China
Africa
Explore chapters and articles related to this topic
Finger Biometric Recognition with Feature Selection
Published in Karm Veer Arya, Robin Singh Bhadoria, The Biometric Computing, 2019
Asish Bera, Debotosh Bhattacharjee, Mita Nasipuri
A traditional biometric system is divided into four key modules, namely, the image acquisition (or sensor) module, feature extraction (and selection) module, template matching module, and decision-making module. Prior to feature computation, a robust preprocessing method is followed for finger segmentation. Then, either the geometric measurements or the shape-based descriptors from the fingers are computed (Dutağaci et al., 2008; Luque-Baena, Elizondo, López-Rubio, Palomo, and Watson, 2013). The pixel-level information is transformed into a descriptor, such as the Fourier descriptors (FDs) (Kang and Wu, 2014), shape context (SC) (Hu et al., 2012), wavelet descriptors (WDs) (Sharma, Dubey, Singh, Saxena, and Singh, 2015), scale-invariant feature transform (SIFT) (Charfi, Trichili, Alimi, and Solaiman, 2014), and others. Both the two types of features, that is, geometric and shape based) are equally imperative to represent an identity uniquely. Notably, a conflation of both types of features can also endow satisfactory results. The defined features (i.e., either geometric or shape based) are not capable enough to contribute adequately in performance evaluation. Hence, only the significant features should be chosen using a suitable feature selection technique in the training phase (Guyon and Elisseeff, 2003; Lashkia and Anthony, 2004). The aim is not only to compute a diverse set of salient features but also to select a minimal subset of discriminative features. The insignificant features are discarded which implies the dimensionality reduction of the input feature set. A pictorial representation of the finger biometric system with the precise tasks during the training and testing phases is shown in Figure 5.1.
Feature Extraction and Learning for Visual Data
Published in Guozhu Dong, Huan Liu, Feature Engineering for Machine Learning and Data Analytics, 2018
Parag S. Chandakkar, Ragav Venkatesan, Baoxin Li
Shape context can be used to match two shapes because corresponding points on two shapes should have similar features. The similarity is measured by the χ2 $ {\chi^2} $ statistic. Suppose that p1 $ {p_1} $ and p2 $ {p_2} $ are two corresponding points on two shapes. Let the normalized shape context histograms of these two points be denoted by h1 $ {h_1} $ and h2 $ {h_2} $ . The χ2 $ {\chi^2} $ statistic is given by C=12∑k=1K(h1-h2)2h1+h2. $$ C = \frac{1}{2}\mathop \sum \limits_{k = 1}^K \frac{{{{({h_1} - {h_2})}^2}}}{{{h_1} + {h_2}}}. $$
3D coronary artery elastic registration based on differential invariant signatures
Published in Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2022
In the field of computer vision and medical imaging, point set registration (or point matching) is one of the most important research issues, which serves as a significant step in registering datasets into a more accurate model, and determines a spatial transformation to align two point sets. Point set registration is now applied in many aspects such as medical image analysis (Rasoulian et al. 2012; Wu et al. 2014; Fang et al. 2019; Zhu et al. 2020), shape matching (Lindner et al. 2014), character recognition (Zhang et al. 2016; Xu et al. 2017), and feature detection, etc. Hierarchical attribute matching mechanism for elastic registration (HAMMER) was introduced in Shen and Davatzikos (2002) which used geometric moment invariants (GMIs) to establish anatomical correspondences in the deformation and was successfully applied to magnetic resonance images of the brain. Belongie et al. (2002) created shape context as a shape descriptor, and then object recognition can be achieved by matching this feature with a priori knowledge of the shape context of the boundary points of the object. Fitzgibbon (2003) introduced a method of registering point sets using general-purpose non-linear optimisation to minimise registration error directly. 2D/3D registration method is established by a point-of-interest tracking network, then the 3D pose of pre-intervention data is estimated through a triangulation layer (Liao et al. 2019).
Fast binary shape categorization
Published in The Imaging Science Journal, 2019
Shape context (SC) [18] is a local shape description in which parts of shapes are divided into landmark points and represented by a coarse histogram of their relative normalized coordinates in the polar system. The descriptor is robust to scale, translation, and rotation. However, its descriptive power decreases for articulated shapes. The inner distance shape context (IDSC) [19], defined as the length of the shortest path between landmark points within the shape silhouette, was proposed to improve the original SC. Its improvement originates from the fact that the Euclidean distance used in SC is replaced by the inner distance, which is robust to articulations. Although this method improved the original SC method, a drawback remains in both methods—i.e. the choice of the number of landmark points. A small number would lead to a poor representation of the shape, whereas a large number would increase the complexity of the method.