Michel Herbin

A clustering method based on the estimation of the probability density function and on the skeleton by influence zones: application to image processing

This paper investigates a new approach to data clustering. The probability density function (p.d.f.) is estimated by using the Parzen window technique. The p.d.f. thresholding permits the segmentation of the data space by influence zones (SKIZ algorithm). A bottom-up thresholding procedure is iterated to refine the segmentation. As a result, a complete partition of the data space is obtained in parallel to the clustering of the data samples. In addition, an estimation of the intrinsic dimensionality of the data set is provided. This approach of clustering is tested with simulated data and applied to color image data.

A 'no-threshold' histogram-based image segmentation method

Although most histogram-based image segmentation methods rely on the identification of a good threshold, we show that thresholding is not mandatory. Instead, we propose the association of grades of membership to each individual pixel, in order to perform probabilistic relaxation in the image space (which realizes some kind of regularization) and finally to obtain the segmented image through defuzzyfication of the relaxed grades of membership.

Estimation of the number of clusters and influences zones

Abstract Whereas estimating the number of clusters is directly involved in the first steps of unsupervised classification procedures, the problem still remains topical. In our attempt to propose a solution, we focalize on procedures that do not make any assumptions on the cluster shapes. Indeed the classification approach we use is based on the estimation of the probability density function (PDF) using the Parzen–Rosenblatt method. The modes of the PDF lead to the construction of influence zones which are intrinsically related to the number of clusters. In this paper, using different sizes of kernel and different samplings of the data set, we study the effects they imply on the relation between influence zones and the number of clusters. This ends up in a proposal of a method for counting the clusters. It is illustrated in simulated conditions and then applied on experimental results chosen from the field of multi-component image segmentation.

Unsupervised Visual Data Mining Using Self-organizing Maps and a Data-driven Color Mapping

This paper presents a new approach for visually mining multivariate datasets and especially large ones. This unsupervised approach proposes to mix a SOM approach and a pixel-oriented visualization. The map is considered as a set of connected pixels, the space filling is driven by the SOM algorithm, and the color of each pixel is computed directly from data using an approach proposed by Blanchard et al. The method visually summarizes the data and helps in understanding its inner structure.

Comparative study of different spatial/spatial-frequency methods (Gabor filters, wavelets, wavelets packets) for texture segmentation/classification

The interest manifested into texture recognition has had a substantial growth since these past few years. Much research is focused on a joint space/spatial-frequency representation of images. This paper describes the comparison between different spatial/spatial-frequency methods involving Gabor filters and wavelets. Two applications are considered: image segmentation and texture classification/recognition. It has been seen that results can depend significantly on the method used to decompose the signal. Two aspects are mainly taken into account: the effect of an over-sampling of the frequency domain by a bank of filters, and the ability of 1D wavelets to segment/classify oriented textures. An improvement using angular filters with separable wavelets is demonstrated, while no noticeable amelioration is observed using Battle-Lemarie wavelets, as compared to Gabor filters.

Are Existing Procedures Enough? Image and Video Quality Assessment: Review of Subjective and Objective Metrics

Images and videos are subject to a wide variety of distortions during acquisition, digitizing, processing, restoration, compression, storage, transmission and reproduction, any of which may result in degradation in visual quality. That is why image quality assessment plays a major role in many image processing applications. Image and video quality metrics can be classified by using a number of criteria such as the type of the application domain, the predicted distortion (noise, blur, etc.) and the type of information needed to assess the quality (original image, distorted image, etc.). In the literature, the most reliable way of assessing the quality of an image or of a video is subjective evaluation [1], because human beings are the ultimate receivers in most applications. The subjective quality metric, obtained from a number of human observers, has been regarded for many years as the most reliable form of quality measurement. However, this approach is too cumbersome, slow and expensive for most applications [2]. So, in recent years a great effort has been made towards the development of quantitative measures. The objective quality evaluation is automated, done in real time and needs no user interaction. But ideally, such a quality assessment system would perceive and measure image or video impairments just like a human being [3]. The quality assessment is so important and is still an active and evolving research topic because it is a central issue in the design, implementation, and performance testing of all systems [4, 5]. Usually, the relevant literature and the related work present only a state of the art of metrics that are limited to a specific application domain. The major goal of this paper is to present a wider state of the art of the most used metrics in several application domains such as compression [6], restoration [7], etc. In this paper, we review the basic concepts and methods in subjective and objective image/video quality assessment research and we discuss their performances and drawbacks in each application domain. We show that if in some domains a lot of work has been done and several metrics were developed, on the other hand, in some other domains a lot of work has to be done and specific metrics need to be developed.

Biometric System Based on Voice Recognition Using Multiclassifiers

In this paper we present a new speaker recognition system based on the fusion of two identification classifiers followed by a verification step. The user pronounces two passwords: the first one is composed by three words uniquely combined from a set of 21 possible words, while the second password represents the name of the user. The first step of the proposed system uses the first password to feed two identification classifiers: a speaker identification system (text independent) and a isolated word identification system (speaker independent). The isolated word identification system is constructed as the fusion of three classifiers, one for each word of the first password. The aim of this first step is to identify a couple speaker/password corresponding to the first password by combining the results of the two identification classifiers. A verification system is then applied on the second password in order to confirm or infirm the identification result (speaker identity) given by the fusion above. Compared with a state of the art speaker recognition system (text dependent) that gives an EER of 4.76%, the first step of the proposed system provides an EER of 0.38%, while the second step an EER of 0.26% for a text independent verification and of 0.13% for a text dependent verification.

Anteriority index for managing fuzzy dates in archæological GIS

During the exploitation of an archæological geographical information system, experts need to evaluate the anteriority in pairs of dates which are uncertain and inaccurate, and consequently represented by fuzzy numbers. To build their hypotheses, they need to have an assessment, taking value in [0, 1], of the relation “lower than” between two FNs. We answer the experts’ need of evaluation by constructing an anteriority index based on the Kerre index. Two applications, which constitute a step in the evaluation of the evolution of Reims during the domination of the Roman Empire, illustrate the use of the anteriority index.

No-reference Image Semantic Quality Approach using Neural Network

Assessment for image quality traditionally needs its original image as a reference but the most of time it is not the case. So, No-Reference (NR) Image Quality Assessment (IQA) seeks to assign quality scores that are consistent with human perception but without an explicit comparison with the reference image. Unfortunately, the field of NR IQA has been largely unexplored. This paper presents a new NR Image Semantic Quality Approach (NR-ISQA) that employs adaptive Neural Networks (NN) to assess the semantic quality of image color. This NN measures the quality of an image by predicting the mean opinion score (MOS) of human observer, using a set of proposed key features especially to describe color. This challenging issues aim at emulating judgment and replacing very complex and time-consuming subjective quality assessment. Two variants of our approach are proposed: the direct and the progressive of the overall quality image. The results show the performances of the proposed approach compared with the human performances.

An innovative approach in structured light systems

This paper presents an integrated 3D face scanning system using the structured light technique. After illuminating the face by a pattern with horizontal colored strips using the De_Bruijn sequence, an image is taken and used to obtain the 3D information. A second image without illumination is used to add the texture to the reconstructed model. The precision of the 3D model depends on the determination of the strip centers. The technique proposed to determine these centers uses a smoothing Gaussian filter with a large kernel applied to the V component in the HSV color space. A classic connection algorithm allows to link the isolated points in the detected strips. The color of the detected strips is determined for the whole connected parts in two steps. Firstly, we use the H component of the HSV color space to determine the color of each set of pixels. Then, we apply a region growing algorithm to assign the colors to the remaining pixels. Each connected and colored part of line is treated alone and matched to a line in the projected pattern. Finally each detected point is triangulated with its corresponding one in the projected pattern to generate the model. Experiment results show a good 3D resolution with this technique.