Inge Gavat

Contextual Descriptors for Scene Classes in Very High Resolution SAR Images

The new generation of spaceborne SAR instruments with meter or submeter resolution finds enormous applications for the observation of urban, industrial, in general of man-made scenes. Thus, targets are not any more observed in isolation, instead the groups of objects, e.g., house, bridge, and road, etc., need to be recognized in their spatial context. This paper proposes a feature extraction method for image patches in order to capture the spatial context. The method is based on the characteristics of the spectra of the SAR data, integrating radiometric, geometric, and texture properties of the SAR image patch. The method is demonstrated for TerraSAR-X High Resolution Spotlight data. To account for the spatial context in which a group of targets is located, it uses an image patch covering typically 200 × 200m2 of the scene. A comparative evaluation of our descriptors and grey-level co-occurrence matrix (GLCM) texture features has been performed over a database of 6916 patches. The method allowed for the robust recognition of over 30 different scene classes, with precision between 50% and 93%. Numerical results show that our method is able to discriminate between scene classes better than GLCM texture parameters.

A Latent Analysis of Earth Surface Dynamic Evolution Using Change Map Time Series

With a continuous increase in the number of Earth Observation satellites, leading to the development of satellite image time series (SITS), the number of algorithms for land cover analysis and monitoring has greatly expanded. This paper offers a new perspective in dynamic classification for SITS. Four similarity measures (correlation coefficient, Kullback-Leibler divergence, conditional information, and normalized compression distance) based on consecutive image pairs from the data are employed. These measures employ linear dependences, statistical measures, and spatial relationships to compute radiometric, spectral, and texture changes that offer a description for the multitemporal behavior of the SITS. During this process, the original SITS is converted to a change map time series (CMTS), which removes the static information from the data set. The CMTS is analyzed using a latent Dirichlet allocation (LDA) model capable of discovering classes with semantic meaning based on the latent information hidden in the scene. This statistical method was originally used for text classification, thus requiring a word, document, corpus analogy with the elements inside the image. The experimental results were computed using 11 Landsat images over the city of Bucharest and surrounding areas. The LDA model enables us to discover a wide range of scene evolution classes based on the various dynamic behaviors of the land cover. The results are compared with the Corinne Land Cover map. However, this is not a validation method but one that adds static knowledge about the general usage of the analyzed area. In order to help the interpretation of the results, we use several studies on forms of relief, weather forecast, and very high resolution images that can explain the wide range of structures responsible for influencing the dynamic inside the resolution cell.

Complex SAR image characterization using space variant spectral analysis

This paper proposes a new parameter based method of SAR image feature extraction and complex image information retrieval. The methodpsilas groundwork is the Fast Fourier Transform, each of the proposed parameters being built on a Fourier Transform basis. We suggest that by the use of several image bands formed of distinct spectral signatures of the original complex image, one can obtain a valid spectral characterization of the SAR image that can be afterwards subject to a clustering algorithm. The classification algorithm proposed in this paper is unsupervised K- means. The main advantages of the algorithm are the simplicity and robustness of the implementation.

A Comparative Study of Feature Extraction Methods Applied to Continuous Speech Recognition in Romanian Language

This paper describes continuous speech recognition experiments on a Romanian language speech database, by using hidden Markov models (EMM). We compare the recognition rates obtained in our ASR system realising front-ends based on features extracted by perceptual variants of cepstral analysis and linear prediction and by simple linear prediction. The best results obtained with 36 coefficients mel-frequency cepstral coefficients (MFCC) are used as basis to rank the front-ends based on LPC. The second rank is very promising for the performance obtained with 5 perceptual linear prediction (PLP) coefficients, obviously better at the last ranked performance of the simple linear prediction coefficients (LPC). We reorganized the database as follows: one database for male speakers, one database for female speakers and one database for both male and female speakers

Mutual information based measure for image content characterization

An image can be decomposed into different elementary descriptors depending on the observer interest. Similar techniques as used to understand words, regarded as molecules, formed by combining atoms, are proposed to describe images based on their information content. In this paper, we use primitive feature extraction and clustering to code the image information content. Our purpose is to describe the complexity of the information based on the combinational profile of the clustered primitive features using entropic measures like mutual information and Kullback-Leibler divergence. The developed method is demonstrated to asses image complexity for further applications to improve Earth Observation image analysis for sustainable humanitarian crisis response in risk reduction.

The ASRS_RL --- A Research Platform for Spoken Language Recognition and Understanding Experiments

In the present paper are described the component stages implemented on the Automatic Speech Recognition System for Romanian Language, ASRS_RL and the speech recognition experiments made with the system. The discussed stages are the feature extraction based on cepstral analysis and linear prediction and the learning strategies implemented on the system for acoustical modelling. The first goal of our research was to develop and experiment a system for continuous speech recognition and understanding in the statistical framework of hidden Markov models (HMM). A second goal was to asses other learning strategies like support vector machines (SVM), artificial neural networks (ANN), and some hybrid structures as alternatives for the HMM. The made experiments in isolated words (digits) and vowels recognition gave good results and proved the competitively of the applied new structures.

Damage Assessment Based on SAR Image Analysis: Flood Scenario for Romanian Eastern Carpathian Region

When natural disasters occur, it is necessary for the authorities to make fast and effective decisions in order to prevent the occurrence of more damage, as well as to find solutions for the affected population that needs to be relocated. Satellite imagery can prove to be a useful instrument in decision support during emergency situations of such nature (floods), and especially SAR data, due to its all weather capabilities. This paper makes an assessment of the utility of satellite radar products (TerraSAR-X and Radarsat) in the frame of emergency situations management. A real case study is presented, where radar data were processed by human specialists on one hand, and automatically on the other hand, using an intelligent information extraction system.

Progress in Speech Recognition for Romanian Language

In this chapter we will present the progress made in automatic speech recognition for Romanian language based on the ASRS_RL (Automatic Speech Recognition System for Romanian Language) research platform. Speech recognition is a research domain with a long history, but despite this fact, still open for new investigations and answers to the not yet finally solved questions. This situation can be explained by the difficulty of the task, underlying on the fact that speech is a human product, with a high degree of correlation in content, but with a great variability in the formal manifestation as an acoustic signal. Great difficulties cause also the imperfection of the audio chain and the noises in the environment. The best-known strategies for speech recognition are the statistical and the connectionist ones, but fuzzy sets can also play an important role. Based on HMM’s the statistical strategies have many advantages, among them being recalled: rich mathematical framework, powerful learning and decoding methods, good sequences handling capabilities, flexible topology for statistical phonology and syntax. The disadvantages lie in the poor discrimination between the models and in the unrealistic assumptions that must be made to construct the HMM’s theory, namely the independence of the successive feature frames (input vectors) and the first order Markov process. Based on artificial neural networks (ANNs), the connectionist strategies for speech recognition have the advantages of the massive parallelism, good adaptation, efficient algorithms for solving classification problems and intrinsic discriminative properties. However, the neural nets have difficulties in handling the temporal dependencies inherent in speech data. The learning capabilities of the statistical and the neural models are very important, classifier built on such bases having the possibility to recognize new, unknown patterns with the experience obtained by training. The introduction of fuzzy sets allows on one hand the so-called fuzzy decisions, on other hand the “fuzzyfication” of input data, often more suitable for recognition of pattern produced by human beings, by speaking, for example. In a fuzzy decision, the recognizer realizes the classification based on the degree of membership to a given class for the pattern to be classified, a pattern belonging in a certain measure to each of the possible classes. This O pe n A cc es s D at ab as e w w w .ite ch on lin e. co m

Knowledge Resources in Automatic Speech Recognition and Understanding for Romanian Language

In this chapter are presented the results obtained in automatic speech recognition and understanding (ASRU) experiments made for Romanian language in the statistical framework, concerning the performance enhancement of two important knowledge resources, namely the acoustical models and the language model. If the ASRU process is for simplicity seen as a two stage process, in the first stage automatic speech recognition (ASR) is done and in the second stage the understanding is accomplished. The acoustical models incorporate knowledge about features statistic in different speech units composing the words and are mostly responsible for the performance of the recognition stage, judged after the WRR (word recognition rate). The language models incorporate knowledge about the word statistic in the phrase and determine mostly the performance of the understanding stage, judged after the PRR (phrase recognition rate). The two considered stages are interrelated and the named performance criteria are interdependent, enhanced WRR leads to PRR enhancement too. In this chapter are exposed methods to enhance the WRR, based on introducing of contextual models like triphones instead monophones or building of gender specialized models (for men, women and children) instead of global models. The methods applied to enhance the PRR are based on introducing of a restrictive finite state grammar instead the permissive word loop grammar or a bigram based language model.

Image disorder characterization based on rate distortion

Rate distortion theory is one of the areas of information transmission theory with important applications in multimodal signal processing, as for example image processing, information bottleneck and steganalysis. This article present an image characterization method based on rate distortion analysis in the feature space. This space is coded using clustering as vector quantization (k-means). Since image information usually cannot be coded by single clusters, because there are image regions corresponding to groups of clusters, the rate and distortion are specifically defined. The rate distortion curve is analyzed, extracting specific features for implementing a database image classification system.