A survey of image descriptors

Muralikrishna, S N and Muniyal, Balachandra and Acharya, Dinesh U (2017) A survey of image descriptors. In: National Workshop-cum-Conference Computer Applications based on Modern Algebra, 01/07/2017, MIT, Manipal.

[img] PDF
793.pdf - Published Version
Restricted to Registered users only

Download (158kB) | Request a copy

Abstract

In any computer vision task feature extraction is an important step. The robustness of any vision system depends on how well the system is able to extract features and represent them. Tasks such as image retrieval, object detection in images/videos, object recognition, tracking of objects in video frames etc., requires some way of describing/representing the whole image in a simpler form, or describing the local regions in the images or shapes of an object in the image, depending on the application. Ideally, these descriptions must be invariant to illumination changes, invariant to transformations such as rotation, scaling etc. Image descriptors are used for this purpose. Descriptors can be classified into categories as global, local and regional features metrics. Processes that involve the use of features usually are divided into two phases: detection and description. First, we detect where a feature may be and then we describe that region in order to provide enough information about this feature. There are several methods for identifying keypoints such as corners, edges, blobs etc., using detectors. Looking at the literature, one can trace back to the development of descriptors as a part of image processing concepts such as texture and statistics. Also, descriptors are used in machine vision methods as local, regional and global features. Texture description can be done through texels which are a set of micro-texture patterns. A numerical description of edge counts, edge magnitude, edge orientation, lines, corner points, and the histogram of gradients can represent a structural pattern within a region. The statistical description uses co-occurrence matrix, gray-level statistical moments etc. Transform based methods such as Fourier descriptors, Wavelets also exist. Development of scale-space theory and local features surrounding interest points in the image called feature descriptors helped in solving problems such as image matching. In the early 2000s scene object modelling based on feature descriptors such as Scale Invariant Feature Transform (SIFT), Speed-Up Robust Feature (SURF), HAAR etc., with complex classifier were proposed. These systems are more accurate and robust compared to earlier systems. Descriptor-based modelling consists of descriptors of four families of feature description techniques, Local Binary Descriptors, Spectra descriptors, Basis space descriptors, Polygon shape descriptors. However, the choice of descriptor and classifier is mostly dependent on the type of application. There is no single point solution to generic problems. We conclude that this area is promising for future research.

Item Type: Conference or Workshop Item (Paper)
Uncontrolled Keywords: Image Descriptors, SIFT, SURF, HAAR, Scale-Space
Subjects: Engineering > MIT Manipal > Computer Science and Engineering
Engineering > MIT Manipal > Information and Communication Technology
Depositing User: MIT Library
Date Deposited: 07 Feb 2019 05:28
Last Modified: 07 Feb 2019 05:28
URI: http://eprints.manipal.edu/id/eprint/153217

Actions (login required)

View Item View Item