Intelligent Systems Reference Library 108 Roumen Kountchev Kazumi Nakamatsu Editors New Approaches in Intelligent Image Analysis Techniques, Methodologies and Applications Intelligent Systems Reference Library Volume 108 Series editors Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland e-mail: [email protected] Lakhmi C. Jain, Bournemouth University, Fern Barrow, Poole, UK, and University of Canberra, Canberra, Australia e-mail: [email protected] About this Series The aim of this series is to publish a Reference Library, including novel advances and developments in all aspects of Intelligent Systems in an easily accessible and well structured form. The series includes reference works, handbooks, compendia, textbooks,well-structuredmonographs,dictionaries,andencyclopedias.Itcontains well integrated knowledge and current information in the field of Intelligent Sys- tems. The series covers the theory, applications, and design methods of Intelligent Systems. Virtually all disciplines such as engineering, computer science, avionics, business, e-commerce, environment, healthcare, physics and life science are included. More information about this series at http://www.springer.com/series/8578 Roumen Kountchev Kazumi Nakamatsu (cid:129) Editors New Approaches in Intelligent Image Analysis Techniques, Methodologies and Applications 123 Editors RoumenKountchev Kazumi Nakamatsu Department ofRadio Communications and Schoolof Human ScienceandEnvironment VideoTechnologies University of Hyogo Technical University of Sofia Himeji Sofia Japan Bulgaria ISSN 1868-4394 ISSN 1868-4408 (electronic) Intelligent Systems Reference Library ISBN978-3-319-32190-5 ISBN978-3-319-32192-9 (eBook) DOI 10.1007/978-3-319-32192-9 LibraryofCongressControlNumber:2016936421 ©SpringerInternationalPublishingSwitzerland2016 Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpart of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilarmethodologynowknownorhereafterdeveloped. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt fromtherelevantprotectivelawsandregulationsandthereforefreeforgeneraluse. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained hereinorforanyerrorsoromissionsthatmayhavebeenmade. Printedonacid-freepaper ThisSpringerimprintispublishedbySpringerNature TheregisteredcompanyisSpringerInternationalPublishingAGSwitzerland Preface Thisbookrepresentstheadvancesinthedevelopmentofnewapproaches,usedfor the intelligent image analysis. It introduces various aspects of the image analysis, related to the theory for their processing, and to some practical applications. The book comprises 11 chapters, whose authors are researchers from different countries: USA, Russia, Bulgaria, Japan, Brazil, Romania, Ukraine, and Egypt. Each chapter is a small monograph, which represents the recent research work oftheauthorsinthecorrespondingscientificarea.Theobjectoftheinvestigationis new methods, algorithms, and models, aimed at the intelligent analysis of signals andimages—singleandsequencesofvariouskinds:natural,medical,multispectral, multi-view, sound pictures, acoustic maps of sources, etc. New Approaches for Hierarchical Image Decomposition, Based on IDP, SVD, PCA, and KPCA InChap.1thebasicmethodsforhierarchicaldecompositionofgrayscaleandcolor images, and of sequences of correlated images are analyzed. New approaches are introduced for hierarchical image decomposition: the Branched Inverse Difference Pyramid(BIDP)andtheHierarchicalSingularValueDecomposition(HSVD)with tree-like computational structure for single images; 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 the evaluation of the computational complexity of the algorithms used for theimplementation ofthese decompositionsisalso given. The basic application areas are defined for efficient image hierarchical decomposition, such as visual information redundancy reduction; noise filtration; color segmenta- tion; image retrieval; image fusion; dimensionality reduction, where the following is executed: the objects classification; search enhancement in large-scale image databases, etc. v vi Preface Intelligent Digital Signal Processing and Feature Extraction Methods The goal of Chap. 2 is to present well-known signal processing methods and the way they can be combined with intelligent systems in order to create powerful feature extraction techniques. In order to achieve this, several case studies are presented to illustrate the power of hybrid systems. The main emphasis is on the instantaneous time–frequency analysis, since it is proven to be a powerful method inseveraltechnicalandscientificareas.Theoldestandmostutilizedmethodisthe Fourier transform, which has been applied in several domains of data processing, but it has very strong limitations due to the constraints it imposes on the analyzed data.Thentheshort-timeFouriertransformandthewavelettransformarepresented astheyprovidebothtemporalandfrequencyinformationasopposedtotheFourier transform. These methods form the basis of most applications, as they offer the possibility of time–frequency analysis of signals. The Hilbert–Huang transform is presentedasanovelsignalprocessingmethod,whichintroducestheconceptofthe instantaneous frequency that can be determined for every time point, making it possible to have a deeper look into different phenomena. Several applications are presented where fuzzy classifiers, support vector machines, and artificial neural networks are used for decision-making. Interconnecting these intelligent methods with signal processing will result in hybrid intelligent systems capable of solving computationally difficult problems. Multi-dimensional Data Clustering and Visualization via Echo State Networks Chapter 3 summarizes the proposed recently approach for multidimensional data clusteringandvisualization.ItusesaspecialkindofrecurrentnetworkscalledEcho StateNetworks(ESN)togeneratemultiple2Dprojectionsofthemultidimensional original data. The 2D projections are subjected to selection based on different criteriadependingontheaimofparticularclusteringtasktobesolved.Theselected projections are used to cluster and/or to visualize the original data set. Several examplesdemonstratethepossiblewaystoapplytheproposedapproachtovariety of multidimensional data sets: steel alloys discrimination by their composition; Earth cover classification from hyperspectral satellite images; working regimes classification of an industrial plant using data from multiple measurements; dis- crimination of patterns of random dot motion on the screen; and clustering and visualization of static and dynamic “sound pictures” by multiple randomly placed microphones. Preface vii Unsupervised Clustering of Natural Images in Automatic Image Annotation Systems Chapter4isdevotedtoautomaticannotationofnaturalimagesjoiningthestrengths of the text-based and the content-based image retrieval. The automatic image annotation is based on the semantic concept models, which are built from large number of patches received from a set of images. In this case, image retrieval is implemented by keywords called Visual Words (VWs) that is similar to text doc- umentretrieval.Thetaskinvolvestwomainstages:alow-levelsegmentationbased on color, texture, and fractal descriptors and a high-level clustering of received descriptorsintotheseparatedclusterscorrespondingtotheVWsset.Theenhanced region descriptor including color, texture, and fractal features has been proposed. For the VWs generation, the unsupervised clustering is a suitable approach. The EnhancedSelf-OrganizingIncrementalNeuralNetwork(ESOINN)waschosendue to its main benefits as a self-organizing structure and online implementation. The preliminary image segmentation permitted to change a sequential order of descriptors entering the ESOINN as associated sets. Such approach simplified, accelerated, and decreased the stochastic variations of the ESOINN. The experi- ments demonstrateacceptableresultsoftheVWsclusteringforanon-large natural image sets. This approach shows better precision values and execution time as compared to the fuzzy c-means algorithm and the classic ESOINN. Also issues of parallel implementation of unsupervised segmentation in OpenMP and Intel Cilk Plus environments were considered for processing of HD-quality images. An Evolutionary Optimization Control System for Remote Sensing Image Processing Chapter 5 provides an evolutionary control system via two Darwinian Particle Swarm Optimizations (DPSO)—one novel application of DPSO—coupled with remote sensing image processing to help in the image data analysis. The remote sensingimageanalysishasbeenatopicofongoingresearchformanyyearsandhas 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 remotesensingandsuchtasksaretraditionallyapproachedonacase-by-casebasis. Intelligentcontrol,however,canstreamlinesomeofthecase-by-casescenariosand allows faster, moreaccurateimageprocessingtosupportthemoreaccurateremote sensing image analysis. viii Preface Tissue Segmentation Methods Using 2D Histogram Matching in a Sequence of MR Brain Images In Chap. 6 a newtransductive learningmethod for tissue segmentation using a 2D histogram modification, applied to Magnetic Resonance (MR) image sequence, is introduced.The2Dhistogramisproducedfromanormalizedsumofco-occurrence matricesofeachMRimage.Twotypesofmodel2Dhistogramsareconstructedfor each subsequence: intra-tissue 2D histogram to separate tissue regions and an inter-tissue edge 2D histogram. First, the MR image sequence is divided into few subsequences, using wave hedges distance between the 2D histograms of the consecutive MR images. The test 2D histogram segments are modified in the confidenceintervalandthemostrepresentativeentriesforeachtissueareextracted, whichare usedforthekNN classification afterdistancelearning. The modification is applied by using LUT and two ways of distance metric learning: large margin nearest neighbor and neighborhood component analysis. 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 corresponding tissues. The proposed algorithmhasbeenevaluatedwithfreeaccessdatasetsandhasshowedresultsthat arecomparabletothestate-of-the-artsegmentationalgorithms,althoughitdoesnot consider specific shape and ridges of brain tissues. Multistage Approach for Simple Kidney Cysts Segmentation in CT Images InChap.7amultistageapproachforsegmentationofmedicalobjectsinComputed Tomography (CT) images is presented. Noise reduction with consecutive applied median filter and wavelet shrinkage packet decomposition, and contrast enhance- ment based on Contrast limited Adaptive Histogram Equalization (CLAHE) are applied in the preprocessing stage. As a next step a combination of two basic methods is used for image segmentation such as the split and merge algorithm, followed by the color-based K-mean clustering. For refining the boundaries of the detected objects, additional texture analysis is introduced based on the 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 than that for the full featureset. Some experimentalresultsare presented, obtained by computer simulation. The experimental results give detailed information about the detected simple renal cysts and their boundaries in the axial plane of the CT images.Theproposedapproachcanbeusedinrealtimeforprecisediagnosisorin disease progression monitoring. Preface ix Audio Visual Attention Models in Mobile Robots Navigation InChap.8,itisproposedtousetheexitingdefinitionsandmodelsforhumanaudio andvisualattention,adaptingthemtothemodelsofmobilerobotsaudioandvisual attention, and combining with the results from mobile robots audio and visual perception in the mobile robots navigation tasks. The mobile robots are equipped withsensitiveaudiovisualsensors(usuallymicrophonearraysandvideocameras). They are the main sources of audio and visual information to perform suitable mobile robots navigation tasks modeling human audio and visual perception. The audio and visual perception algorithms are widely used, separately or in audio visualperception,inmobilerobotnavigation,forexampletocontrolmobilerobots motion in applications like people and objects tracking, surveillance systems, etc. The effectiveness and precision of the audio and visual perception methods in mobile robots navigation can be enhanced combining audio and visual perception with audio and visual attention. There exists relative sufficient knowledge describing the phenomena of human audio and visual attention. Local Adaptive Image Processing Three methods for 2D local adaptive image processing are presented in Chap. 9. In the first one, the adaptation is based on the local information from the four neighborhoodpixelsoftheprocessedimageandtheinterpolationtypeischangedto zeroorbilinear.Theanalysisofthelocalcharacteristicsofimagesinsmallareasis 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 gen- eralized 2D Last Mean Square (LMS) error-diffusion filter for image quantization. The thresholds for comparing the input image levels are calculated from the gray values dividing the normalized histogram of the input halftone image into equal parts. In the third one, the adaptive line prediction is based on the 2D 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. 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 objective criteria and the subjective observation. The given experimental results, fromthesimulationforeachofthedevelopedalgorithms,suggestthattheeffective useoflocalinformationcontributestominimizetheprocessingerror.Themethods are suitable for different types of images (fingerprints, contour images, cartoons, medical signals, etc.). The developed algorithms have low computational com- plexity and are suitable for real-time applications.