Aleksey Fadeev

Comparison of different classification algorithms for landmine detection using GPR

The Edge Histogram Detector (EHD) is a landmine detection algorithm that has been developed for ground penetrating radar (GPR) sensor data. It has been tested extensively and has demonstrated excellent performance. The EHD consists of two main components. The first one maps the raw data to a lower dimension using edge histogram based feature descriptors. The second component uses a possibilistic K-Nearest Neighbors (pK-NN) classifier to assign a confidence value. In this paper we show that performance of the baseline EHD could be improved by replacing the pK-NN classifier with model based classifiers. In particular, we investigate two such classifiers: Support Vector Regression (SVR), and Relevance Vector Machines (RVM). We investigate the adaptation of these classifiers to the landmine detection problem with GPR, and we compare their performance to the baseline EHD with a pK-NN classifier. As in the baseline EHD, we treat the problem as a two class classification problem: mine vs. clutter. Model parameters for the SVR and the RVM classifiers are estimated from training data using logarithmic grid search. For testing, soft labels are assigned to the test alarms. A confidence of zero indicates the maximum probability of being a false alarm. Similarly, a confidence of one represents the maximum probability of being a mine. Results on large and diverse GPR data collections show that the proposed modification to the classifier component can improve the overall performance of the EHD significantly.

Adaptive Edge Histogram Descriptor for Landmine Detection using GPR

The Edge Histogram Detector (EHD) is a landmine detection algorithm for sensor data generated by ground penetrating radar (GPR). It uses edge histograms for feature extraction and a possibilistic K-Nearest Neighbors (K-NN) rule for confidence assignment. To reduce the computational complexity of the EHD and improve its generalization, the K-NN classifier uses few prototypes that can capture the variations of the signatures within each class. Each of these prototypes is assigned a label in the class of mines and a label in the class of clutter to capture its degree of sharing among these classes. The EHD has been tested extensively. It has demonstrated excellent performance on large real world data sets, and has been implemented in real time versions in hand-held and vehicle mounted GPR. In this paper, we propose two modifications to the EHD to improve its performance and adaptability. First, instead of using a fixed threshold to decide if the edge at a certain location is strong enough, we use an adaptive threshold that is learned from the background surrounding the target. This modification makes the EHD more adaptive to different terrains and to mines buried at different depths. Second, we introduce an additional training component that tunes the prototype features and labels to different environments. Results on large and diverse GPR data collections show that the proposed adaptive EHD outperforms the baseline EHD. We also show that the edge threshold can vary significantly according to the edge type, alarm depth, and soil conditions.

Dominant Texture Descriptors for image classification and retrieval

In this paper, we propose a generic approach for representing image texture features in a compact and intuitive way. Our approach, called Dominant Texture Descriptor (DTD), is inspired by the dominant color descriptor. It is based on clustering the local texture features and identifying the dominant components and their spatial distribution. We also present an enhanced version of the DTD (eDTD) that encodes the spatial distribution of the pixels within each dominant component. We illustrate this approach for the case of two well-known descriptors, namely, the MPEG-7 Edge Histogram, and Ga- bor texture. The performance of the proposed texture feature representation is illustrated by using it to classify a collection of 900 color images. Experimental results are compared with those obtained using the traditional approaches. We show that our representation is more compact, interpretable, and could improve classification results by 10%-20%, especially for images with non-homogeneous texture.

Significance of MPEG-7 Textural Features for Improved Mass Detection in Mammography

The purpose of the study is to investigate the significance of MPEG-7 textural features for improving the detection of masses in screening mammograms. The detection scheme was originally based on morphological directional neighborhood features extracted from mammographic regions of interest (ROIs). Receiver operating characteristics (ROC) was performed to evaluate the performance of each set of features independently and merged into a back-propagation artificial neural network (BPANN) using the leave-one-out sampling scheme (LOOSS). The study was based on a database of 668 mammographic ROIs (340 depicting cancer regions and 328 depicting normal parenchyma). Overall, the ROC area index of the BPANN using the directional morphological features was Az=0.85plusmn0.01. The MPEG-7 edge histogram descriptor-based BPNN showed an ROC area index of Az=0.71plusmn0.01 while homogeneous textural descriptors using 30 and 120 channels helped the BPNN achieve similar ROC area indexes of Az=0.882plusmn0.02 and Az=0.877plusmn0.01 respectively. After merging the MPEG-7 homogeneous textural features with the directional neighborhood features the performance of the BPANN increased providing an ROC area index of Az=0.91plusmn0.01. MPEG-7 homogeneous textural descriptor significantly improved the morphology-based detection scheme

Quality evaluation of fast morphing interpolation model for 3D volume reconstruction

The purpose of this study is to evaluate a 3D volume reconstruction model for volume rendering. The model is conducted using brain MRI data of Visible Human Project. Particularly MRI T1 data were used. The quality of the developed model is compared with linear interpolation technique. By applying our morphing technique recursively, taking progressively smaller subregions within a region, a high quality and accuracy interpolation is obtained. The presented algorithm is robust and has 20 adjustable parameters for use with different modalities. The main advantages of this morphing algorithm are: 1) applicability to general configurations of planes in 3D space, 2) automated behavior, 3) applicability to CT scans with no changes in the algorithm and software. Subsequently, to visualize data, a specialized volume rendering card (TeraRecon VolumePro 1000) was used. To represent data in 3D space, special software was developed to convert interpolated CT slices to 3D objects compatible with the VolumePro card. Quantitative and visual comparison between the proposed model and linear interpolation clearly demonstrates the superiority of the proposed model. Evaluation is performed by removing slices from the original stack of 2D images and using them as reference for error comparison among alternative approaches. Error analysis using average Mean Square and Absolute error clearly demonstrates improved performance

Ensemble Possibilistic K-NN for Functional Clustering of Gene Expression Profiles in Human Cancers Challenge

This paper describes the Ensemble Possibilistic K-NN algorithm for classification of gene expression profiles into three major cancer categories. In fact, a modification of forward feature selection is proposed to identify relevant feature subsets allowing for multiple possibilistic K-nearest neighbors (pK-NNs) rule experts. First, individual features are ranked according to their performance on training data and subsets of features identified using greedy approach. Each subset has significantly lower dimensionality than the original feature vector. Second, each subset is associated with pK-NN expert and the final classification decision is based on combining results produced by all experts.

Dominant Audio Descriptors for Audio Classification and Retrieval

In this paper, we propose a new general low-level feature representation for audio signals. Our approach, called Dominant Audio Descriptor is inspired by the MPEG-7 Dominant Color Descriptor. It is based on clustering timelocal features and identifying dominant components. The features used to illustrate this approach are the well-known Mel Frequency Cepstral Coefficients. The performance of the proposed framework is evaluated on audio classification and retrieval tasks. In particular, the experiments are performed on a benchmark music data set. The results are compared to those previously obtained on the same data base. We show that our approach improved classification and retrieval results by more then 3%, and for the case of retrieval reached almost perfect retrieval rate of 99:36%. In addition, the paper presents comparative results against several state of the art classifiers, such as Hidden Markov Models, Support Vector Machines and k-Nearest Neighbors.

Semantic video indexing using context-dependent fusion

We present a novel method for fusing the results of multiple semantic video indexing algorithms that use different types of feature descriptors and different classification methods. This method, called Context-Dependent Fusion (CDF), is motivated by the fact that the relative performance of different semantic indexing methods can vary significantly depending on the video type, context information, and the high-level concept of the video segment to be labeled. The training part of CDF has two main components: context extraction and algorithm fusion. In context extraction, the low-level audio-visual descriptors used by the different classification algorithms are combined and used to partition the descriptors space into groups of similar video shots, or contexts. The algorithm fusion component identifies a subset of classification algorithms (local experts) for each context based on their relative performance within the context. Results on the TRECVID-2002 data collections show that the proposed method can identify meaningful and coherent clusters and that different labeling algorithms can be identified for the different contexts. Our initial experiments have indicated that the context-dependent fusion outperforms the individual algorithms. We also show that using simple visual descriptors and a simple K-NN classifier, the CDF approach provides results that are comparable to the state-of-the-art methods in semantic indexing.

Transformation of Relational Features for Use with Conventional Classifiers

In this paper, we address the problem of transforming relational features into an Euclidian space so that standard classification methods that assume that data is in a vector form could be used. Our approach has three main steps. First, a relational matrix that represents the pair-wise dissimilarities between all objects is constructed. Second, a fuzzy relational clustering algorithm is used to partition the data into groups of similar objects. Third, the relational data features are mapped to a unit hyper-cube space where each object is represented by its membership vectors in all clusters. The proposed method is validated by comparing the performance of several classifiers with different feature sets on the original and the transformed spaces. We show that the transformed space conserves the discriminative information of the original features. We also show that, using the transformed space, a richer set of standard classifiers could be used.

Adapted morphing model for 3D volume reconstruction applied to abdominal CT images

The purpose of this study was to develop a 3D volume reconstruction model for volume rendering and apply this model to abdominal CT data. The model development includes two steps: (1) interpolation of given data for a complete 3D model, and (2) visualization. First, CT slices are interpolated using a special morphing algorithm. The main idea of this algorithm is to take a region from one CT slice and locate its most probable correspondence in the adjacent CT slice. The algorithm determines the transformation function of the region in between two adjacent CT slices and interpolates the data accordingly. The most probable correspondence of a region is obtained using correlation analysis between the given region and regions of the adjacent CT slice. By applying this technique recursively, taking progressively smaller subregions within a region, a high quality and accuracy interpolation is obtained. The main advantages of this morphing algorithm are 1) its applicability not only to parallel planes like CT slices but also to general configurations of planes in 3D space, and 2) its fully automated nature as it does not require control points to be specified by a user compared to most morphing techniques. Subsequently, to visualize data, a specialized volume rendering card (TeraRecon VolumePro 1000) was used. To represent data in 3D space, special software was developed to convert interpolated CT slices to 3D objects compatible with the VolumePro card. Visual comparison between the proposed model and linear interpolation clearly demonstrates the superiority of the proposed model.