Papers by Jorma Laaksonen
We have developed an experimental system called PicSOM for retrieving images similar to a given s... more We have developed an experimental system called PicSOM for retrieving images similar to a given set of reference images in large unannotated image databases. The technique is based on a hierarchical variant of the Self-Organizing Map (SOM) called the Tree Structured Self-Organizing Map (TS-SOM). Given a set of reference images, PicSOM is able to retrieve another set of images which
Analyzing Low-Level Visual Features Using Content-Based Image Retrieval
International Conference on Neural Information Processing, 2000
This paper describes how low-level statistical vi- sual features can be analyzed in our content-b... more This paper describes how low-level statistical vi- sual features can be analyzed in our content-based image retrieval system named PicSOM. The low- level visual features used in the system are all sta- tistical by nature. They include average color, color moments, contrast-type textural feature, and edge histogram and Fourier transform based shape fea- tures. Other features can be added easily.
Using Long-Term Learning to Improve Efficiency of Content-Based Image Retrieval
Pattern Recognition in Information Systems, 2003
Content-based image retrieval (CBIR) is an emerging re- search field, studying retrieval of image... more Content-based image retrieval (CBIR) is an emerging re- search field, studying retrieval of images from unannotated databases. In CBIR, images are indexed on the basis of low-level statistical fea- tures that can be automatically derived from the images. Due to the gap between high-level semantic concepts and low-level visual features, the performance of CBIR applications often remains quite modest. One
Results on a comparison of adaptive recognition techniques for on-line recognition of handwritten... more Results on a comparison of adaptive recognition techniques for on-line recognition of handwritten Latin alphabets are presented. The emphasis is on ve adaptive classi cation strategies described in this paper. The strategies are based on rst generating a user-independent set of prototype characters and then modifying this set in order to adapt it to each user's personal writing style. The initial set is formed by a simple clustering algorithm. The modi cation of the prototype set is performed using three modes of operation: 1) new prototypes are added, 2) existing prototypes are reshaped to better match the input, and 3) prototypes which produce false classi cations are removed. The classi cation decision uses the k-Nearest Neighbor (k-NN) rule for the distances between the unknown character and the stored prototypes. The distances are calculated by using template matching with Dynamic Time Warping (DTW). The reshaping of the existing prototypes is performed by utilizing a modi ed version of the Learning Vector Quantization (LVQ) algorithm. The presented experiments show that the recognition system is able to adapt well to the user's writing style with only a few { say one hundred { handwritten characters.
PicSOM A Framework for Content-Based Image Database Retrieval using Self-Organizing Maps
Scandinavian Conference on Image Analysis, 1999
We have developed an image retrieval systemwhich uses Tree Structured Self-Organizing Maps(TS-SOM... more We have developed an image retrieval systemwhich uses Tree Structured Self-Organizing Maps(TS-SOMs) as the method for retrieving imagessimilar to a given set of reference images in adatabase. It also provides a framework for theresearch on algorithms and methods for contentbasedretrieval of images. A novel technique introducedin this paper facilitates automatic combinationof the responses from multiple TS-SOMs andtheir hierarchical levels. The
Int. Conference on Artificial Neural Networks, 1998
Results on a comparison of adaptive recognition techniques for on-linerecognition of handwritten ... more Results on a comparison of adaptive recognition techniques for on-linerecognition of handwritten Latin alphabets are presented. The classi?-cation strategies compared are based on ?rst compressing or distilling alarge database of handwritten characters to a small set of character prototypes.Each adaptive classi?er then either modi?es the original prototypesor conditionally adds new prototypes when they become availablefrom the user of the system.
Learning Vector Quantization (LVQ) is a group of algorithms applicable to statistical pattern rec... more Learning Vector Quantization (LVQ) is a group of algorithms applicable to statistical pattern recognition, in which the classes are described by a relatively small number of codebook vectors, properly placed within each zone such that the decision borders are approximated by the nearest-neighbor rule. The LVQ PAK program package contains all programs necessary for the correct application of certain Learning Vector Quantization algorithms in an arbitrary statistical classi cation or pattern recognition task, as well as a program for the monitoring of the codebook vectors at any time during the learning process. The rst version 1.0 of this program package was published in 1991 and since then the package has been updated regularly to include latest improvements in the LVQ implementations. This report that contains the last documentation was prepared for bibliographical purposes.
Lecture Notes in Computer Science, 2003
Content-based image retrieval (CBIR) addresses the problem of assisting a user to retrieve images... more Content-based image retrieval (CBIR) addresses the problem of assisting a user to retrieve images from unannotated databases, based on features that can be automatically derived from the images. Today, there exists several CBIR systems based on different methods. Only few attemps to benchmark these have been made, although the usefulness of benchmarking is undeniable in the development of different algorithms. In this paper we publish our benchmarking results of two CBIR systems with different implementation methods. The CBIR systems in question are GIFT (University of Geneva) and PicSOM (Helsinki University of Technology). The results clearly show that our PicSOM system, which we earlier have not been able to benchmark against other CBIR systems, comes off well in the comparison. Also, the results indicate that tests based on a single ground truth class are not enough for fair system comparisons.
Lecture Notes in Computer Science, 2004
In content-based image retrieval (CBIR), the images in a database are indexed on the basis of low... more In content-based image retrieval (CBIR), the images in a database are indexed on the basis of low-level statistical features that can be automatically derived from the images. Due to the semantic gap, the performance of CBIR systems often remains quite modest especially on broad image domains. One method for improving the results is to incorporate automatic image classification methods to the CBIR system. The resulting subsets can be indexed separately with features suitable for those particular images or used to limit an image query only to certain promising image subsets. In this paper, a method for supporting different types of image subsets within a generic framework based on multiple parallel Self-Organizing Maps and binary clusterings is presented.
Analysing the Structure of Semantic Concepts in Visual Databases
Lecture Notes in Computer Science, 2011
In this paper we study how the Self-Organizing Map (SOM) can be used in analysing the structure o... more In this paper we study how the Self-Organizing Map (SOM) can be used in analysing the structure of semantic concepts in visual data. We investigate two data sets with concept labels provided by humans, and unlabelled data for which we utilise automatically detected concept membership scores by using models trained on a labelled data set. By arranging the concept memberships of visual objects as components of a vector, they can be used as the feature space for training a SOM. A visual and qualitative analysis of the SOM distributions of different concepts is augmented with a quantitative analysis based on measuring the Earth Mover’s Distance between the vector distributions on the 2D SOM surface. In particular we study the PASCAL VOC 2007 and TRECVID 2010 databases, which are two large image and video data sets.
Subspace dimension selection and averaged learning subspace method in handwritten digit classification
Lecture Notes in Computer Science, 1996
. We present recent improvements in using subspace classi?ersin recognition of handwritten digits... more . We present recent improvements in using subspace classi?ersin recognition of handwritten digits. Both non-trainable CLAFIC andtrainable ALSM methods are used with four models for initial selectionof subspace dimensions and their further error-driven re?nement. Theresults indicate that these additions to the subspace classi?cation schemenoticeably reduce the classi?cation error.1 IntroductionIn a recent study [1], we demonstrated that subspace classi?ers, especially theerror...
LVQ PAK: A program package for the correct application of Learning Vector Quantization algorithms
International Joint Conference on Neural Networks, 1992
. This paper is an overview of the program package LVQ PAK, which has been developedfor convenien... more . This paper is an overview of the program package LVQ PAK, which has been developedfor convenient and effective application of Learning Vector Quantization algorithms. Two newfeatures are included: fast conflict-free initial distribution of codebook vectors into the class zones,and the optimized-learning-rate algorithm OLVQ1.1 IntroductionStatistical classification or pattern recognition of stochastic vectorial samples can be performedby the Learning Vector Quantization
Techniques for Still Image Scene Classification and Object Detection
Lecture Notes in Computer Science, 2006
In this paper we consider the interaction between different semantic levels in still image scene ... more In this paper we consider the interaction between different semantic levels in still image scene classification and object detection problems. We present a method where a neural method is used to produce a tentative higher-level semantic scene representation from low-level statistical visual features in a bottom-up fashion. This emergent representation is then used to refine the lower-level object detection results.
Lecture Notes in Computer Science, 2004
This paper presents a task-based user evaluation of two contentbased image database browsing syst... more This paper presents a task-based user evaluation of two contentbased image database browsing systems. The performance of the two systems is compared to that of a commercial image database management program, which does not employ content-based information. Experimental results show that content-based cues improve the efficiency of the browsing considerably. Guidelines for system design are derived from the user feedback.
IFIP International Federation for Information Processing, 2006
An art installation was on display in the Centre Pompidou National Museum of Modern Art in Paris,... more An art installation was on display in the Centre Pompidou National Museum of Modern Art in Paris, were visitors could contribute with their own personal objects, adding keyword descriptions and quantified semantic features such as age or hardness. The data was projected in real-time onto a Self-Organizing Map (SOM) and shown in the gallery. In this paper we analyze the same data by extracting visual features from the images and organize the image collection with multiple SOMs. We show how this mapping facilitates the emergence of semantic associations between visual, textual and modalities by studying the distributions of the different feature vectors on the SOMs.
Abstract. Standard ,performance ,tests of information ,retrieval systems ,are based on measuring ... more Abstract. Standard ,performance ,tests of information ,retrieval systems ,are based on measuring precision at fixed recall levels and averaging results over a large set of test topics. It has recently been demonstrated,how,seemingly,equal average,performance ,may ,obscure ,important ,differences between ,search methods. The in-depth analysis of individual ,queries is needed ,for revealing the hidden differences. This paper introduces the motivation, structure and
In this article the use of statistical, low-level shape features in content-based image retrieval... more In this article the use of statistical, low-level shape features in content-based image retrieval is studied. The emphasis is on such techniques which do not demand object segmentation. PicSOM, the image retrieval system used in the experiments, requires that features are represented by constant-sized feature vectors for which the Euclidean distance can be used as a similarity measure. The shape features suggested here are edge histograms and Fourier-transform-based features computed from the image after edge detection in Cartesian or polar coordinate planes. The results show that both local and global shape features are important clues of shapes in an image.
Content-Based Image Retrieval Using Self-Organizing Maps
Lecture Notes in Computer Science, 1999
We have developed an image retrieval system named PicSOMwhich uses Tree Structured Self-Organizin... more We have developed an image retrieval system named PicSOMwhich uses Tree Structured Self-Organizing Maps (TS-SOMs) asthe method for retrieving images similar to a given set of reference images.A novel technique introduced in the PicSOM system facilitates automaticcombination of the responses from multiple TS-SOMs and theirhierarchical levels. This mechanism aims at adapting to the user's preferencesin selecting which images resemble each
Overview of the ImageCLEF 2007 Object Retrieval Task
Lecture Notes in Computer Science, 2008
Classification with learning k-nearest neighbors
Proceedings of International Conference on Neural Networks (ICNN'96), 1996
The nearest neighbor (NN) classifiers, especially the k-NN algorithm, are among the simplest and ... more The nearest neighbor (NN) classifiers, especially the k-NN algorithm, are among the simplest and yet most efficient classification rules and are widely used in practice. We introduce three adaptation rules that can be used in iterative training of a k-NN classifier. This is a novel approach both from the statistical pattern recognition and the supervised neural network learning points of
Uploads
Papers by Jorma Laaksonen