China
Africa
Explore chapters and articles related to this topic
A study for classification technique of cracks in concrete slabs by using pattern recognition
Published in Marc A. Maes, Luc Huyse, Reliability and Optimization of Structural Systems, 2020
Y. Mikumo, K. Yasuda, M. Hirokane, H. Furuta, Y. Kusunose
The digital images of cracks are classified into different damage levels by using the extracted characteristics. The pattern classification involves the techniques for assigning a pattern to its respective class. In this study, it is required to classify the digital images of cracks into the different classes of damage levels. There are many variants of the classification algorithm allowing for faster convergence and more accurate representation. In this study, we used LVQ (Learning Vector Quantization) algorithm for the classifier, which is a type of artificial neural network. LVQ was developed by Kohonen and is based on the self-organizing map. In the LVQ algorithms, the vector quantization is not used to approximate to density functions of the class samples, but to directly define the class borders according to the nearest-neighbour rule. We used the LVQ_PAK package developed in the Technological University of Helsinki.
Firefly Algorithm
Published in A Vasuki, Nature-Inspired Optimization Algorithms, 2020
The firefly algorithm is suitable for non-linear, unimodal, and multimodal optimization problems and is found to be more efficient than GA and PSO. FA is also effective in solving multi-objective optimization problems. Discrete versions of the firefly algorithm have been proposed and are available in the literature with proven good performance. Hybrid optimization algorithms where the FA has been applied in combination with other nature-inspired algorithms have also been proposed. In [3] the firefly algorithm has been applied for the optimization of queueing systems that are used for the analysis and solution of complex problems related to the field of computer science as well as in industries. Vector quantization is a popular technique for image compression. The Linde–Buzo–Gray (LGB) algorithm is normally used to construct the codebook for VQ. In [4] the firefly algorithm has been applied along with the LGB algorithm to construct the optimal codebook. The optimal codebook design is one which maximizes the fitness function for all input vectors. The firefly-LBG algorithm has been compared with other state-of-the-art algorithms, and it has been found to outperform the other optimization algorithms. FA has been found to be suitable for clustering, image classification, feature selection, and for other computer science applications such as graph coloring, network routing, and the famous traveling salesman problem.
Medical Image Processing Environment
Published in Jiří Jan, Medical Image Processing, Reconstruction and Analysis, 2019
The original (spatial) domain block-oriented compression is realized via vector quantization. The ordered set of N × N pixel values is considered a vector in the vector space RN2. Based on statistical analysis of the distribution of the vectors in the space, the space is subdivided into areas such that each area contains similar vectors. A vector that represents reasonably well all the vectors belonging to a certain area is chosen as its representative; finding suitable representatives and respective area borders (the codebook design) is a complex step, possibly performed by self-learning algorithms or neural networks with unsupervised learning. Once the codebook is defined, the vector quantization consists of replacing any input vector by the representative of the area to which the vector belongs. This procedure aims at obtaining a shorter transformed alphabet (of representatives), instead of a much more extensive alphabet of all possible N2-sized vectors. Clearly, the representatives do not describe the elementary images exactly, being only their approximations; the chosen number and positioning of representatives influence the mean error as well as the compression ratio. However, it seems that block-oriented compression via frequency domain is easier and more effective.
Faults and fractures detection using a combination of seismic attributes by the MLP and UVQ artificial neural networks in an Iranian oilfield
Published in Petroleum Science and Technology, 2022
Reza Mohammadi, Mohammad Reza Bakhtiari
Classification, or unsupervised classification, is a sort of cluster analysis that aims to discover rules for categorizing items given a group of pre-classified events. Classification is crucial to learning algorithms, data mining, and pattern identification (Steinbach, Ertöz, and Kumar 2004). UVQ is derived from the learning vector quantizer (LVQ). Vector quantization is a prominent application of competitive learning for data encoding and compression. (Haykin and Lippmann 1994; Hertz, Krogh, and Palmer 2018). The vector representation utilizes combined seismic attributes as test data sets (Ashraf, Zhang, Anees, Nasir Mangi, et al. 2020). Without depending on known examples, the network can be trained by patterns in data sets on the basis of similarities across samples (Mousavi et al. 2019). The UVQ determines the average distance between data points and cluster centers. A modest mean distance indicates that the clusters adequately represent the data. When the unsupervised network is trained, the positions of the vectors are modified to minimize the mean Euclidean distance between each data point and its nearest vector (Bougrain and Alexandre 1998; Mousavi et al. 2019; Hussein, Stewart, and Wu 2020). The UVQ clusters the input data using this way. What the resultant clusters signify is yet unknown. Hence, interpretation must be performed after unsupervised learning (Aminzadeh and De Groot 2006). Figure 3 shows a schematic of all the descriptions, the data analysis procedures, and the study's workflow.
A learning- and scenario-based MPC design for nonlinear systems in LPV framework with safety and stability guarantees
Published in International Journal of Control, 2023
Yajie Bao, Hossam S. Abbas, Javad Mohammadpour Velni
To express the joint uncertainties of matrix functions and scheduling signals by scenario trees, we generate model paths by sampling the scheduling signals and simulating the BNN model and apply reduction techniques to the model paths for generating representative scenarios while preserving the statistical properties of uncertainty quantified by the BNN model. In particular, we use MC sampling methods and K-means, a clustering method in machine learning, to generate scenarios, as the uncertainties are described by a BNN model such that the propagation of uncertainties is intractable to analyse. In particular, MC methods are employed to sample models from the BNN model for selected scheduling trajectories. While Lemma 2.3 claims there exists a scalar such that the trajectories of the sampled models contain the system trajectory, can be too large for online optimisation of the SMPC problems. Instead, we apply K-means clustering to the models to reduce the number of scenarios. K-means clustering is a vector quantisation method which partitions observations/samples into C disjoint clusters by minimising the within-cluster sum-of-squares variances (squared Euclidean distances) , and each cluster is described by the mean (a.k.a. centroid) of the samples in the cluster. We use the cluster centroids of the models sampled from the BNN model as the models of scenarios. However, the C scenarios may lose the property of the models in Lemma 2.3.