Neural Classifiers - Neural Networks and Micromechanics - page 26

Fig. 3.1 Samples of real-world images
system (74 features). We used associative-projective neural networks for recogni-
tion [ 4 ], achieving a recognition rate of 79.9 %. In Fig. 3.1 two examples of our
images are presented.
In 1996, A. Goltsev developed an assembly neural network for texture segmen-
tation [ 5 , 6 ] and used it for real scene analysis. Texture recognition algorithms are
used in different areas, for example, in the textile industry for detection of fabric
defects [ 7 ]. In the electronic industry, texture recognition is important to character-
ize the microstructure of metal films deposited on flat substrates [ 8 ], and to
automate the visual inspection of magnetic disks for quality control [ 9 ]. Texture
recognition is used for foreign object detection (for example, contaminants in food,
such as pieces of stone, fragments of glass, etc.) [ 10 ]. Aerial texture classification is
applied to resolve difficult figure-ground separation problems [ 11 ].
Different approaches were developed to solve the problem of texture recogni-
tion. Leung et al. [ 12 ] proposed textons (representative texture elements) for texture
description and recognition. The vocabulary of textons corresponds to the charac-
teristic features of the image. They tested this algorithm on the Columbia-Utrecht
Reflectance and Textures (CUReT) database [ 13 , 14 ]. This approach has disadvan-
tages: it needs many parameters to be set manually and it is computationally
complex. Another approach is based on the micro-textons, which are extracted by
means of a multiresolution local binary pattern operator (LBP). LBP is a grayscale
invariant primitive statistic of texture. This method was tested on CUReT database
and performed well in both experiments and analysis of outdoor scene images.
Many statistical texture descriptors were based on a generation of co-occurrence
matrices. In [ 9 ] the texture co-occurrence of n th order rank was proposed. This
matrix contains statistics of the pixel under investigation and its surrounding pixels.
The co-occurrence operator can be used to map the binary image, too. For example,
in [ 15 ], the method to extract texture features in terms of the occurrence of n conjoint
pixel values was combined with a single-layer neural network. There are many
investigations in the application of neural networks for texture recognition [ 16 , 17 ].
To test the developed system [ 17 ], texture images from [ 18 ] were used.
The reasons to choose a system based on neural network architecture include its
significant properties of adaptiveness and robustness to texture variety.