Digit recognition using trispectral features

Features derived from the trispectra of DFT magnitude slices are used for multi-font digit recognition. These features are insensitive to translation, rotation, or scaling of the input. They are also robust to noise. Classification accuracy tests were conducted on a common data base of 256× 256 pixel bilevel images of digits in 9 fonts. Randomly rotated and translated noisy versions were used for training and testing. The results indicate that the trispectral features are better than moment invariants and affine moment invariants. They achieve a classification accuracy of 95% compared to about 81% for Hu's (1962) moment invariants and 39% for the Flusser and Suk (1994) affine moment invariants on the same data in the presence of 1% impulse noise using a 1-NN classifier. For comparison, a multilayer perceptron with no normalization for rotations and translations yields 34% accuracy on 16× 16 pixel low-pass filtered and decimated versions of the same data.

Pattern recognition using invariants defined from higher order spectra - one-dimensional inputs

A new approach to pattern recognition using invariant parameters based on higher order spectra is presented. In particular, invariant parameters derived from the bispectrum are used to classify one-dimensional shapes. The bispectrum, which is translation invariant, is integrated along straight lines passing through the origin in bifrequency space. The phase of the integrated bispectrum is shown to be scale and amplification invariant, as well. A minimal set of these invariants is selected as the feature vector for pattern classification, and a minimum distance classifier using a statistical distance measure is used to classify test patterns. The classification technique is shown to distinguish two similar, but different bolts given their one-dimensional profiles. Pattern recognition using higher order spectral invariants is fast, suited for parallel implementation, and has high immunity to additive Gaussian noise. Simulation results show very high classification accuracy, even for low signal-to-noise ratios.

Toward automated power line corridor monitoring using advanced aircraft control and multisource feature fusion

The conventional manual power line corridor inspection processes that are used by most energy utilities are labor-intensive, time consuming and expensive. Remote sensing technologies represent an attractive and cost-effective alternative approach to these monitoring activities. This paper presents a comprehensive investigation into automated remote sensing based power line corridor monitoring, focusing on recent innovations in the area of increased automation of fixed-wing platforms for aerial data collection, and automated data processing for object recognition using a feature fusion process. Airborne automation is achieved by using a novel approach that provides improved lateral control for tracking corridors and automatic real-time dynamic turning for flying between corridor segments, we call this approach PTAGS. Improved object recognition is achieved by fusing information from multi-sensor (LiDAR and imagery) data and multiple visual feature descriptors (color and texture). The results from our experiments and field survey illustrate the effectiveness of the proposed aircraft control and feature fusion approaches.

Supervised latent dirichlet allocation models for efficient activity representation

Local spatio-temporal features with a Bag-of-visual words model is a popular approach used in human action recognition. Bag-of-features methods suffer from several challenges such as extracting appropriate appearance and motion features from videos, converting extracted features appropriate for classification and designing a suitable classification framework. In this paper we address the problem of efficiently representing the extracted features for classification to improve the overall performance. We introduce two generative supervised topic models, maximum entropy discrimination LDA (MedLDA) and class- specific simplex LDA (css-LDA), to encode the raw features suitable for discriminative SVM based classification. Unsupervised LDA models disconnect topic discovery from the classification task, hence yield poor results compared to the baseline Bag-of-words framework. On the other hand supervised LDA techniques learn the topic structure by considering the class labels and improve the recognition accuracy significantly. MedLDA maximizes likelihood and within class margins using max-margin techniques and yields a sparse highly discriminative topic structure; while in css-LDA separate class specific topics are learned instead of common set of topics across the entire dataset. In our representation first topics are learned and then each video is represented as a topic proportion vector, i.e. it can be comparable to a histogram of topics. Finally SVM classification is done on the learned topic proportion vector. We demonstrate the efficiency of the above two representation techniques through the experiments carried out in two popular datasets. Experimental results demonstrate significantly improved performance compared to the baseline Bag-of-features framework which uses kmeans to construct histogram of words from the feature vectors.

Enhancing human action recognition with region proposals

Deep convolutional network models have dominated recent work in human action recognition as well as image classification. However, these methods are often unduly influenced by the image background, learning and exploiting the presence of cues in typical computer vision datasets. For unbiased robotics applications, the degree of variation and novelty in action backgrounds is far greater than in computer vision datasets. To address this challenge, we propose an “action region proposal” method that, informed by optical flow, extracts image regions likely to contain actions for input into the network both during training and testing. In a range of experiments, we demonstrate that manually segmenting the background is not enough; but through active action region proposals during training and testing, state-of-the-art or better performance can be achieved on individual spatial and temporal video components. Finally, we show by focusing attention through action region proposals, we can further improve upon the existing state-of-the-art in spatio-temporally fused action recognition performance.

Activity recognition using binary tree SVM

This paper presents an effective classification method based on Support Vector Machines (SVM) in the context of activity recognition. Local features that capture both spatial and temporal information in activity videos have made significant progress recently. Efficient and effective features, feature representation and classification plays a crucial role in activity recognition. For classification, SVMs are popularly used because of their simplicity and efficiency; however the common multi-class SVM approaches applied suffer from limitations including having easily confused classes and been computationally inefficient. We propose using a binary tree SVM to address the shortcomings of multi-class SVMs in activity recognition. We proposed constructing a binary tree using Gaussian Mixture Models (GMM), where activities are repeatedly allocated to subnodes until every new created node contains only one activity. Then, for each internal node a separate SVM is learned to classify activities, which significantly reduces the training time and increases the speed of testing compared to popular the `one-against-the-rest' multi-class SVM classifier. Experiments carried out on the challenging and complex Hollywood dataset demonstrates comparable performance over the baseline bag-of-features method.

Spectral unmixing with estimated adaptive endmember index using Extended Support Vector Machine

The most difficult operation in the flood inundation mapping using optical flood images is to separate fully inundated areas from the ‘wet’ areas where trees and houses are partly covered by water. This can be referred as a typical problem the presence of mixed pixels in the images. A number of automatic information extraction image classification algorithms have been developed over the years for flood mapping using optical remote sensing images. Most classification algorithms generally, help in selecting a pixel in a particular class label with the greatest likelihood. However, these hard classification methods often fail to generate a reliable flood inundation mapping because the presence of mixed pixels in the images. To solve the mixed pixel problem advanced image processing techniques are adopted and Linear Spectral unmixing method is one of the most popular soft classification technique used for mixed pixel analysis. The good performance of linear spectral unmixing depends on two important issues, those are, the method of selecting endmembers and the method to model the endmembers for unmixing. This paper presents an improvement in the adaptive selection of endmember subset for each pixel in spectral unmixing method for reliable flood mapping. Using a fixed set of endmembers for spectral unmixing all pixels in an entire image might cause over estimation of the endmember spectra residing in a mixed pixel and hence cause reducing the performance level of spectral unmixing. Compared to this, application of estimated adaptive subset of endmembers for each pixel can decrease the residual error in unmixing results and provide a reliable output. In this current paper, it has also been proved that this proposed method can improve the accuracy of conventional linear unmixing methods and also easy to apply. Three different linear spectral unmixing methods were applied to test the improvement in unmixing results. Experiments were conducted in three different sets of Landsat-5 TM images of three different flood events in Australia to examine the method on different flooding conditions and achieved satisfactory outcomes in flood mapping.

Mixed pixel analysis and assessment for flood mapping

The most difficult operation in flood inundation mapping using optical flood images is to map the ‘wet’ areas where trees and houses are partly covered by water. This can be referred to as a typical problem of the presence of mixed pixels in the images. A number of automatic information extracting image classification algorithms have been developed over the years for flood mapping using optical remote sensing images, with most labelling a pixel as a particular class. However, they often fail to generate reliable flood inundation mapping because of the presence of mixed pixels in the images. To solve this problem, spectral unmixing methods have been developed. In this thesis, methods for selecting endmembers and the method to model the primary classes for unmixing, the two most important issues in spectral unmixing, are investigated. We conduct comparative studies of three typical spectral unmixing algorithms, Partial Constrained Linear Spectral unmixing, Multiple Endmember Selection Mixture Analysis and spectral unmixing using the Extended Support Vector Machine method. They are analysed and assessed by error analysis in flood mapping using MODIS, Landsat and World View-2 images. The Conventional Root Mean Square Error Assessment is applied to obtain errors for estimated fractions of each primary class. Moreover, a newly developed Fuzzy Error Matrix is used to obtain a clear picture of error distributions at the pixel level. This thesis shows that the Extended Support Vector Machine method is able to provide a more reliable estimation of fractional abundances and allows the use of a complete set of training samples to model a defined pure class. Furthermore, it can be applied to analysis of both pure and mixed pixels to provide integrated hard-soft classification results. Our research also identifies and explores a serious drawback in relation to endmember selections in current spectral unmixing methods which apply fixed sets of endmember classes or pure classes for mixture analysis of every pixel in an entire image. However, as it is not accurate to assume that every pixel in an image must contain all endmember classes, these methods usually cause an over-estimation of the fractional abundances in a particular pixel. In this thesis, a subset of adaptive endmembers in every pixel is derived using the proposed methods to form an endmember index matrix. The experimental results show that using the pixel-dependent endmembers in unmixing significantly improves performance.

Characterisation of Landscape with Forest Fragmentation Dynamics

Land cover (LC) and land use (LU) dynamics induced by human and natural processes play a major role in global as well as regional patterns of landscapes influencing biodiversity, hydrology, ecology and climate. Changes in LC features resulting in forest fragmentations have posed direct threats to biodiversity, endangering the sustainability of ecological goods and services. Habitat fragmentation is of added concern as the residual spatial patterns mitigate or exacerbate edge effects. LU dynamics are obtained by classifying temporal remotely sensed satellite imagery of different spatial and spectral resolutions. This paper reviews five different image classification algorithms using spatio-temporal data of a temperate watershed in Himachal Pradesh, India. Gaussian Maximum Likelihood classifier was found to be apt for analysing spatial pattern at regional scale based on accuracy assessment through error matrix and ROC (receiver operating characteristic) curves. The LU information thus derived was then used to assess spatial changes from temporal data using principal component analysis and correspondence analysis based image differencing. The forest area dynamics was further studied by analysing the different types of fragmentation through forest fragmentation models. The computed forest fragmentation and landscape metrics show a decline of interior intact forests with a substantial increase in patch forest during 1972-2007.

Gestão Ambiental apoiada por SIG na estrada da Colônia, Ilha Grande (RJ): uma visão do futuro.

Os SIG Sistemas de Informação Geográfica vêm sendo cada vez mais estudados como ferramentas facilitadoras de análises territoriais com o objetivo de subsidiar a gestão ambiental. A Ilha Grande, que pertence ao município de Angra dos Reis, localiza-se na baía de Ilha Grande no sul do estado do Rio de Janeiro e constitui-se no recorte espacial de análise. Apresenta uma dinâmica ambiental complexa que se sobrepõem principalmente aos usos de proteção ambiental e de atividade turística em uma porção do território em que as normatizações legais são difíceis de serem aplicadas, pois são reflexos de interesses que se manifestam em três esferas do poder a municipal, a estadual e a federal. O objetivo principal desta pesquisa é a realização do processamento digital de imagem para auxiliar a gestão territorial da Ilha Grande. Em foco, a estrada Abraão - Dois Rios, que liga Abraão (local de desembarque dos turistas, principal núcleo da Ilha) a Dois Rios (local de visitação por estudantes e pesquisadores, núcleo que abrigava o presídio, atualmente abriga sede do centro de pesquisa e museu da Universidade do Estado do Rio de Janeiro), ambos protegidos por diferentes categorias de unidades de conservação. A metodologia fundamenta-se no processamento digital de imagem através da segmentação e da classificação supervisionada por pixel e por região. O processamento deu-se a partir da segmentação (divisão de uma imagem digital em múltiplas regiões ou objetos, para simplificar e/ou mudar a representação de uma imagem) e dos processos de classificações de imagem, com a utilização de classificação por pixel e classificação por regiões (com a utilização do algoritmo Bhattacharya). As segmentações e classificações foram processadas no sistema computacional SPRING versão 5.1.7 e têm como objetivo auxiliar na análise de uso da Terra e projetar cenários a partir da identificação dos pontos focais de fragilidade encontrados ao longo da estrada Abraão-Dois Rios, propensos a ocorrências de movimentos de massa e que potencializam o efeito de borda da floresta e os impactos ambientais. A metodologia utilizada baseou-se em análise de campo e comparações de tecnologias de classificação de imagens. Essa estrada eixo de ligação entre os dois núcleos tem significativa importância na história da Ilha, nela circulam veículos, pesados e leves, de serviço, pedestres e turistas. Como resultados da presente foram gerados os mapas de classificação por pixel, os mapas de classificação por região, o mapa fuzzy com a intersecção dos mapas de classificação supervisionada por região e os mapas com os locais coletados em campo onde são verificadas ocorrências de movimentos de massa nas imagens ALOS, 2000, IKONOS, 2003 e ortofotografias, 2006. Esses mapas buscam servir de apoio à tomada de decisões por parte dos órgãos locais responsáveis.

Applicabilité de la texture couleur à la différentiation des classes d’occupation du territoire sur des images satellitales multispectrales

La texture est un élément clé pour l’interprétation des images de télédétection à fine résolution spatiale. L’intégration de l’information texturale dans un processus de classification automatisée des images se fait habituellement via des images de texture, souvent créées par le calcul de matrices de co-occurrences (MCO) des niveaux de gris. Une MCO est un histogramme des fréquences d’occurrence des paires de valeurs de pixels présentes dans les fenêtres locales, associées à tous les pixels de l’image utilisée; une paire de pixels étant définie selon un pas et une orientation donnés. Les MCO permettent le calcul de plus d’une dizaine de paramètres décrivant, de diverses manières, la distribution des fréquences, créant ainsi autant d’images texturales distinctes. L’approche de mesure des textures par MCO a été appliquée principalement sur des images de télédétection monochromes (ex. images panchromatiques, images radar monofréquence et monopolarisation). En imagerie multispectrale, une unique bande spectrale, parmi celles disponibles, est habituellement choisie pour générer des images de texture. La question que nous avons posée dans cette recherche concerne justement cette utilisation restreinte de l’information texturale dans le cas des images multispectrales. En fait, l’effet visuel d’une texture est créé, non seulement par l’agencement particulier d’objets/pixels de brillance différente, mais aussi de couleur différente. Plusieurs façons sont proposées dans la littérature pour introduire cette idée de la texture à plusieurs dimensions. Parmi celles-ci, deux en particulier nous ont intéressés dans cette recherche. La première façon fait appel aux MCO calculées bande par bande spectrale et la seconde utilise les MCO généralisées impliquant deux bandes spectrales à la fois. Dans ce dernier cas, le procédé consiste en le calcul des fréquences d’occurrence des paires de valeurs dans deux bandes spectrales différentes. Cela permet, en un seul traitement, la prise en compte dans une large mesure de la « couleur » des éléments de texture. Ces deux approches font partie des techniques dites intégratives. Pour les distinguer, nous les avons appelées dans cet ouvrage respectivement « textures grises » et « textures couleurs ». Notre recherche se présente donc comme une analyse comparative des possibilités offertes par l’application de ces deux types de signatures texturales dans le cas spécifique d’une cartographie automatisée des occupations de sol à partir d’une image multispectrale. Une signature texturale d’un objet ou d’une classe d’objets, par analogie aux signatures spectrales, est constituée d’une série de paramètres de texture mesurés sur une bande spectrale à la fois (textures grises) ou une paire de bandes spectrales à la fois (textures couleurs). Cette recherche visait non seulement à comparer les deux approches intégratives, mais aussi à identifier la composition des signatures texturales des classes d’occupation du sol favorisant leur différentiation : type de paramètres de texture / taille de la fenêtre de calcul / bandes spectrales ou combinaisons de bandes spectrales. Pour ce faire, nous avons choisi un site à l’intérieur du territoire de la Communauté Métropolitaine de Montréal (Longueuil) composé d’une mosaïque d’occupations du sol, caractéristique d’une zone semi urbaine (résidentiel, industriel/commercial, boisés, agriculture, plans d’eau…). Une image du satellite SPOT-5 (4 bandes spectrales) de 10 m de résolution spatiale a été utilisée dans cette recherche. Puisqu’une infinité d’images de texture peuvent être créées en faisant varier les paramètres de calcul des MCO et afin de mieux circonscrire notre problème nous avons décidé, en tenant compte des études publiées dans ce domaine : a) de faire varier la fenêtre de calcul de 3*3 pixels à 21*21 pixels tout en fixant le pas et l’orientation pour former les paires de pixels à (1,1), c'est-à-dire à un pas d’un pixel et une orientation de 135°; b) de limiter les analyses des MCO à huit paramètres de texture (contraste, corrélation, écart-type, énergie, entropie, homogénéité, moyenne, probabilité maximale), qui sont tous calculables par la méthode rapide de Unser, une approximation des matrices de co-occurrences, c) de former les deux signatures texturales par le même nombre d’éléments choisis d’après une analyse de la séparabilité (distance de Bhattacharya) des classes d’occupation du sol; et d) d’analyser les résultats de classification (matrices de confusion, exactitudes, coefficients Kappa) par maximum de vraisemblance pour conclure sur le potentiel des deux approches intégratives; les classes d’occupation du sol à reconnaître étaient : résidentielle basse et haute densité, commerciale/industrielle, agricole, boisés, surfaces gazonnées (incluant les golfs) et plans d’eau. Nos principales conclusions sont les suivantes a) à l’exception de la probabilité maximale, tous les autres paramètres de texture sont utiles dans la formation des signatures texturales; moyenne et écart type sont les plus utiles dans la formation des textures grises tandis que contraste et corrélation, dans le cas des textures couleurs, b) l’exactitude globale de la classification atteint un score acceptable (85%) seulement dans le cas des signatures texturales couleurs; c’est une amélioration importante par rapport aux classifications basées uniquement sur les signatures spectrales des classes d’occupation du sol dont le score est souvent situé aux alentours de 75%; ce score est atteint avec des fenêtres de calcul aux alentours de11*11 à 15*15 pixels; c) Les signatures texturales couleurs offrant des scores supérieurs à ceux obtenus avec les signatures grises de 5% à 10%; et ce avec des petites fenêtres de calcul (5*5, 7*7 et occasionnellement 9*9) d) Pour plusieurs classes d’occupation du sol prises individuellement, l’exactitude dépasse les 90% pour les deux types de signatures texturales; e) une seule classe est mieux séparable du reste par les textures grises, celle de l’agricole; f) les classes créant beaucoup de confusions, ce qui explique en grande partie le score global de la classification de 85%, sont les deux classes du résidentiel (haute et basse densité). En conclusion, nous pouvons dire que l’approche intégrative par textures couleurs d’une image multispectrale de 10 m de résolution spatiale offre un plus grand potentiel pour la cartographie des occupations du sol que l’approche intégrative par textures grises. Pour plusieurs classes d’occupations du sol un gain appréciable en temps de calcul des paramètres de texture peut être obtenu par l’utilisation des petites fenêtres de traitement. Des améliorations importantes sont escomptées pour atteindre des exactitudes de classification de 90% et plus par l’utilisation des fenêtres de calcul de taille variable adaptées à chaque type d’occupation du sol. Une méthode de classification hiérarchique pourrait être alors utilisée afin de séparer les classes recherchées une à la fois par rapport au reste au lieu d’une classification globale où l’intégration des paramètres calculés avec des fenêtres de taille variable conduirait inévitablement à des confusions entre classes.

On Recurrent and Deep Neural Networks

L'apprentissage profond est un domaine de recherche en forte croissance en apprentissage automatique qui est parvenu à des résultats impressionnants dans différentes tâches allant de la classification d'images à la parole, en passant par la modélisation du langage. Les réseaux de neurones récurrents, une sous-classe d'architecture profonde, s'avèrent particulièrement prometteurs. Les réseaux récurrents peuvent capter la structure temporelle dans les données. Ils ont potentiellement la capacité d'apprendre des corrélations entre des événements éloignés dans le temps et d'emmagasiner indéfiniment des informations dans leur mémoire interne. Dans ce travail, nous tentons d'abord de comprendre pourquoi la profondeur est utile. Similairement à d'autres travaux de la littérature, nos résultats démontrent que les modèles profonds peuvent être plus efficaces pour représenter certaines familles de fonctions comparativement aux modèles peu profonds. Contrairement à ces travaux, nous effectuons notre analyse théorique sur des réseaux profonds acycliques munis de fonctions d'activation linéaires par parties, puisque ce type de modèle est actuellement l'état de l'art dans différentes tâches de classification. La deuxième partie de cette thèse porte sur le processus d'apprentissage. Nous analysons quelques techniques d'optimisation proposées récemment, telles l'optimisation Hessian free, la descente de gradient naturel et la descente des sous-espaces de Krylov. Nous proposons le cadre théorique des méthodes à région de confiance généralisées et nous montrons que plusieurs de ces algorithmes développés récemment peuvent être vus dans cette perspective. Nous argumentons que certains membres de cette famille d'approches peuvent être mieux adaptés que d'autres à l'optimisation non convexe. La dernière partie de ce document se concentre sur les réseaux de neurones récurrents. Nous étudions d'abord le concept de mémoire et tentons de répondre aux questions suivantes: Les réseaux récurrents peuvent-ils démontrer une mémoire sans limite? Ce comportement peut-il être appris? Nous montrons que cela est possible si des indices sont fournis durant l'apprentissage. Ensuite, nous explorons deux problèmes spécifiques à l'entraînement des réseaux récurrents, à savoir la dissipation et l'explosion du gradient. Notre analyse se termine par une solution au problème d'explosion du gradient qui implique de borner la norme du gradient. Nous proposons également un terme de régularisation conçu spécifiquement pour réduire le problème de dissipation du gradient. Sur un ensemble de données synthétique, nous montrons empiriquement que ces mécanismes peuvent permettre aux réseaux récurrents d'apprendre de façon autonome à mémoriser des informations pour une période de temps indéfinie. Finalement, nous explorons la notion de profondeur dans les réseaux de neurones récurrents. Comparativement aux réseaux acycliques, la définition de profondeur dans les réseaux récurrents est souvent ambiguë. Nous proposons différentes façons d'ajouter de la profondeur dans les réseaux récurrents et nous évaluons empiriquement ces propositions.

Estimation of Effective Impervious Surface Area of Cochin using Satellite Images

Urbanization refers to the process in which an increasing proportion of a population lives in cities and suburbs. Urbanization fuels the alteration of the Land use/Land cover pattern of the region including increase in built-up area, leading to imperviousness of the ground surface. With increasing urbanization and population pressures; the impervious areas in the cities are increasing fast. An impervious surface refers to an anthropogenic ally modified surface that prevents water from infiltrating into the soil. Surface imperviousness mapping is important for the studies related to water cycling, water quality, soil erosion, flood water drainage, non-point source pollution, urban heat island effect and urban hydrology. The present study estimates the Total Impervious Area (TIA) of the city of Kochi using high resolution satellite image (LISS IV, 5m. resolution). Additionally the study maps the Effective Impervious Area (EIA) by coupling the capabilities of GIS and Remote Sensing. Land use/Land cover map of the study area was prepared from the LISS IV image acquired for the year 2012. The classes were merged to prepare a map showing pervious and impervious area. Supervised Maximum Likelihood Classification (Supervised MLC),which is a simple but accurate method for image classification, is used in calculating TIA and an overall classification accuracy of 86.33% was obtained. Water bodies are 100% pervious, whereas urban built up area are 100% impervious. Further based on percentage of imperviousness, the Total Impervious Area is categorized into various classes

Semantic-level Visual Content Descriptor combining visual and textual cues as collaterals

In this paper, we introduce a novel high-level visual content descriptor which is devised for performing semantic-based image classification and retrieval. The work can be treated as an attempt to bridge the so called “semantic gap”. The proposed image feature vector model is fundamentally underpinned by the image labelling framework, called Collaterally Confirmed Labelling (CCL), which incorporates the collateral knowledge extracted from the collateral texts of the images with the state-of-the-art low-level image processing and visual feature extraction techniques for automatically assigning linguistic keywords to image regions. Two different high-level image feature vector models are developed based on the CCL labelling of results for the purposes of image data clustering and retrieval respectively. A subset of the Corel image collection has been used for evaluating our proposed method. The experimental results to-date already indicates that our proposed semantic-based visual content descriptors outperform both traditional visual and textual image feature models.