoff between computation time and lower variance. This paper provides an algorithm that also runs in deterministic time O(kNlogM) but achieves a lower variance of min(M/k,M−c)(M−c)/k. For all score values c that are less than M−(M/k), our variance is essentially a factor of k smaller than in previous work, and for M−(M/k)<c⩽M it matches the previous workʼs variance. As in the previous work, our estimator is unbiased, and we make no assumption about the probabilistic characteristics of the input, or about the size of the alphabet, and our solution extends to string matching with classes, class complements, “never match” and “always match” symbols, to the weighted case and to higher dimensions.

Use of convolution and filtration to extract information from the DCM graphic files

The paper describes the problem of using digital filters on the DICOM graphic files. An attempt has been made to improve the process of extracting information from graphic files generated by medical measuring instruments.

Randomly curved runs interrupted by tumbling: A model for bacterial motion

Small bacteria are strongly buffeted by Brownian forces that make completely straight runs impossible. A model for bacterial motion is formulated in which the effects of fluctuational forces and torques on the run phase are taken into account by using coupled Langevin equations. An integrated description of the motion, including runs and tumbles, is then obtained by the use of convolution and Laplace transforms. The properties of the velocity-velocity correlation function, of the mean displacement, and of the two relevant diffusion coefficients are examined in terms of the bacterial sizes and of the magnitude of the propelling forces. For bacteria smaller than E. coli, the integrated diffusion coefficient crosses over from a jump-dominated to a rotational-diffusion-dominated form.

On the convergence of convolved vector subdivision schemes

The aim of this paper is to discuss the use of convolution to build new families of refinable functions and thus new families of subdivision algorithms. In particular, we first generalise a result given in [C. Conti, K. Jetter, in: A. Cohen, R. Rabut, L.L. Schumaker (Eds.), Curves and Surfaces Fitting, Saint Malo, 1999, Vanderbilt University Press, Nashville, TN, 2000, pp. 135–142] where, following a too restrictive definition of convergence, it is proved that convolution of two refinable function vectors, which give rise to convergent subdivision schemes, generates a convolved scheme which is again convergent. Then, we discuss the use of convolution to construct refinable bases suitable to interpolate directions given by difference of control points. This turns out to be useful, for example, in algebraic numerical grid generation methods such as the one discussed in [C. Conti, R. Morandi, in: Proceedings of MASCOT/02, IMACS Ser. Comput. Appl. Math., IMACS, New Brunswick, NJ, 2003, pp. 75–82].