![](https://csdnimg.cn/release/download_crawler_static/11178905/bg11.jpg)
1.1 Research Topics of Multidimensional Night-Vision Information Understanding 3
Image feature representation is key to computer vision, image analysis and artifi-
cial intelligence. A good image representation method can provide a strong expres-
sion mechanism, not only saving computer storage space but also reducing image
processing complexity. Feature representation strongly influences image classifica-
tion and recognition. Unfortunately, current feature representation techniques do not
meet the needs of actual image analysis applications, because many key problems in
feature representation have not been well solved. For example, original image fea-
tures are high dimensional, whereas image content is complex and various. Images
are often affected by light, scale, rotation, translation, occlusion and other noises.
Presently, the research of image feature representation focuses on bottom feature
extraction by artificial design and learning-based feature representation. The fea-
tures extracted by artificial design are used to describe image content and similarity.
Methods include global and local feature representation.
All information in the region of interest and background area are extracted from
the whole image as global features. Global information of image includes colour,
texture, shape, etc. Corresponding feature vectors are formed by processing the basic
information of each feature point. The image is then mapped to a feature vector set.
The most representative global features are the histogram of oriented gradient (HOG)
in Dalal and Triggs (2005), local binary pattern (LBP) in Ojala et al. (2002), and
colour histogram in Rubner et al. (2010). For image classification, global features
make good use of image context information. Thus, feature extraction is very fast.
These advantages make the global feature widely used in classification systems, such
as for scene classification.
Global features focus on whole-image information, ignoring the target object.
Additionally, image representation, based on global features, is usually presented
as statistical data. It is more redundancy, but, under the circumstances, it is too
sensitive to geometric transformation (e.g. rotation and zooming), image brightness
changes, target occlusion, target displacement and deformation. This severely limits
its application to image classification and recognition.
Later, a primitive feature i s introduced to flexibly describe the internal informa-
tion and details in images. This local feature can find invariant parts in the same
class of ever-changing images. Image representations based on local features are
increasingly popular. Thus, many researchers have proposed image representation
methods, including the word-bag mode of Li and Perona (2005) (i.e., bag-of-words),
the soft quantisation algorithm of Philbin et al. (2008) and Liu et al. (2011), the
locally restricted linear coding algorithm of Wang et al. (2010) and the spatial pyra-
mid algorithm of Lazebnik et al. (2006).
Visual data is represented by video and spectral images, characterised by huge
amounts of data, high-resolution information, unstructured data, interpretation vari-
ety, ambiguity and uncertainty. Similarly, image data often possess very high dimen-
sions, again making image representation and artificial design waste time and energy.
The design and extraction process of manual features is very difficult. Thus, it cannot
be directly applied to current high-dimensional image analysis and processing. Thus,
to find a reasonable method of representing image data, the non-artificial design has