Grayscale Image Segmentation With Quantum-Inspired Multilayer Self-Organizing Neural Network Architecture Endorsed by Context Sensitive Thresholding

Introduction

Image processing using image segmentation is a difficult task to recover the objects from multilevel images. So many research approaches have been taken out to reconstruct the object from the multilevel background, but it still remains a fallacy to recover the true objects ideally. The multilayer self organizing neural network (MLSONN) architecture is unable to extract the gray scale objects from the blurred and noisy atmosphere which is designed by Ghosh et al., 1993 (Ghosh et al., 1993) and is used to extract the binary objects efficiently. Here, interconnection weights of different layers viz. between input to hidden layer and hidden to output layer are updated by means of fuzzy measures. This architecture is limited only for the bi-level sigmoid activation function to segment the binary images. The authors (Pal et al., 1993) have described the color image segmentation technique using fuzzy and non-fuzzy methods considering segmentation of range images and neural network based approaches. The authors (Pantofaru et al., 2005) have presented the result of the mean shift segmentation and the efficient graph-based segmentation techniques for objective evaluation algorithm describing the three components viz. the Correctness, the Stability with respect to parameter choice and the Stability with respect to image choice. If these characteristics are fully satisfied by the segmentation technique, then it can be more useful by larger systems. The authors have considered the pixel location and color feature for each and every image for this segmentation algorithm using the Berkeley segmentation database. The authors (Bhattacharyya et al., 2007) have described on the true color image segmentation by self supervised PSONN architecture using multilevel sigmoidal activation function. Regarding this proposed architecture, it is the extension version of standard single self organizing neural network architecture (SONN) and comprises input or source layer, three middle layers for segmentation of three primary color components and the output layer or sink layer. To segment the color image for first object recovery in large image database, the probability of pixel distribution is implemented (Kang et al., 2008). After incorporating three channel images of R, G and B from the given image, and then applying pixel distribution is taken out using similarity measures using the well known defined distribution function Weibull, Exponential, Beta, Gamma, Normal, and Uniform. Using the measurement of sum least of square error, to fit the image to the distribution. Under consideration of minimum amount of error, image is quantized to gray levels for three channels of distribution using threshold value and then these three channel values are fused together to get the desired information. Few years’ latter authors (Bhattacharyya et al., 2010) proposed multilevel image segmentation using a MUSIG activation function which is more efficient to extract the multilevel images by means of functionally modifying the network. The MUSIG activation function is characterized by ignoring heterogeneity of the image information content for understanding equal and fixed class responses. The authors De S. et al. (De S. et al., 2010) described under consideration of the heterogeneity of the image information content in the segmented images by applying optimized MUSIG (OptiMUSIG) activation function. OptiMUSIG activation function is used in another way to segment the true color image segmentation (De S. et al, 2012) on optimized class responses on self organizing neural network architecture. The authors De S. et al. considered generic based optimized segmentation method. Gray scale image is segmented (De S. et al., 2012) on optimized class responses without considering the heterogeneity of image information content by optimization of MUSIG (OptiMUSIG) activation function. It may or may not generate good quality of segmented outputs. There are so many research works have been done to recover the object from the different images. Segmentation is one of the approaches where object is reconstructed from the image. The gray scale image is segmented to recover the object by means of context sensitive thresholding implementation technique. In this chapter, the authors propose a gray scale image segmentation method with a quantum inspired based neural network architecture having multilayer self organization nature with endorsed by context sensitive thresholding approaches, which is more efficient to extract the multilevel gray scale objects. The objective of this chapter is to extend the functional modification of QMLSONN architecture so as to segment the gray scale multilevel images by adapting the context sensitive thresholding using multilevel sigmoidal activation function (MUSIG) activation function as shown in Figure 1. Quantum computer has the ability to segment the gray scale images very efficiently as the extraction time is very less as compared to the classical computation technique. Quantum bit is able to do this job tactically using the quantum superposition principle. In this chapter, authors use the segmentation technique using a context sensitive thresholding concept to improve the image intensity information. In the process of image segmentation, image is subdivided i.e. segmented into different classes corresponding to correlate in different regions according to the required features of interest. Here the pixels of compactness of groups are taken together as a whole for the stipulated attributes of similar properties. Authors (Yogamangalam et al., 2013) have described the different image segmentation techniques, mentioning Markov Random Field (MRF) is the strongest and the simplest method for cancelling the noise to recover the object from the noisy atmospheric image. According to the authors’ overview, the segmented parts are enclosed to cancel the noise more easily from the noisy image. Binary object is extracted from the noisy environment is efficiently done using quantum version of MLSONN architecture, proposed by Bhattacharyya et al. (Bhattacharyya et al., 2014).