Panagiotis Barmpoutis

Spatio-Temporal Flame Modeling and Dynamic Texture Analysis for Automatic Video-Based Fire Detection

Every year, a large number of wildfires all over the world burn forested lands, causing adverse ecological, economic, and social impacts. Beyond taking precautionary measures, early warning and immediate response are the only ways to avoid great losses. To this end, in this paper we propose a computer vision approach for fire-flame detection to be used by an early-warning fire monitoring system. Initially, candidate fire regions in a frame are defined using background subtraction and color analysis based on a nonparametric model. Subsequently, the fire behavior is modeled by employing various spatio-temporal features, such as color probability, flickering, spatial, and spatio-temporal energy, while dynamic texture analysis is applied in each candidate region using linear dynamical systems and a bag-of-systems approach. To increase the robustness of the algorithm, the spatio-temporal consistency energy of each candidate fire region is estimated by exploiting prior knowledge about the possible existence of fire in neighboring blocks from the current and previous video frames. As a final step, a two-class support vector machine classifier is used to classify the candidate regions. Experimental results have shown that the proposed method outperforms existing state-of-the-art algorithms.

Classification of Multidimensional Time-Evolving Data Using Histograms of Grassmannian Points

In this paper, we address the problem of classifying multidimensional time-evolving data in dynamic scenes. To take advantage of the correlation between the different channels of data, we introduce a generalized form of a stabilized higher order linear dynamical system (sh-LDS) and we represent the multidimensional signal as a third-order tensor. In addition, we show that the parameters of the proposed model lie on a Grassmann manifold and we attempt to address the classification problem through study of the geometric properties of the sh-LDS’s space. Moreover, to tackle the problem of nonlinearity of the observation data, we represent each multidimensional signal as a cloud of points on the Grassmann manifold and we create a codebook by identifying the most representative points. Finally, each multidimensional signal is classified by applying a bag-of-systems approach having first modeled the variation of the class of each codeword on its tangent space instead of the sh-LDS’s space. The proposed methodology is evaluated in three different application domains, namely, video-based surveillance systems, dynamic texture categorization, and human action recognition, showing its great potential.

Higher Order Linear Dynamical Systems for Smoke Detection in Video Surveillance Applications

In this paper, we consider the problem of multi-dimensional dynamic texture analysis, and we introduce a new higher order linear dynamical system (h-LDS) descriptor. The proposed h-LDS descriptor is based on the higher order decomposition of the multidimensional image data and enables the analysis of dynamic textures by using information from various image elements. In addition, we propose a methodology for its application to video-based early warning systems that focus on smoke identification. More specifically, the proposed methodology enables the representation of video subsequences as histograms of h-LDS descriptors produced by the smoke candidate image patches in each subsequence. Finally, to further improve the classification accuracy, we propose the combination of multidimensional dynamic texture analysis with the spatiotemporal modeling of smoke by using a particle swarm optimization approach. The ability of the h-LDS to analyze the dynamic texture information is evaluated through a multivariate comparison against the standard LDS descriptor. The experimental results that use two video datasets have shown the great potential of the proposed smoke detection method.

Grading of invasive breast carcinoma through Grassmannian VLAD encoding

In this paper we address the problem of automated grading of invasive breast carcinoma through the encoding of histological images as VLAD (Vector of Locally Aggregated Descriptors) representations on the Grassmann manifold. The proposed method considers each image as a set of multidimensional spatially-evolving signals that can be efficiently modeled through a higher-order linear dynamical systems analysis. Subsequently, each H&E (Hematoxylin and Eosin) stained breast cancer histological image is represented as a cloud of points on the Grassmann manifold, while a vector representation approach is applied aiming to aggregate the Grassmannian points based on a locality criterion on the manifold. To evaluate the efficiency of the proposed methodology, two datasets with different characteristics were used. More specifically, we created a new medium-sized dataset consisting of 300 annotated images (collected from 21 patients) of grades 1, 2 and 3, while we also provide experimental results using a large dataset, namely BreaKHis, containing 7,909 breast cancer histological images, collected from 82 patients, of both benign and malignant cases. Experimental results have shown that the proposed method outperforms a number of state of the art approaches providing average classification rates of 95.8% and 91.38% with our dataset and the BreaKHis dataset, respectively.

Extracting Dynamics from Multi-dimensional Time-evolving Data using a Bag of Higher-order Linear Dynamical Systems

In this paper we address the problem of extracting dynamics from multi-dimensional time-evolving data. To this end, we propose a linear dynamical model (LDS), which is based on the higher order decomposition of the observation data. In this way, we are able to extract a new descriptor for analyzing data of multiple elements coming from of the same or different data sources. Each sequence of data is modeled as a collection of higher order LDS descriptors (h-LDSs), which are estimated in equally sized temporal segments of data. Finally, each sequence is represented as a term frequency histogram following a bag-ofsystems approach, in which h-LDSs are used as feature descriptors. For evaluating the performance of the proposed methodology to extract dynamics from time evolving multidimensional data and using them for classification purposes in various applications, in this paper we consider two different cases: dynamic texture analysis and human motion recognition. Experimental results with two datasets for dynamic texture analysis and two datasets for human action recognition demonstrate the great potential of the proposed

Classification of Nuclei in Follicular Lyphoma Tissue Sections Using Different Stains and Bayesian Networks

Automated centroblast (CB) detection in Follicular Lymphoma (FL) tissue samples has recently attracted significant research interest. Most of the methods described in the literature are based on the use of Hematoxilin and Eosin (H&E) stain. However, the automated detection of CBs from H&E stained images remains a challenging issue. To this end, this paper presents a novel approach which is based on the use of both PAX5 and H&E stains in tissue sections sliced at the thickness of 1μm. The goal of PAX5 is three-fold: to facilitate the segmentation of nuclei, to remove a number of follicular dendritic cells and finally to extract morphological characteristics of nuclei. Furthermore, the use of H&E stain enables us to extract textural information related to histological characteristics used by pathologists in diagnosis of FL grading. In our method we propose a novel algorithm for the separation of overlapped nuclei inspired by the clustering of large scale visual vocabularies. Finally, aiming to model pathologists’ knowledge used in FL grading, we use a Bayesian Network classifier to combine the morphological and textural characteristics. Experiments conducted on a dataset of ten pairs of PAX5 and H&E images demonstrate the potential of the proposed approach providing an average detection rate of 93.46%.

Wood species recognition through multidimensional texture analysis

Abstract Wood recognition is a crucial task for wood sciences and industries, since it leads to the identification of the anatomical features and physical properties of wood. Traditionally, the recognition process relies almost exclusively on human experts, who are based on various characteristics of wood, such as color, structure and texture. However, there are numerous types of wood species in the nature that are difficult to be identified even by experienced scientists. Towards this end, in this paper we propose a novel approach for automated wood species recognition through multidimensional texture analysis. By taking advantage of the fact that static wood images contain periodic spatially-evolving characteristics, we introduce a new spatial descriptor considering each wood image as a collection of multidimensional signals. More specifically, the proposed methodology enables the representation of wood images as concatenated histograms of higher order linear dynamical systems produced by vertical and horizontal image patches. The final classification of images, i.e., histogram representations, into wood species, is performed using a Support Vector Machines (SVM) classifier. For the evaluation of the proposed method, a dataset, namely “WOOD-AUTH”, consisting of more than 4200 wood images (from cross, radial and tangential sections of normal wood structure) of twelve common wood species existing in Greek territory, was created. Experimental results presented in this paper show the great potential of the proposed methodology, which, despite a small number of misclassification cases with regards to both anatomically similar and different species, outperforms a number of state of the art approaches, yielding a classification rate of 91.47% in wood cross sections.