New Approaches in Intelligent Image Analysis

The contemporary forms of image representation vary depending on the application. There are well-known mathematical methods for image representation, which comprise: matrices, vectors, determined orthogonal transforms, multi-resolution pyramids, Principal Component Analysis (PCA) and Independent Component Analysis (ICA), Singular Value Decomposition (SVD), wavelet sub-band decompositions, hierarchical tensor transformations, nonlinear decompositions through hierarchical neural networks, polynomial and multiscale hierarchical decompositions, multidimensional tree-like structures, multi-layer perceptual and cognitive models, statistical models, etc. In this chapter are analyzed the basic methods for hierarchical decomposition of grayscale and color images, and of sequences of correlated images of the kind: medical, multispectral, multi-view, etc. Here is also added one expansion and generalization of the ideas of the authors from their previous publications, regarding the possibilities for the development of new, efficient algorithms for hierarchical image decompositions with various purposes. In this chapter are presented and analyzed the following four new approaches for hierarchical image decomposition: the Branched Inverse Difference Pyramid (BIDP), based on the Inverse Difference Pyramid (IDP); the Hierarchical Singular Value Decomposition (HSVD) with tree-like computational structure; the Hierarchical Adaptive Principle Component Analysis (HAPCA) for groups of correlated images; and the Hierarchical Adaptive Kernel Principal Component Analysis (HAKPCA) for color images. In the chapter are given the algorithms, used for the implementation of these decompositions, and their computational complexity is evaluated. Some experimental results, related to selected applications are also given, and various possibilities for the creation of new hybrid algorithms for hierarchical decomposition of multidimensional images are specified. On the basis of the results obtained from the executed analysis, the basic application areas for efficient image processing are specified, such as: reduction of the information surplus; noise filtration; color segmentation; image retrieval; image fusion; dimensionality reduction for objects classification; search enhancement in large scale image databases, etc.

Intelligent systems comprise a large variety of applications, including ones based on signal processing. This field benefits from considerable popularity, especially with recent advances in artificial intelligence, improving existing processing methods and providing robust and scalable solutions to existing and new problems. This chapter builds on well-known signal processing techniques, such as the short-time Fourier and wavelet transform, and introduces the concept of instantaneous frequency along with implementation details. Applications featuring the presented methods are discussed in an attempt to show how intelligent systems and signal processing can work together. Examples that highlight the cooperation between signal analysis and fuzzy c-means clustering, neural networks and support vector machines are being presented.

The chapter summarizes the proposed recently approach for multidimensional data clustering and visualization. It uses a special kind of recurrent networks called Echo state networks (ESN) to generate multiple two-dimensional (2D) projections of the multidimensional original data. For this purpose equilibrium states of all neurons in the ESN are exploited. In order to fit the neurons equilibriums to the data an algorithm for tuning internal weights of the ESN called Intrinsic Plasticity (IP) is applied. Next 2D projections are subjected to selection based on different criteria in dependence on the aim of particular clustering task to be solved. The selected projections are used to cluster and/or to visualize the original data set. Several examples demonstrate possible ways to apply the proposed approach to variety of multidimensional data sets, namely: steel alloys discrimination by their composition; Earth cover classification from hyper spectral satellite images; working regimes classification of an industrial plant using data from multiple measurements; discrimination of patterns of random dot motion on the screen; and clustering and visualization of static and dynamic “sound pictures” taken by multiple randomly placed microphones.

Remote sensing image analysis has been a topic of ongoing research for many years and has led to paradigm shifts in the areas of resource management and global biophysical monitoring. Due to distortions caused by variations in signal/image capture and environmental changes, there is not a definite model for image processing tasks in remote sensing and such tasks are traditionally approached on a case-by-case basis. Intelligent control, however, can streamline some of the case-by-case scenarios and allow for faster, more accurate image processing to aid in more accurate remote sensing image analysis. This chapter will provide an evolutionary control system via two Darwinian particle swarm optimizations—one a novel application of DPSO—coupled with remote sensing image processing to help in the analysis of image data.

MR brain image sequences are characterized by a specific structure and intra- and inter-image correlation but most of the existing histogram segmentation methods do not consider them. We address this issue by proposing a method for tissue segmentation using 2D histogram matching (TS-2DHM). Our 2D histogram is produced from a sum of co-occurrence matrices of each MR image. Two types of model 2D histograms are constructed: an intra-tissue 2D histogram for separate tissue regions and an inter-tissue edge 2D histogram. Firstly, we divide a MR image sequence into a few subsequences using wave hedges distance between 2D histograms of consecutive MR images. Then we save and clear out inter-tissue edge entries in each test 2D histogram, match the test 2D histogram segments in a percentile interval and extract the most representative entries for each tissue, which are used for kNN classification after distance learning. We apply the matching using LUT and two ways of distance metric learning: LMNN and NCA. Finally, segmentation of the test MR image is performed using back projection with majority vote between the probability maps of each tissue region, where the inter-tissue edge entries are added with equal weights to the corresponding main tissues. The proposed algorithm has been evaluated with IBSR 18 and 20, and BrainWeb data sets and showed results comparable with state-of-the-art segmentation algorithms, although it does not consider specific shape and ridges of brain tissues. Its benefits are modest execution time, robustness to outliers and adaptation to different 2D histogram distributions.

In the chapter is presented a multistage approach for segmentation of medical objects in Computed Tomography (CT) images. Noise reduction with consecutive applied median filter and wavelet shrinkage packet decomposition, and contrast enhancement based on Contrast limited adaptive histogram equalization (CLAHE) are applied in preprocessing stage. As a next step is used a combination of 2 basic methods for image segmentation such as split and merge algorithm, following by color based K-mean clustering. For refining the boundaries of the detected objects additional texture analysis is introduced based on limited Haralick’s feature set and morphological filters. Due to the diminished number of components for the feature vectors the speed of the segmentation stage is higher in comparison with the full feature set. Some experimental results are presented, obtained by computer simulation in the MATLAB environment. The experimental results give detailed information about detected simple renal cysts and their boundaries in axial plane of CT images which are presented in native, arterial and venous phases. The proposed approach can be used in real time for precise diagnosis or in monitoring the disease progression.

The mobile robots are equipped with sensitive audio visual sensors, usually microphone arrays and video cameras. They are the main sources of audio visual information to perform suitable mobile robots navigation tasks, modeling the human audio visual perception. The results from the audio and visual perception algorithms are widely used, separate or in conjunction (audio visual perception) in the mobile robots navigation, for example to control mobile robots motion in applications like people and objects tracking, surveillance systems, etc. The effectiveness and precision of the audio visual perception methods in the mobile robots navigation can be enhanced combining audio visual perception with audio visual attention. Sufficient relative knowledge exists, describing the phenomena of human audio and visual attention. Such approaches are usually based on a lot of physiological, psychological, medical and technical experimental investigations relating the human audio and visual attention, with the human audio and visual perception with the leading role of the brain activity. Of course, the results from these investigations are very important, but not sufficient for the mobile robots audio visual attention modeling, mainly because of brain missing in mobile robots audio visual perception systems. Therefore, in this chapter is proposed to use the existing definitions and models for human audio and visual attention, adapting them to the models of mobile robots audio and visual attention and combining with the results from the mobile robots audio and visual perception in the mobile robots navigation tasks.

Three methods for two-dimensional local adaptive image processing are presented in this chapter. In the first one, the adaptation is based on the local information from the four neighborhood pixels of the processed image and the interpolation type is changed to zero or bilinear. An analysis of local characteristics of images in small areas is presented from which the optimal selection of thresholds for dividing into homogeneous and contour blocks is made and the interpolation type is changed adaptively. In the second one, the adaptive image halftoning is based on the generalized two-dimensional LMS error-diffusion filter for image quantization. The thresholds for comparing of input image levels are calculated from the gray values dividing the normalized histogram of the input halftone image into equal parts. The third one—the adaptive line prediction is based on two-dimensional LMS adaptation of coefficients of the linear prediction filter for image coding. An analysis of properties of 2D LMS filters in different directions was made. As a result of the performed mathematical description in the presented methods, three algorithms for local adaptive image processing was developed. The principal block schemes of the developed algorithms are presented. An evaluation of the quality of the processed images was made on the base of the calculated PSNR, SNR, MSE and the subjective observation. The given experimental results from the simulation in MATLAB environment for each of the developed algorithms, suggest that the effective use of local information contributes to minimize the processing error. The methods are extremely suitable for different types of images (for example: fingerprints, contour images, cartoons, medical signals, etc.). The developed algorithms have low computational complexity and are suitable for real-time applications.

Access control is a set of regulations that governs access to certain areas or information. By access we mean entering a specific area, or logging on a machine. The access regulated by a set of rules that specifies who is allowed to get access and what is the restrictions on such access. Across the years several access control systems have been developed. Due to the rapid advancement in technology over the past years, older systems are now easily by passed, thus the need to have new methods of access control. Biometrics is referred to as an authentication technique that relies on a computer system to electronically validate a measurable biological characteristic that is physically unique and cannot be duplicated. Biometrics has been used for ages as access control security system. In this chapter we will present several biometric techniques their usage, advantages and disadvantages.

The work is devoted to the generation of acoustic maps. The experimental work considers the possibility to increase the resolution of the maps. The work uses two-dimensional microphone array with randomly spaced elements to generate acoustic maps of sources located in its near-field region. In this region, the wavefront is not flat and phase of input signals depends on the direction of arrival and the range as well. The input signals are partially distorted by indoor multipath propagation and related interference of sources emissions. For acoustic mapping with the improved resolution, an algorithm in the frequency domain is proposed. The algorithm is based on the modified method of Capon. Acoustic maps of point-like noise sources are generated. The maps are compared with the maps generated using other standard methods including built-in equipment software. The resolution improvement is up to 2.7 times. The obtained results are valuable in the estimation of the direction of arrival for Noise Exposure Monitoring.