Search

International audience; Documenting the inter-annual variability and the long-term trend of the glacier snow line altitude is highly relevant to document the evolution of glacier mass changes. Automatically identifying the snow line on glaciers is challenging; recent developments in machine learning approaches show promise to tackle this issue. This manuscript presents a proof of concept of machine learning approaches applied to multi-spectral images to detect the snow line and quantify its average altitude. The tested approaches include the combination of different image processing and classification methods, and takes into account cast shadows. The efficiency of these approaches is evaluated on mountain glaciers in the European Alps by comparing the results with manually annotated data. Solutions provided by the different approaches are robust when compared to the ground truth’s snow lines, with a Pearson’s correlation ranging from 79% to 96% depending on the method. However, the tested approaches may fail when snow lines are not continuous or exhibit a strong change of elevation. The major advantage over the state of the art is that the proposed approach does not require one calibration per glacier.

Multi-spectral semantic segmentation has shown great advantages under poor illumination conditions, especially for remote scene understanding of autonomous vehicles, since the thermal image can provide complementary information for RGB image. However, methods to fuse the information from RGB image and thermal image are still under-explored. In this paper, we propose a simple but effective module, add–multiply fusion (AMFuse) for RGB and thermal information fusion, consisting of two simple math operations—addition and multiplication. The addition operation focuses on extracting cross-modal complementary features, while the multiplication operation concentrates on the cross-modal common features. Moreover, the attention module and atrous spatial pyramid pooling (ASPP) modules are also incorporated into our proposed AMFuse modules, to enhance the multi-scale context information. Finally, in the UNet-style encoder–decoder framework, the ResNet model is adopted as the encoder. As for the decoder part, the multi-scale information obtained from our proposed AMFuse modules is hierarchically merged layer-by-layer to restore the feature map resolution for semantic segmentation. The experiments of RGBT multi-spectral semantic segmentation and salient object detection demonstrate the effectiveness of our proposed AMFuse module for fusing the RGB and thermal information.

Precisely detecting solar active regions (AR) from multi-spectral images is a challenging task yet important in understanding solar activity and its influence on space weather. A main challenge comes from each modality capturing a different location of these 3D objects, as opposed to more traditional multi-spectral imaging scenarios where all image bands observe the same scene. We present a multi-task deep learning framework that exploits the dependencies between image bands to produce 3D AR detection where different image bands (and physical locations) each have their own set of results. Different feature fusion strategies are investigated in this work, where information from different image modalities is aggregated at different semantic levels throughout the network. This allows the network to benefit from the joint analysis while preserving the band-specific information. We compare our detection method against baseline approaches for solar image analysis (multi-channel coronal hole detection, SPOCA for ARs (Verbeeck et al. Astron Astrophys 561:16, 2013)) and a state-of-the-art deep learning method (Faster RCNN) and show enhanced performances in detecting ARs jointly from multiple bands. We also evaluate our proposed approach on synthetic data of similar spatial configurations obtained from annotated multi-modal magnetic resonance images.

Over the past few years, deep learning algorithms have held immense promise for better multi-spectral (MS) optical remote sensing image (RSI) analysis. Most of the proposed models, based on convolutional neural network (CNN) and fully convolutional network (FCN), have been applied successfully on computer vision images (CVIs). However, there is still a lack of exploration of spectra correlation in MS RSIs. In this study, a deep neural network with a spectrum separable module (DSSM) is proposed for semantic segmentation, which enables the utilization of MS characteristics of RSIs. The experimental results obtained on Zurich and Potsdam datasets prove that the spectrum-separable module (SSM) extracts more informative spectral features, and the proposed approach improves the segmentation accuracy without increasing GPU consumption.

Multi-spectral imaging (MSI) is a novel non-invasive tool for visualizing the entire span of the eye, from the internal limiting membrane to the choroid. However, spatial misalignments can be frequently observed in sequential MSI images because the eye saccade movement is usually faster than the MSI image acquisition speed. Therefore, registering MSI images is necessary for computer-based analysis of retinal degeneration via MSI. In this paper, we propose an early deep learning framework for achieving an accurate registration of MSI images in a group-wise fashion. The framework contains three parts: a template construction based on principal component analysis, a deformation field calculation, and a spatial transformation. The framework is uniquely capable of resolving two key challenges, i.e., the “multi-modal” characteristics in MSI images for the acquisition with different spectra and the requirement of joint registration of the sequential images. Our experimental results demonstrate the superior performance of our framework compared to several representative state-of-the-art techniques in both speed and accuracy.

International audience; Precisely detecting solar Active Regions (AR) from multi-spectral images is a challenging task yet important in understanding solar activity and its influence on space weather. A main challenge comes from each modality capturing a different location of these 3D objects, as opposed to more traditional multi-spectral imaging scenarios where all image bands observe the same scene. We present a multi-task deep learning framework that exploits the dependencies between image bands to produce 3D AR detection where different image bands (and physical locations) each have their own set of results. We compare our detection method against baseline approaches for solar image analysis (multi-channel coronal hole detection, SPOCA for ARs (Verbeeck et al., 2013)) and a state-of-the-art deep learning method (Faster RCNN) and show enhanced performances in detecting ARs jointly from multiple bands.

������������ �������������������� ���� �������������� �������������������� ���������������� �������������������������� �� �������������������������� �������������������� �������������������� ������ �������������� ���������� ���������������������������� ���������������������� �� �������������� ���������������������������������� �������������� ���������������������������� ������������������������ ������������ ����������������������. ������������������������ �� ������������ ���������� �������������� ���� ����������������-���������� ���������������������� �������������� �� ������������������ �������������������� �������������������� ���� ���������������������������������� ��������������. ������-�������������������� ���������������������� ���������������������������� ����������������, �������������� ������������������������ ���������� ������������������������ ��������������-������ ���������������������������� ���������������������� ���� ������������ ���� ������������������ �� �������������������� ����������������������, ������������������ ���� ���������� ���������������������� ����������������������. �� ���������������� �������������� �������������������� ���������� ������������������ �� ������������������������������ �������������������� ���������������������� ������������������������ ������������������������ �������������������� ������������. ������������ ���������������� �������������������������� ���������������������� ���� ������������ �������������������� ���������� ������������������ ���������������������� �������������������� ���������� �������������������� ������������-���������� ������������ ���������������� �� �������������� ������������ ��������������, ���������������������� �������������� ���������������������� �������������������� ������-������������ �� �������������� ���������������� ���������������������� ������������������. ������������������ ���������������� ���������������������������� ���������� �������������� ���������������������� �������������������������� �������������������������� �� ������������������ ������������ ������������������������, ������������������������������ ���� ������������������ �������������� �������������������� ������������������������ �� ���� �������������� ������������������. ���������������� ������������������ ���������� ���������� �������������������������������� �������������������� �������������������������� ������������������������ ������������������ ������������ �������������������� �������������������� �� ������������ ������������, �� ���������� �������������������� ������������������������ ������������ ������������ ���������������������������� ������������������������, �� ������ ���������� �� ���������� �������������������������������� ����������������������., The article is concerned with the study of possibility to identify anomalous surveillance in multi-spectral images through distinctive-feature clustering, as well as by using the developed method for environmental monitoring. This article���s topic is significant since there are many problems associated with identifying anomalous surveil-lance in time series of images bound mainly to the need for detailed study of each picture. The paper proposes a method for estimating anomalous surveillance using summarized characteristics of satellite images, such as brightness entropy and Minkowski fractal dimension and others. Hierarchical clustering is executed based on the above characteristics and using Gower's distance breaking down as metrics. That enables to identify pictures with anomalous objects based on exponent for sum of squares of metric distances between image attributes in-side the cluster and average silhouette width, which make it possible to select optimal number of clusters and evaluate the quality of breaking down results. Anomalies��� detection is done by analyzing hierarchical clustering results and detecting dendrogram branches located at initial levels of dendrogram scheme and without branch-ing. The authors did work on selecting distance metric when performing clustering and on studying characteris-tics of breaking down in case of partitional and agglomerative clustering. The examples of using methods as well as, distinctive features, advantages and disadvantages are presented. Computer mathematical modeling is also done using Daisy (), Diana(), clusplot() packages, R language and RStudio development environment.

Copernicus program via its Sentinel missions is making earth observation more accessible and affordable for everybody. Sentinel-2 images provide multi-spectral information every 5 days for each location. However, the maximum spatial resolution of its bands is 10m for RGB and near-infrared bands. Increasing the spatial resolution of Sentinel-2 images without additional costs, would make any posterior analysis more accurate. Most approaches on super-resolution for Sentinel-2 have focused on obtaining 10m resolution images for those at lower resolutions (20m and 60m), taking advantage of the information provided by bands of finer resolutions (10m). Otherwise, our focus is on increasing the resolution of the 10m bands, that is, super-resolving 10m bands to 2.5m resolution, where no additional information is available. This problem is known as single-image super-resolution and deep learning-based approaches have become the state-of-the-art for this problem on standard images. Obviously, models learned for standard images do not translate well to satellite images. Hence, the problem is how to train a deep learning model for super-resolving Sentinel-2 images when no ground truth exist (Sentinel-2 images at 2.5m). We propose a methodology for learning Convolutional Neural Networks for Sentinel-2 image super-resolution making use of images from other sensors having a high similarity with Sentinel-2 in terms of spectral bands, but greater spatial resolution. Our proposal is tested with a state-of-the-art neural network showing that it can be useful for learning to increase the spatial resolution of RGB and near-infrared bands of Sentinel-2. This work was partially supported by the Public University of Navarre under project PJUPNA13 and Tracasa Instrumental S.L. (OTRI 2018 901 073 and OTRI 2019 901 091).

One of the most challenging problems in precision agriculture is to correctly identify and separate crops from the soil. Current precision farming algorithms based on artificially intelligent networks use multi-spectral or hyper-spectral data to derive radiometric indices that guide the operational management of agricultural complexes. Deep learning applications using these big data require sensitive filtering of raw data to effectively drive their hidden layer neural network architectures. Threshold techniques based on the normalized difference vegetation index (NDVI) or other similar metrics are generally used to simplify the development and training of deep learning neural networks. They have the advantage of being natural transformations of hyper-spectral or multi-spectral images that filter the data stream into a neural network, while reducing training requirements and increasing system classification performance. In this paper, to calculate a detailed crop/soil segmentation based on high resolution Digital Surface Model (DSM) data, we propose the redefinition of the radiometric index using a directional mathematical filter. To further refine the analysis, we feed this new radiometric index image of about 3500 × 4500 pixels into a relatively small Convolution Neural Network (CNN) designed for general image pattern recognition at 28 × 28 pixels to evaluate and resolve the vegetation correctly. We show that the result of applying a DSM filter to the NDVI radiometric index before feeding it into a Convolutional Neural Network can potentially improve crop separation hit rate by 65%.

In this study, the dynamics of two vegetation indices, the normalized differential vegetative index (NDVI) and the variant of the NDVI that uses the green band (GNDVI) in a rice growing of the variety fedearroz 2000 in reproduction phase, are analyzed. These indices were calculated through the geoprocessing of multi-spectral aerial images taken by a drone or UAVs, with the aim of identifying which zones of the crops are under stress, healthy or dense. The rice growing had an area of approximately 4,1 hectares and its location corresponds to the farm El Faro in the footpath Campo Hermoso within the municipal district of San José de Cúcuta – Norte de Santander. For this research, two flights were carried out, one at the beginning of the reproduction phase dated September 4th 2016 and the second one at the end corresponding to October 8th 2016; these flights were performed with a Iris+ 3DR drone, a canon S100 camera was implemented as a catch images sensor converted into NDVI by using a NGB filter (Near infrared, Green and Blue). As a result, 4 mosaics are shown, one NDVI and one GNDVI on September 4th 2016 and one NDVI and one GNDVI on October 8th 2016, each one of them were classified according to the characteristics observed in field in zones under stress or with low development, healthy and dense zones. Finally, a NDVI dynamic analysis was completed., En este estudio se analiza la dinámica de dos índices de vegetación, el índice vegetativo diferencial normalizado (NDVI) y la variante del NDVI que utiliza la banda verde (GNDVI) en un cultivo de arroz de la variedad fedearroz 2000 en fase de reproducción. Estos índices se calcularon mediante el geoprocesamiento de imágenes aéreas multiespectrales tomadas por un dron o UAVs, con el objetivo de identificar qué zonas del cultivo están bajo estrés, sanas o densas. El cultivo de arroz tenía un área aproximada de 4,1 hectáreas y su ubicación corresponde a la finca El Faro en la vereda Campo Hermoso dentro del municipio de San José de Cúcuta - Norte de Santander. Para esta investigación se realizaron dos vuelos, uno al inicio de la fase de reproducción fechado el 4 de septiembre de 2016 y el segundo al final correspondiente al 8 de octubre de 2016; estos vuelos se realizaron con un dron Iris+ 3DR, se implementó una cámara canon S100 como sensor de captura de imágenes convertidas en NDVI mediante el uso de un filtro NGB (Near infrared, Green and Blue). Como resultado se muestran 4 mosaicos, uno de NDVI y otro de GNDVI el 4 de septiembre de 2016 y otro de NDVI y otro de GNDVI el 8 de octubre de 2016, cada uno de ellos fue clasificado según las características observadas en campo en zonas bajo estrés o con bajo desarrollo, zonas sanas y densas. Finalmente, se completó un análisis dinámico del NDVI.

International audience; Skin ulcers (SU) are ones of the most frequent causes of consultation in primary health-care units (PHU) in tropical areas. However, the lack of specialized physicians in those areas, leads to improper diagnosis and management of the patients. There is then a need to develop tools that allow guiding the physicians toward a more accurate diagnosis. Multi-spectral imaging systems are a potential non-invasive tool that could be used in the analysis of skin ulcers. With these systems it is possible to acquire optical images at different wavelengths which can then be processed by means of mathematical models based on optimization approaches. The processing of those kind of images leads to the quantification of the main components of the skin. In the case of skin ulcers, these components could be correlated to the different stages of wound healing during the follow-up of a skin ulcer. This article presents the processing of a skin ulcer multi-spectral image. The ulcer corresponds to Leishmaniasis which is one of the diseases the most prominent in tropical areas. The image processing is performed by means of a light-tissue interaction model based on the distribution of the skin as a semi-infinite layer. The model, together with an optimization approach allows quantifying the main light-absorbing and scattering skin-parameters in the visible and near-infrared range. The results show significant differences between healthy and unhealthy area of the image.

WADI water leak detection surveillance service consists in the application of optical remote sensing on manned and unmanned aerial platforms. This innovative and cost-effective technology will be tested on two pilot sites located in the Provence region (France) and in Alqueva (Portugal).

Face recognition systems based on ’conventional’ images have reached a significant level of maturity with some practical successes. However, their performance may degrade under poor and/or changing illumination. Multispectral imagery represents a viable alternative to conventional imaging in the search for a robust and practical identification system. Multi- spectral imaging (MI) can be defined as a ’collection of several monochrome images of the same scene, each of them taken with additional receptors sensitive to other frequencies of the visible light or to frequencies beyond the visible light like the infrared region of electro- magnetic continuum. Each image is referred to as a band or a channel. However, one weakness of MI is that they may significantly increase the system processing time because of the huge quantity of data to be mined; in some cases, hundreds of MI are taken for each subject. In this thesis, we propose to solve this problem by developing new approaches to select the set of best visible spectral bands for face matching. For this purpose, the problem of best spectral bands selection is formulated as an optimization problem where spectral bands are constrained to maximize the recognition accuracy under challenging imaging conditions. We reduce the redundancy of both spectral and spatial information without losing valuable details needed for the object recognition, discrimination and classification. We have investigated several mathematic and optimization tools widely used in the field of image processing. One of the approaches we have proposed formulated the problem of best spectral bands selection as a pursuit problem where weights of importance were affected to each spectral band and the vector of all weights was constrained to be sparse with most of its elements are zeros. In another work, we have assigned to each spectral band a linear discriminant analysis (LDA) based weak classifier. Then, all weak classifiers were boosted together using an Adaboost process. From this later, each weak classifier obtained a weight that characterizes its importance and hence the quality of the corresponding spectral band. Several other techniques were also used for best spectral bands selection including but not limited to mixture of Gaussian based modeling, multilinear sparse decomposition, image quality factors, local descriptors like SURF and HGPP, likelihood ratio and so on. These different techniques enabled to build systems for best spectral bands selection that are either static with the same bands are selected for all the subjects or dynamic with each new subject get its own set of best bands. This latter category, dynamic systems, is an original component of our work that, to the best of our knowledge, has not been proposed before; all existing systems are only static. Finally, the proposed algorithms were compared to state-of-the-art algorithms developed for face recognition purposes in general and specifically for best spectral bands selection.; Les performances des systèmes de reconnaissance des visages en utilisant des images RGB baissent rapidement quand ils sont appliqués dans des conditions d’illumination extrêmes. L’utilisation des images multispectrales représente une alternative prometteuse pour résoudre ce problème. Dans cette thèse on s’intéresse à l’utilisation des images multispectrales visibles pour la reconnaissance des visages humains. Les images multispectrales visibles sont des images capturées à des longueurs d’ondes différentes du spectre visible (band spectral) qui s’étend de 480nm à 720nm. Ces images représentent des caractéristiques qui favorisent la reconnaissance des visages humains dans des conditions particulières comme la présence d’excès d’illumination incidente sur le visage photographié. Notre travail consiste à exploiter ces caractéristiques sur des stages différentes: optimiser le choix du nombre de bandes spectrales à utiliser, optimiser les longueurs d’ondes choisies, optimiser les techniques de fusion des informations extraites à partir des différentes bandes spectrales pour avoir plus d’informations utiles et moins d’informations bruits. Plusieurs nouvelles approches ont été proposées dans le cadre de ce travail avec des résultats encourageants en termes de performances. Ces approches ont exploité plusieurs outils mathématiques pour resoudre les différents problèmes rencontrés, en particulier la formulation de la sélection des bandes spectrales optimales sous formes de problèmes d’optimisation où nous avons utilisé le « basis pursuit algorithm » pour déterminer un vecteur de poids sparse pour représenter l’importance des différentes bandes. Dans d’autres problèmes d’optimisation, nous avons attribué à chaque bande un classifieur faible, puis combiné les classifieurs faibles avec dif- férents poids associés selon l’importance. La méthode Adaboost a été utilisée pour trouver la combinaison optimale. D’autres techniques ont introduites d’une manière originale la dé- composition multilinéaire des images de visage pour formuler une sorte de base de données caractérisant les bandes spectrales. Cette base de données a été utilisée avec les nouvelles images, ou image test, pour déterminer les bandes les plus robustes contre une variation importante d’illumination. Le travail présenté dans le cadre de cette thèse est une petite contribution à la reconnaissance des visages en utilisant des images multispectrales, qui est une approche d’actualité, mais qui nécessite encore plus de développement afin de maximiser ses performances.

Face recognition systems based on ’conventional’ images have reached a significant level of maturity with some practical successes. However, their performance may degrade under poor and/or changing illumination. Multispectral imagery represents a viable alternative to conventional imaging in the search for a robust and practical identification system. Multi- spectral imaging (MI) can be defined as a ’collection of several monochrome images of the same scene, each of them taken with additional receptors sensitive to other frequencies of the visible light or to frequencies beyond the visible light like the infrared region of electro- magnetic continuum. Each image is referred to as a band or a channel. However, one weakness of MI is that they may significantly increase the system processing time because of the huge quantity of data to be mined; in some cases, hundreds of MI are taken for each subject. In this thesis, we propose to solve this problem by developing new approaches to select the set of best visible spectral bands for face matching. For this purpose, the problem of best spectral bands selection is formulated as an optimization problem where spectral bands are constrained to maximize the recognition accuracy under challenging imaging conditions. We reduce the redundancy of both spectral and spatial information without losing valuable details needed for the object recognition, discrimination and classification. We have investigated several mathematic and optimization tools widely used in the field of image processing. One of the approaches we have proposed formulated the problem of best spectral bands selection as a pursuit problem where weights of importance were affected to each spectral band and the vector of all weights was constrained to be sparse with most of its elements are zeros. In another work, we have assigned to each spectral band a linear discriminant analysis (LDA) based weak classifier. Then, all weak classifiers were boosted together using an Adaboost process. From this later, each weak classifier obtained a weight that characterizes its importance and hence the quality of the corresponding spectral band. Several other techniques were also used for best spectral bands selection including but not limited to mixture of Gaussian based modeling, multilinear sparse decomposition, image quality factors, local descriptors like SURF and HGPP, likelihood ratio and so on. These different techniques enabled to build systems for best spectral bands selection that are either static with the same bands are selected for all the subjects or dynamic with each new subject get its own set of best bands. This latter category, dynamic systems, is an original component of our work that, to the best of our knowledge, has not been proposed before; all existing systems are only static. Finally, the proposed algorithms were compared to state-of-the-art algorithms developed for face recognition purposes in general and specifically for best spectral bands selection., Les performances des systèmes de reconnaissance des visages en utilisant des images RGB baissent rapidement quand ils sont appliqués dans des conditions d’illumination extrêmes. L’utilisation des images multispectrales représente une alternative prometteuse pour résoudre ce problème. Dans cette thèse on s’intéresse à l’utilisation des images multispectrales visibles pour la reconnaissance des visages humains. Les images multispectrales visibles sont des images capturées à des longueurs d’ondes différentes du spectre visible (band spectral) qui s’étend de 480nm à 720nm. Ces images représentent des caractéristiques qui favorisent la reconnaissance des visages humains dans des conditions particulières comme la présence d’excès d’illumination incidente sur le visage photographié. Notre travail consiste à exploiter ces caractéristiques sur des stages différentes: optimiser le choix du nombre de bandes spectrales à utiliser, optimiser les longueurs d’ondes choisies, optimiser les techniques de fusion des informations extraites à partir des différentes bandes spectrales pour avoir plus d’informations utiles et moins d’informations bruits. Plusieurs nouvelles approches ont été proposées dans le cadre de ce travail avec des résultats encourageants en termes de performances. Ces approches ont exploité plusieurs outils mathématiques pour resoudre les différents problèmes rencontrés, en particulier la formulation de la sélection des bandes spectrales optimales sous formes de problèmes d’optimisation où nous avons utilisé le « basis pursuit algorithm » pour déterminer un vecteur de poids sparse pour représenter l’importance des différentes bandes. Dans d’autres problèmes d’optimisation, nous avons attribué à chaque bande un classifieur faible, puis combiné les classifieurs faibles avec dif- férents poids associés selon l’importance. La méthode Adaboost a été utilisée pour trouver la combinaison optimale. D’autres techniques ont introduites d’une manière originale la dé- composition multilinéaire des images de visage pour formuler une sorte de base de données caractérisant les bandes spectrales. Cette base de données a été utilisée avec les nouvelles images, ou image test, pour déterminer les bandes les plus robustes contre une variation importante d’illumination. Le travail présenté dans le cadre de cette thèse est une petite contribution à la reconnaissance des visages en utilisant des images multispectrales, qui est une approche d’actualité, mais qui nécessite encore plus de développement afin de maximiser ses performances.

After decades of mining and industrialization in Qatar, it is important to estimate their impact on soil pollution with toxic metals. The study utilized 300 topsoil (0–30 cm) samples, multi-spectral images (Landsat 8), spectral indices and environmental variables to model and map the spatial distribution of arsenic (As), chromium (Cr), nickel (Ni), copper (Cu), lead (Pb) and zinc (Zn) in Qatari soils. The prediction model used condition-based rules generated in the Cubist tool. In terms of R2 and the ratio of performance to interquartile distance (RPIQ), the models showed good predictive capabilities for all elements. Of all of the prediction results, Cu had the highest R 2 = 0.74, followed by As > Pb > Cr > Zn > Ni. This study found that all of the models only chose images from January and February as predictors, which indicates that images from these two months are important for soil toxic metals’ monitoring in arid soils, due to the climate and the vegetation cover during this season. Topsoil maps of the six toxic metals were generated. The maps can be used to prioritize the choice of remediation measures and can be applied to other arid areas of similar environmental/socio-economic conditions and pollution causes.

According to clinical protocols, skin diseases are quantified by dermatologists throughout a treatment period, and then a statistical test on these measures allows to evaluate a treatment efficacy. The first step of this process it to classify pathological interest areas. This task is challenging due to the high variability of the images in one clinical data set. In this report, we first review algorithms that exist in the literature and adapt them to our problem. Then we choose the more appropriate algorithm to design a classification strategy. Thereby, we propose to use data reduction combined with SVM to do a first classification of the disease. Then we associate the obtained classification map with a segmentation map in an "interactive classification tool" in order to compromise between operator dependency and algorithm robustness.

This paper presents the usefulness and utility of three complementary digitisation techniques for the dynamic rendering of colour, roughness and shape of easel paintings. The 2D multi-spectral digitisation and the 3D modelling of paintings are used for colour accuracy rendering and measurement, colour characterisation, pigment identification, for high definition visualisation of the painting and also for the measurement of the shape of the support and the paint-layer roughness. The experimentation was performed on several paintings. The results show the advantages of each technique. Several viewers were developed to handle multi-spectral images and 3D models.

This PhD focuses on algorithms in the field of Mathematical Morphology and Image Processing, from the point of view of modern programming techniques. Computer science is often considered as a simple witness of hardware improvements dictated by Moore's law. However, software techniques evolve as well, and new programming tools such as metaprogramming are now available for the scientific community. Programming is of paramount interest for image processing; for this task, meta-programming brings significant improvements both in scientific terms by providing a powerful mean of abstraction, and in terms of practicability by providing portability, development centralizing, error reduction, etc. The work presented in this thesis is structured around the elaboration of a general algorithmic framework for - morphological - image processing. Furthermore, examples drawn from concrete industrial applications (in the fields of visual surveillance and car security) are described in order to illustrate the use of some of these developments. Prior to any algorithmic development, the thesis first proposes a model identifying underlying mathematical notions and their connections: data structures - images, graphs, topology, orders, and priority queues - are revisited with the meta-programming paradigm. This model clearly distinguishes the different actors and reaches the targeted abstraction level. Using specific mechanisms, the automation of numerous tasks is enabled with the compiler. Algorithms are then closer in description to the original mathematical formulations. These developments open for a wide area of research, including N dimensional and hyperspectral images which we further investigate in the following. The support of both N dimensional imaging and generic programming philosophy led us to define an exact distance transform algorithm. The hypotheses assumed regarding the underlying distance are few (homogeneity in space and convexity of the unit ball), which allows the use of a wide class of functions. Following a theoretical study, an algorithm is proposed based on ordered propagation. The error-proneness on many examples in 4D space (Euclidian, L5, oriented, non-isometric) is also illustrated. A different approach uses morphological distance transforms, popular in the field of mathematical morphology. This approach has recently seen a significant extension from binary to grey-scale images: Beucher's « quasi-distance ». In this thesis, an algorithm is proposed which features lower complexity than the one proposed originally for this purpose. The handling of colour, or more generally of hyperspectral images (i.e. vectorial data), is rather delicate in the field of Mathematical Morphology. This thesis tackles this issue, and proposes three different approaches: the use of colour metrics, then of local statistics and finally of algebraic lattices by means of lexicographical orders. The programming framework introduced in the document suits all the needs of such approaches. Colour metrics enable the definition of morphological gradients in colour spaces. However, this definition is computationally expensive and practically unusable for wide neighbourhoods. In this case, local statistics provide an interesting alternative solution. Focus was put on circular statistics in HLS colour space, which led to a chromatic gradient. Lexicographical orders also provide an algebraic framework suitable for mathematical morphology in colour spaces. Using the framework proposed in this thesis, such orders do not require a fundamental revisiting of the existing algorithms. Operators based on orders and geodesy (extrema, reconstructions, granulometries ...) are extended with a very low cost in development. These abstract considerations are illustrated within two concrete applications: skin colour characterization robust to illumination changes (automotive security context), and motion detection (visual surveillance). Finally, this thesis deals with the issue of segmentation - more precisely, the watershed algorithm. An efficient implementation of the watershed uses hierarchically ordered queues. However the original algorithm using this method leads to some bias. The algorithm proposed in this thesis corrects these biases. Again, a great benefit comes from the proposed framework and, as a result, the algorithm is bound neither to space nor to relief data: watersheds in 4D, on real or colour images are now possible. Region construction is then modified in order to have a finer control on segmentations from a few numbers of markers. The first modification brings external information (e.g. colour or statistical consistency) using a cost function, computed either on the contour or the interior of the region. The second modification is based on contours' local geometrical configuration and simulates a viscous flooding., Au cours de ces travaux de thèse, nous nous sommes intéressés d'un point de vue global aux algorithmes en Traitement d'Image et plus particulièrement en Morphologie Mathématique, selon certaines techniques nouvelles de programmation. L'évolution matérielle des moyens informatiques suit les prédictions de la loi de Moore. Cependant, une évolution parallèle, d'ordre logicielle, met à la disposition de la recherche scientifique des moyens de programmation nouveaux, dont la méta-programmation. Les avantages sont considérables, tant en terme scientifique par les possibilités offertes, qu'en termes simplement pratiques (portabilité, capitalisation des développements, réduction des erreurs, etc.). La présentation des travaux est structurée autour de la conception d'une bibliothèque de traitement - morphologique - d'image. Les différents aspects sont illustrés en partie par des exemples pris dans les domaines de la vidéosurveillance et de la sécurité automobile, et issus de projets industriels. Nous présentons dans un premier temps le cadre informatique utilisé pour l'écriture algorithmique. Afin de rendre efficace l'utilisation des nouvelles techniques de programmation, une étude préalable des notions mathématiques en Morphologie Mathématique - images, graphes, relations d'ordre, voisinages, éléments structurant, ... -, ainsi que des outils informatiques associés, est réalisée. La séparation correcte des rôles permet en outre l'écriture des structures indépendamment de la nature des données qu'ils contiennent, l'automatisation de nombreuses opérations par le compilateur, et une écriture algorithmique fidèle à une formulation mathématique. La conjonction de ces développements ouvre un grand champ d'exploration comme celui émanant des images nD et hyperspectrales, dont nous nous proposons d'explorer certains aspects. Le support des images nD associé à la programmation générique a sollicité le développement d'un algorithme de transformée exacte en distance. Les hypothèses sur la fonction distance sont faibles (homogénéité dans l'espace et convexité de la boule unité associée) afin d'utiliser les mêmes développements pour une large classe de fonction. Suite à une étude théorique, nous proposons un algorithme de calcul basé sur des propagations. Le même algorithme est utilisé pour l'ensemble des illustrations (fonctions de distance - L2, L5, orienté, non isométrique, ... - sur des images 4D). Les transformées morphologiques en distance sont d'approche totalement différente et d'usage courant en morphologie mathématique. Elles connaissent actuellement de nouveaux développements grâce à l'extension numérique proposée par Beucher: les « quasi-distances ». Nous proposons un algorithme de calcul rapide de ces distances. La couleur et plus généralement les images multispectrales (données vectorielles) sont d'une manipulation délicate en morphologie mathématique. Nous présentons trois approches complémentaires: l'utilisation de métriques couleurs, des statistiques locales et enfin les relations d'ordre lexicographique. Notre cadre informatique et algorithmique est parfaitement adapté à ces trois types de traitement. Le cadre métrique permet d'étendre la définition du gradient morphologique aux espaces couleurs, et plusieurs métriques dans Lab et HLS sont envisagées. Cette formulation est cependant coûteuse en termes de calcul et devient impraticable lorsque le voisinage utilisé pour le gradient s'agrandit. L'usage de statistiques locales permet de contourner ce problème. Nous nous sommes particulièrement intéressés à des statistiques circulaires dans HLS, ce qui nous a amené à la définition d'un gradient chromatique dans cet espace. Enfin, l'utilisation de relation d'ordre lexicographique étend le cadre algébrique classique à la couleur, sans modification fondamentale des algorithmes. Dans cette optique, nous verrons quels sont les moyens à notre disposition pour étendre la plupart des opérateurs (extrema, reconstruction, granulométries, ...) en maintenant un coût de développement bas. Deux études illustrent ces développements : la caractérisation chromatique de la peau, robuste aux changements d'illumination (contexte automobile), et la détection des zones de mouvement (vidéosurveillance). Le dernier sujet d'intérêt concerne la segmentation, et plus particulièrement l'algorithme de ligne de partage des eaux. L'implémentation de référence à l'aide de files d'attente hiérarchiques conduit à certains biais que nous corrigeons. L'algorithme proposé étant générique, nous l'appliquons sur des images de dimension 4, sur des reliefs en précision flottante ou couleur. Nous modifions ensuite la construction des bassins versants de manière à contourner certaines difficultés rencontrées lors de la segmentation avec un nombre faible de marqueurs. La première modification injecte dans le processus de propagation une information extérieure exprimée sous forme de fonction de coût. Cette fonction concerne aussi bien le contour que l'intérieur de la région en cours de construction. La seconde modification utilise une contrainte locale et rend le fluide visqueux. Des analogies sont établies entre ces nouvelles propagations et les équations d'évolution de courbe à l'aide de dérivées partielles.

International audience; Image segmentation on outdoor images is a difficult task because it comes up against the complexity and the variability of objects to detect and also of natural phenomena as shadows, highlights, partially overlapping objects. These difficulties conducts us to develop an adapted segmentation tool to this context. Furthermore, accordingly to the development of multispectral sensors, it seems interesting to propose a methodology adapted to multispectral images. Indeed, the segmentation process takes advantage of complementary informations. Our approach consists firstly in carrying out an image over-segmentation y use of a region growing algorithm and secondly, in labeling the obtained regions in c+1 classes with c equal to the number of class object to detect. The c+1 label allows us to distinguish regions having ambiguous properties (undetermined regions). To remove this ambiguity, a collaborative contour/region approach is applied on these undetermined regions. This approach has been tested on a set of cabbage images at different growth levels. In each image, all pixels have been labeled by hand for validation process. The results obtained by this method are compared with those obtained by a direct region growing segmentation.; L'imagerie multispectrale apparait comme un outil adapté à l'analyse des images prises en milieu extérieur. En effet, les procedés de segmentation peuvent etre basés sur les attributs des pixels qui reflètent certaines propriétés spectrales spécifiques des objets naturels. Cependant, certains outils de segmentation effectuent une classification directe des pixels dans l'espace des attributs sans tenir compte des relations spatiales. Nous proposons dans cet article une approche de segmentation d'images multispectrales prenant en compte cette dimension à un niveau région. Dans un premier temps, les régions homog´enes de l`images sont déterminées à l`aide d`un algorithme de segmentation type croissance de région. Les régions ainsi déterminées sont ensuite labellisées en tenant compte des attributs des pixels. Cependant, certaines d`entre elles restent indéterminées. Ces indéterminations seront par la suite levées à l'aide de règles de d´ecision prenant en compte les labels des r´egions adajacentes et leurs frontières communes. Les résultats obtenus sur des images RGB prisent en milieu extérieur pour la discrimination de plantes de choux et d`adventices sont ensuite présentées et discutées.

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior The work developed was built upon the following three basic steps: the first one related to the social, economic and cultural aspects of the settled people, to recover their origins, their dreams and intentions regarding the land. The second one related to data survey and environment analysis; and the third and last step integrates data and social, economic and cultural information with the environment to generate a pre-project of parcelling equivalently productive plots. The social, economic and cultural aspects, survied by means of questionnaires and interviews, showed that the majority of the settled people (94%) came from the Quebra Anzol microregion, from these, 88% were from the rural area. These data reinforce the fall of the mith of lack of familiarity with the hoe. The liking for farm working, the disemployment and the lack of land to work on were the main reasons for the settled people of Quebra Anzol to join the struggle for land movement. The research also showed that the delay to parcel out the land and the lack of rural credit are the greatest problems for the settlements. In relation to the intentions of land use, the main ones are agriculture, specially the cultivation of maize, beans, rice, coffee and fruit. Yet the hopes regarding the land are the possibility of a better life and helps for children s study. The survey and mapping of limitant environmental factors to agricultural soil use of the Quebra Anzol PA, used, as main data source, the field survey and the Landsat TM-7 data, which presented a good performance and quality in the classification process. The mapping of environmental factors showed that 96.74% of the area designed for the direct soil use have a strong fertility deficiency, which constitutes a restraint factor for use in the level A management, which is dependent on the natural soil fertility. The use of money and technology, anticipated for levels B and C managements, allows for the soil fertility improvement, increasing from 18.06 to 617.71 ha the area to used for crops. Nevertheless, the research points out the lack of capital for investiments as a factor which hinders the decrease of the limitation degree imposed by the deficiency in soil fertility. The existing soil fertility conditions and the lack of capital are a risk to the survival and permanence of the settled people in the Quebra Anzol PA, thus being a sound reason for a favourable credit policy and an effective establishement of actions for technical assistance to enable the survival, the permanence and the development of the settled people on the area. The collective representation, the space ordination, and the mapping of the areas s natural resources allowed the settled people to give their opinion, to suggest, to disagree, to agree, and, most of all, to know the steps of the process, promoting the satisfaction and the acceptance of the pre-project by the majority of the settled people. The integration of data and information supplied by the geographic information system allowed that, along the research, data and information given by the settled people could be aggregated to the process, adding practicality and eliminating the subjectivity in the elaboration of basic maps. The pre-project worked out together with the settled people was delivered to the INCRA, being approved and established. O trabalho desenvolvido compôs-se de três etapas básicas, sendo a primeira referente aos aspectos sócio-econômicos e culturais do assentado, o resgate da sua origem, seu sonho e pretensão para com a terra; a segunda refere-se ao levantamento de dados e análise do ambiente, e a terceira e última etapa integra os dados e informações sócio- econômicas e culturais com o ambiente objetivando a geração de um anteprojeto de parcelamento com lotes equiprodutivos. Os aspectos sócio-econômicos e culturais, levantados por meio de questionários e entrevistas, permitiram observar que a maioria dos assentados (94%) é proveniente da microrregião do Quebra Anzol, e 88% têm origem no campo, dados que reforçam a queda do mito da falta de intimidade do assentado para com a enxada. O gosto do trabalho pela terra, o desemprego e a falta de terra para trabalhar foram os principais motivos que levaram os assentados do Quebra Anzol a aderirem ao movimento de luta pela terra. A pesquisa também aponta a demora do parcelamento e a falta de crédito rural como os maiores problemas do assentamento. Em relação as pretensões de uso da terra, estas referem-se com maior freqüência a agricultura, sobressaindo-se a cultura de milho, feijão, arroz, café e pomar. Já as expectativas em relação à terra referemse a possibilidade de uma vida melhor e de ajuda nos estudos dos filhos. O levantamento e mapeamento dos fatores ambientais limitantes ao uso agrícola no solo do PA Quebra Anzol utilizou como principais fontes de dados o levantamento de campo e a imagem TM/Landsat-7, a qual apresentou bom desempenho e qualidade no processo de classificação. O mapeamento dos fatores ambientais mostrou que 96,74% da área destinada ao uso direto do solo apresenta forte deficiência de fertilidade, constituindo fator de restrição de uso no manejo de nível A, dependente da fertilidade natural dos solos. O emprego de capital e tecnologia, previstos para os níveis de manejo B e C, torna viável a melhoria da fertilidade do solo, aumentando de 18,06 ha para 617,71 ha a área com aptidão para lavoura. No entanto, a pesquisa aponta a falta de capital para investimento como fator que inviabiliza a diminuição do grau de limitação imposto pela deficiência de fertilidade. As atuais condições de fertilidade do solo e a falta de capital comprometem a sobrevivência e a permanência dos assentados no PA Quebra Anzol, razão pela qual prima-se por uma política de crédito favorável e uma implantação efetiva de ações relacionadas à assistência técnica, tornando possível a sobrevivência, a permanência e o desenvolvimento do assentados na área. A representação coletiva, a ordenação do espaço, e o mapeamento dos recursos naturais da área permitiram aos assentados opinar, sugerir, discordar, concordar e, sobretudo, conhecer as etapas do processo, promovendo a satisfação e a aceitação do anteprojeto pela maioria dos assentados. A integração dos dados e informações, promovida pelo sistema de informação geográfica, permitiu que ao longo da pesquisa, dados e informações fornecidos pelos assentados, fossem agregados ao processo, fornecendo praticidade e eliminando a subjetividade na elaboração de mapas básicos. O anteprojeto elaborado em conjunto com os assentados foi entregue ao INCRA, sendo na seqüência aprovado e implantado.

In this work, a multi-resolution procedure based on a generalized Laplacian pyramid (GLP), with p : q (i.e. rational) scale factor, is proposed to merge image data of any resolution and represent them at any scale. The GLP-based data fusion is shown to be slightly superior to those of a similar scheme based on the discrete wavelet transform, (WT) according to a set of parameters established in the literature. Not only fused images look sharper than their original versions, but also textured regions are enhanced without losing their spectral signatures. The pyramid-generating filters can be easily designed for data of any resolutions, differently from the WT, whose filter-bank design is non-trivial when the ratio between the scales of the images to be merged is not a power of two. Eventually, remotely sensed images from LandSat TM and from Panchromatic SPOT are fused together. The resulting bands capture multi-spectral features with enhanced contrast and texture, and an increased spatial resolution, thereby expediting automatic analyses for contextual interpretation of the environment.

In the last years it has become possible to combine high spatial resolution imagery, quick turnaround times and low operational complexity and costs in order to generate useful remote sensing products for vegetation monitoring applied to viticulture and the environment. Remote sensing platforms based on unmanned aerial vehicles (UAVs) represent a tool for this purpose, providing low-cost approaches to meet the critical requirements of spatial, spectral, and temporal resolutions. An open-source/low-cost UAV has been developed and tested in real field conditions proving to be an effective and inexpensive tool for mapping vineyard vigour areas of vineyards and thereby providing information relevant for precision viticulture applications. Two additional preliminary research activities have tested the capacity of the UAV system to evaluate diseases and bio char effects on vine vigour. Future prospects point to, the development of advanced UAV systems that allow a greater payload, and the consequent integration with more complex sensors in terms of weight and operation, such as thermal and hyper spectral cameras.