Pere Marti-Puig

A new otolith image contour descriptor based on partial reflection

In this paper we propose a new contour descriptor for reconstructing fish otolith contours that uses half the number of coefficients needed by the classical elliptical Fourier descriptors (EFDs) for the same accuracy. The efficiency of the proposed shape descriptor has been tested with two different species, Liza aurata and Liza ramada, belonging to the family Mugilidae (http://aforo.cmima.csic.es), and two populations (from the USA and Canada) of the family Merlucciidae. These groups are characterized by high similarity between species; therefore, accurate, detailed shape analyses of their otoliths can help to identify and discriminate morphologically close species or different populations. For comparative purposes the descriptor was also tested with specimens of Mullus barbatus (Family Mullidae). For a certain number of coefficients (<50) the new descriptor clearly outperforms the reconstruction accuracy of the EFD.

Stereo Video Surveillance Multi-agent System: New Solutions for Human Motion Analysis

This article presents a distributed agent-based system that can process the visual information obtained by stereoscopic cameras. The system is embedded within a global project whose objective is to develop an intelligent environment for location and identification within dependent environments that merges with other types of technologies. In this kind of environments, vision algorithms are very costly and require a lot of time to produce a response, which is highly inconvenient since many applications can require action to be taken in real time. A multi-agent system (MAS) can automate the process of analyzing images obtained by cameras, and optimize the procedure. This study presents a MAS that can process stereoscopic images to detect and classify people by combining a series of novel techniques.The article shows in detail the combination of techniques used to perform the detection process. The process can be subdivided into human detection, human tracking, and human behavior understanding. With the addition of a case-based reasoning (CBR) model, the system can also incorporate reasoning capabilities. The system was tested under different conditions and environments.

Mathematical morphology preprocessing to mitigate AWGN effects: improving pitch tracking performance in hard noise conditions

In this paper we show how a nonlinear preprocessing of speech signal -with high noise- based on morphological filters improves the performance of robust algorithms for pitch tracking (RAPT). This result happens for a very simple morphological filter. More sophisticated ones could even improve such results. Mathematical morphology is widely used in image processing in where it has found a great amount of applications. Almost all its formulations derived in the two-dimensional framework are easily reformulated to be adapted to one-dimensional context.

Score function for voice activity detection

In this paper we explore the use of non-linear transformations in order to improve the performance of an entropy based voice activity detector (VAD). The idea of using a non-linear transformation comes from some previous work done in speech linear prediction (LPC) field based in source separation techniques, where the score function was added into the classical equations in order to take into account the real distribution of the signal. We explore the possibility of estimating the entropy of frames after calculating its score function, instead of using original frames. We observe that if signal is clean, estimated entropy is essentially the same; but if signal is noisy transformed frames (with score function) are able to give different entropy if the frame is voiced against unvoiced ones. Experimental results show that this fact permits to detect voice activity under high noise, where simple entropy method fails.

New parameterisation method for three-dimensional otolith surface images

The three-dimensional (3-D) otolith shapes recently included in the Analisi de FORmes d’Otolits (AFORO) database are defined by means of clouds of points across their surfaces. Automatic retrieval and classification of natural objects from 3-D databases becomes a difficult and time-consuming task when the number of elements in the database becomes large. In order to simplify that task we propose a new method for compacting data from 3-D shapes. The new method has two main steps. The first is a subsampling process, the result of which can always be interpreted as a closed curve in the 3-D space by considering the retained points in an appropriate order. The subsampling preserves morphological information, but greatly reduces the number of points required to represent the shape. The second step treats the coordinates of the 3-D closed curves as periodic functions. Therefore, Fourier expansions can be applied to each coordinate, producing more information compression into a reduced set of points. The method can reach very high information compression factors. It also allows reconstruction of the 3-D points resulting from the subsampling process in the first step. This parameterisation method is able to capture 3-D information relevant to classification of fish species from their otoliths, providing a greater percentage of correctly classified specimens compared with the previous two-dimensional analysis.

Otoliths identifiers using image contours EFD

In this paper we analyze the characteristics of an experimental otolith identification system based on image contours described with Elliptical Fourier Descriptors (EFD). Otoliths are found in the inner ear of fishes. They are formed by calcium carbonate crystals and organic materials of proteic origin. Fish otolith shape analysis can be used for sex, age, population and species identification studies, and can provide necessary and relevant information for ecological studies. The system we propose has been tested for the identification of three different species, Engraulis encrasicholus, Pomadasys incisus belonging to the different families (Engroulidae and Haemolidae), and two populations of the species Merluccius merluccius (from CAT and GAL) from the family Merlucciidae. The identification of species from different families could be carried out quite easily with some simple class identifiers -i.e based on Support Vector Machine (SVM) with linear Kernel-; however, to identify these two populations that are characterized by a high similarity in their global form; a more accurate, and detailed shape representation of the otoliths are required, and at the same time the Otolith identifiers have to deal with a bigger number of descriptors. That is the principal reason that makes a challenging task both the design and the training of an otolith identification system, with a good performance on both cases.

Image Processing to Detect and Classify Situations and States of Elderly People

Monitoring and tracking of elderly people using vision algorithms is an strategy gaining relevance to detect anomalous and potentially dangerous situations and react immediately. In general vision algorithms for monitoring and tracking are very costly and take a lot of time to respond, which is highly inconvenient since many applications can require action to be taken in real time. A multi-agent system (MAS) can establish a social model to automate the tasks carried out by the human experts during the process of analyzing images obtained by cameras. This study presents a detector agent integrated in a MAS that can process stereoscopic images to detect and classify situations and states of elderly people in geriatric residences by combining a series of novel techniques. We will talk in details about the combination of techniques used to perform the detection process, subdivided into human detection, human tracking ,and human behavior understanding, and where there is a case-based reasoning (CBR) model that allows the system to add reasoning capabilities.

Radix-4 FFT algorithms with ordered input and output data

In this work we derive two families of radix-4 factorizations for the FFT (Fast Fourier Transform) that have the property that both inputs and outputs are addressed in natural order. These factorizations are obtained from another two families of radix-2 algorithms that have the same property. The radix-4 algorithms obtained have the same mathematical complexity (number of multiplications and additions) that Cooley-Tukey radix-4 algorithms but avoid de bit-reversal ordering applied to the input or at the output.

On-line Ham Grading using pattern recognition models based on available data in commercial pig slaughterhouses

The thickness of the subcutaneous fat in hams is one of the most important factors for the dry-curing process and largely determines its final quality. This parameter is usually measured in slaughterhouses by a manual metrical measure to classify hams. The aim of the present study was to propose an automatic classification method based on data obtained from a carcass automatic classification equipment (AutoFom) and intrinsic data of the pigs (sex, breed, and weight) to simulate the manual classification system. The evaluated classification algorithms were decision tree, support vector machines (SVM), k-nearest neighbour and discriminant analysis. A total of 4000 hams selected by breed and sex were classified as thin (0-10 mm), standard (11-15 mm), semi-fat (16-20 mm) and fat (>20 mm). The most reliable model, with a percentage of success of 73%, was SVM with Gaussian kernel, including all data available. These results suggest that the proposed classification method can be a useful online tool in slaughterhouses to classify hams.

Effects of the pre-processing algorithms in fault diagnosis of wind turbines

Abstract The wind sectors pends roughly 2200M€ in repair the wind turbines failures. These failures do not contribute to the goal of reducing greenhouse gases emissions. The 25–35% of the generation costs are operation and maintenance services. To reduce this amount, the wind turbine industry is backing on the Machine Learning techniques over SCADA data. This data can contain errors produced by missing entries, uncalibrated sensors or human errors. Each kind of error must be handled carefully because extreme values are not always produced by data reading errors or noise. This document evaluates the impact of removing extreme values (outliers) applying several widely used techniques like Quantile, Hampel and ESD with the recommended cut-off values. Experimental results on real data show that removing outliers systematically is not a good practice. The use of manually defined ranges (static and dynamic) could be a better filtering strategy.