مقاله انگلیسی رایگان در مورد نشانه مبتنی بر شبکه عصبی تصادفی – الزویر 2019

Dynamic texture analysis has been the focus of intensive research in recent years. Thus, this paper presents an innovative and highly discriminative dynamic texture analysis method, whose signature is composed of the weights of the output layer of a randomized neural network after a training procedure. This training is performed by using the pixels of slices of each orthogonal plane of the video (XY, YT, and XT) as input feature vectors and corresponding output labels. The obtained video signature provided an accuracy of 97.05%, 98.54%, 97.74% and 96.51% on the UCLA-50 classes, UCLA-9 classes, UCLA-8 classes and Dyntex++, respectively. These results, when compared to other dynamic texture analysis methods, demonstrate that our descriptors are very effective and that our proposed approach can contribute significantly to the field of dynamic texture analysis.

Introduction

Dynamic texture analysis is an important research area of computer vision responsible for extracting meaningful characteristics from dynamic texture videos. This field has gained much attention due to the range of applications, such as monitoring of traffic in highway (Chan & Vasconcelos, 2005; Derpanis & Wildes, 2011), human activity recognition (Kellokumpu, Zhao, & Pietikäinen, 2008), facial expression recognition (Zhao & Pietikainen, 2007), medical videos analysis (Brieu et al., 2010), crowd analysis and management (Chan, Morrow, & Vasconcelos, 2009), among others. Although the understanding and perception of dynamic textures are easy to humans, their formal definition and description using computational methods are a hard task (Gonçalves & Bruno, 2013a). Unlike traditional texture images, dynamic textures are sequences of images with texture patterns that represent a dynamic object or process and present certain stationary properties in space and time (Doretto, Chiuso, Wu, & Soatto, 2003). Therefore, dynamic textures can be defined as an extension of traditional texture images to the spatial and temporal domain, which correspond to the appearance and motion characteristics, respectively (Gonçalves & Bruno, 2013b). Examples of dynamic textures are sea waves, boiling water, waterfall, metal corrosion process and fire. The addition of the time domain causes new challenges in the characterization task, since it is necessary to combine appearance and motion information (e.g. some methods analyze textures based on motion only), and to process it with low computational complexity. To overcome this, many approaches have been proposed, each one investigating characteristics of the dynamic texture video in a different way