Contributed articleInhibitory connections in the assembly neural network for texture segmentation
Introduction
The task of texture segmentation of natural images is important for solving the problems of object–ground separation, recognition of object shapes, and analysis of natural scenes. The solution to this task is important for robotics and for automation of visual information processing in many different areas, e.g. in medicine, mapping, etc.
Most of the studies of texture recognition and segmentation use heuristic algorithms (Nothdurft, 1985; Voorhees & Poggio, 1988; Bovik et al., 1990; Wu & Chen, 1993; You & Cohen, 1993; Caelli & Reye, 1993; Hild & Shirai, 1993; Shen et al., 1993), but it is the neural network that performs texture recognition and segmentation in the brain of man and animals. Following Hebb's theory, intellectual functions of the mammalian brain are based on the structural and functional organization of neural networks in the form of neural assemblies (Hebb, 1949; Milner, 1957). In neural network-based models neural networks mainly perform classification of textures and nonlinear transforffiaiion of filtered input data (Oje, 1989;Malik & Perona, 1990; Kussul et al., 1991;Van Hulle & Tollenaere, 1993). Inhibitory connections are seldom used in neural models for texture recognition and segmentation. However, it is known that inhibitory connections take part in information processing of the brain (e.g. Eccles, 1969; Hubel, 1988). In (Goltsev, 1996) a neural network with assembly organization is described; the network is intended for texture segmentation of natural images. The present article continues this study. Inhibitory connections are formed in the assembly network, increasing the recognition rate and reliability of the recognition process of the assembly network.
Section snippets
Description of the assembly network
The model is designed to solve the following task. Consider a natural image containing several texture regions, each of which constitutes a texture class. In each region, a comparatively small number of image patches (training samples of this texture class) are chosen by the teacher for the network's learning. Image samples of the training set of each texture class are distributed among all typical positions within the corresponding region. The problem is to identify the membership of test set
Formation of the excitatory connections
The network consists of N neurons, numbered from 1 to N. The neuron numbers are denoted by indices i and j in the formulae presented later. For formal description of modification of the assembly network structure, the input and output binary vectors are introduced to represent network neurons by its one-valued components. At the intersections of corresponding lines and columns of connection matrixes, the weights of connections may change as a result of training.
Let there exist U texture
Description of the network dynamics
The activity of each neuron is calculated at every time step, t, synchronously with all neurons of the network. The output of each neuron has only two values: 0 or 1. A long binary vector of neural activity S(x) is used for description of activity dynamics of the network at the stage of texture recognition of the xth image patch; it represents outputs of all network neurons. Let us denote the output of the ith neuron at the tth time step by .
At the zeroth time step of a recognition process
Formation of negative features
During training the excitatory connections of some texture class, a cumulation of the feature sets extracted from all training samples of this texture class is collected by the model as well. For this purpose, at every training patch xm, of the mth texture class, the model performs a disjunction of the vector G(xm) and a special short binary vector Wm (with all zero-valued components at the beginning of the process). Let the training set of the mth texture class consists of M patches. During
Training the inhibitory connections
In order to train the inhibitory connections directed from the negative features formed in vectors H, the model successively passes all image patches of the training set again. The inhibitory connections, from each training patch's texture class subnetwork, to another subnetwork, are trained by means of a special procedure. This procedure, described later, is performed at each texture patch of the training set. As a result of the inhibitory connections' training process only those neurons which
Restrictions on the negative features
Let us consider the restrictions in using the negative features. The task to the network implies that the training set for each texture class must cover all typical texture samples of this class. The reason for this requirement is evident. The network ought to be trained for all feature values and their combinations which may be met in the texture class. First, this means that the inner structure of assemblies, which consists of the excitatory connections, ought to be formed based on the
Results of the computer simulation
A simulation program for the network described above has been created and used to process photographs digitized to 192×160 pixels with 32 grey levels. Fig. 5(a)Fig. 6(a)Fig. 7(a)Fig. 8(a) show samples of the original images [the same as in (Goltsev, 1996)] used in the experiments. The model was intended to recognize four texture classes, therefore, the network consisted of four subnetworks. There were a total 2048 neurons in the network. One full-connected connection matrix was used, containing
Discussion
There is a necessity to clarify some points relating to the assembly network. The recurrent, dynamic algorithm of the assembly network convergence, which is described above, makes this assembly network a dynamic and nonlinear device. In the present model the network convergence is performed by means of an associative process of neural activity re-distribution between the assemblies. This algorithm demonstrates associative capacities of the assembly network. Associative properties of neural
Conclusion
In conclusion, the following should be remarked. The inhibitory connections are useful in those cases when it is necessary to have high reliability of recognition of a part or all recognized classes, and the fully representative training set is available. The inhibitory connections are a tool, by means of which it is possible to prevent completely the recognition errors in a limited number of classes if, of course, any negative features exist in these classes.
The inhibitory connections do not
References (26)
- et al.
Statistical neurodynamics of associative memory
Neural Networks
(1988) - et al.
On the classification of image regions by color, texture and shape
Pattern Recognition
(1993) Brightness image segmentation by a neuron-like network
Neurocomputing
(1992)An assembly neural network for texture segmentation
Neural Networks
(1996)- et al.
Extraction of texture elements from images of shaded scenes
Pattern Recognition
(1993) Sensitivity for structure gradient for texture discrimination tasks
Vision Research
(1985)- et al.
A texture based distance measure for classification
Pattern Recognition
(1993) - et al.
A modular artificial neural network for texture processing
Neural Networks
(1993) - et al.
Multi-threshold dimension vector for texture analysis and its application to liver tissue classification
Pattern Recognition
(1993) - et al.
Classification and segmentation of rotated and scaled textured images using texture `turned' masks
Pattern Recognition
(1993)
Multichannel texture analysis using localized spatial filters
IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-
Cited by (9)
Investigation of efficient features for image recognition by neural networks
2012, Neural NetworksModular neural networks with Hebbian learning rule
2009, NeurocomputingCitation Excerpt :A modular neural network of several modules may have connections not only inside the modules, but between neurons of different modules as well. Such a connection structure is considered for the case of only inhibitory connections in [15]. The subsequent part of the paper describes a modular neural network with modifiable inhibitory as well as excitatory connections between neurons of different modules named as NN2.
Secondary learning in the assembly neural network
2004, NeurocomputingCitation Excerpt :The simplification of the recognition algorithm makes the present assembly network much faster. The present assembly neural network does not need any regulator of neural activity such as Milner's structure of internal inhibitory connections [33,34], Braitenberg's central extra-network regulator [6], and the system of excitation-inhibition [3,4,26,17,19]. This circumstance is very important.
Neural networks and micromechanics
2010, Neural Networks and MicromechanicsTexture recognition with random subspace neural classifier
2005, WSEAS Transactions on Circuits and Systems