Goran Zajic

Adaptive Content-Based Image Retrieval with Relevance Feedback

Retrieval of images, based on similarities between feature vectors of querying image and those from database, is considered. The searching procedure was performed through the two basic steps: an objective one, based on the Euclidean distances and a subjective one based on the users relevance feedback. Images recognized from user as the best matched to a query are labeled and used for updating the query feature vector through a RBF (radial basis function) neural network. The searching process is repeated from such subjectively refined feature vectors. In practice, several iterative steps are sufficient, as confirmed by intensive simulations

Classification of Prolapsed Mitral Valve versus Healthy Heart from Phonocardiograms by Multifractal Analysis

Phonocardiography has shown a great potential for developing low-cost computer-aided diagnosis systems for cardiovascular monitoring. So far, most of the work reported regarding cardiosignal analysis using multifractals is oriented towards heartbeat dynamics. This paper represents a step towards automatic detection of one of the most common pathological syndromes, so-called mitral valve prolapse (MVP), using phonocardiograms and multifractal analysis. Subtle features characteristic for MVP in phonocardiograms may be difficult to detect. The approach for revealing such features should be locally based rather than globally based. Nevertheless, if their appearances are specific and frequent, they can affect a multifractal spectrum. This has been the case in our experiment with the click syndrome. Totally, 117 pediatric phonocardiographic recordings (PCGs), 8 seconds long each, obtained from 117 patients were used for PMV automatic detection. We propose a two-step algorithm to distinguish PCGs that belong to children with healthy hearts and children with prolapsed mitral valves (PMVs). Obtained results show high accuracy of the method. We achieved 96.91% accuracy on the dataset (97 recordings). Additionally, 90% accuracy is achieved for the evaluation dataset (20 recordings). Content of the datasets is confirmed by the echocardiographic screening.

Global Image Search vs. Regional Search in CBIR Systems

The global image search and regional image search are compared by using content-based image retrieval system with user relevance feedback. It was expectable that regional search can minimize the effect of the background to the image retrieval. Images from database are partitioned into regular rectangular regions: 4times4 non-overlapped (NOV) regions and 3times3 overlapped (OV) regions, and a feature vectors are determined for whole images and for regions. Four CBIR scenarios are considered: global search, search based on 4times4 NOV regions, based on 3times3 OV regions and based on arbitrary cropped part of a query image. System is tested over images from Corel IK dataset.

Comparison of CBIR Systems with Different Number of Feature Vector Components

Content-based image retrieval (CBIR) systems with user relevance feedback are considered. The influence of the type and the number of feature vector (FV) components on the retrieval efficiency was investigated. We compared a CBIR system with a very small number of FV components (only 25 components describing color and texture) with a system with a high-dimensional FV inspired by MPEG-7 (556 coordinates describing color, texture and line directions), as well as with a system using feature vector reduction (FVR) of about 90% (with about 50 FV components from the full-length 556-component FVs). The systems are tested over the annotated Corel IK and Corel 60K datasets. Simulation results show that a decreased number of FV components does not have significant influence on the quality of image retrieval, while the processing time is reduced compared to CBIR with full-length FV and/or FVR.A consolidated view of the content within different data repositories would facilitate useful operations such as advanced informational discovery. This paper proposes a methodology and an architecture that uses semantic annotations to enable such holistic views across domain specific information sources. The potential for the personalization of this annotated information is described, and the future work necessary to implement the system is elaborated on.

Minor Component Analysis (MCA) Applied to Image Classification in CBIR Systems

A content-based image retrieval system with query image classification prior to retrieving procedure is proposed. Query image is compared to representative patterns of image classes, not to all images from database, accelerating thus initial retrieving step. Such procedure is possible when images from database are grouped into classes with similar content. Classification is performed using minor component (MC) analysis. Since it is expectable that MCs mainly depend on image details, not on an image background, this approach seems to be more efficient than classic CBIR. Minor components may be calculated by using single-layer neural network. The efficiency of proposed system is tested over images from Corel dataset

Second generation wavelets: Advantages in cardiosignal processing

In this paper, the cardiosignal analysis using second generation wavelets and lifting structure based on clinically-relevant extracted information is described. The information extraction can be performed by different spectral (such as multifractal) and spectrogram (time-frequency) representations. Here, the novelties are related mainly to the knowledge-controlled analysis and noise reduction in phonocardiograms. We suggest adaptive (signal-dependent) wavelet-based processing able to simultaneously preserve relevant high-frequency details and reduce noise.

The influence of the feature vector content on the CBIR system efficiency

The influence of the feature vector (FV) content on the CBIR (content-based image retrieval) system efficiency was considered. By using two different FVs and applying three different learning methods, it was shown that the efficiency of retrieving depends on both the FV content and the learning method, independently.

The use of unlabeled data in image retrieval with relevance feedback

This paper describes a content-based image retrieval (CBIR) system which makes use of both labeled images, annotated by the user, and unlabeled images available in the database. The system initially retrieves images objectively closest to the query image. The user then subjectively labels retrieved images as relevant or irrelevant. Although such relevance feedback from the user is an effective way of bridging the semantic gap between objective and subjective similarity, it is also very time consuming, requiring huge human effort. Often, the number of labeled images is very small. In an inductive approach the labeled set of images is used for training a CBIR system while the large set of unlabeled images remains unused. In this paper we exploit the transductive support vector machine (SVM) algorithm as a way of taking advantage of unlabeled data in CBIR. Our findings are compared to the results of an inductive SVM. We draw some conclusions as to when the use of unlabeled data might be helpful. The considered systems are tested over images from the Corel 1K dataset.

Searching image database based on content

The modern age is characterized by great professional and private multi-media production of which the largest percentage are images. Search a large number of images from the users perspective is almost always based on content. The aim of this work is to create a database search algorithm for images that is content based. Starting from the features matrix, algorithm is based on working with color, texture and shape. Using the five characteristics of color, four characteristics of texture, and one feature for shape, the initial group of images, by applying different levels of ranking and reduction, are reduced to a small number of representative images in relation to query image. Using a neural network model and the group obtained images, the result obtained by artificial intelligence is shown to the user. Interaction with the user, in every iteration, search result can be further filtered or modified in accordance with selected images.

Shot-change detection based on multifractal analysis

This paper presents a new algorithm for detection of abrupt shot changes in video sequence based on Multifractal Analysis (MA). Video sequence is decoded into series of frames, and low level feature for texture and color are extracted from each of them. Content difference between adjacent frames is calculated based on correlation of frame low level features. MA is used for description of content difference arrays in multifractal singularity domain (α - domain) with emphasized local singularities. The detection of abrupt changes is realized using fixed thresholds technique and singularity shape analysis. Proposed algorithm is tested on more than 600 000 frames, with high efficiency of detection.