Ilya Levner

Classification-Driven Watershed Segmentation

This paper presents a novel approach for creation of topographical function and object markers used within watershed segmentation. Typically, marker-driven watershed segmentation extracts seeds indicating the presence of objects or background at specific image locations. The marker locations are then set to be regional minima within the topological surface (typically, the gradient of the original input image), and the watershed algorithm is applied. In contrast, our approach uses two classifiers, one trained to produce markers, the other trained to produce object boundaries. As a result of using machine-learned pixel classification, the proposed algorithm is directly applicable to both single channel and multichannel image data. Additionally, rather than flooding the gradient image, we use the inverted probability map produced by the second aforementioned classifier as input to the watershed algorithm. Experimental results demonstrate the superior performance of the classification-driven watershed segmentation algorithm for the tasks of 1) image-based granulometry and 2) remote sensing

Ore image segmentation by learning image and shape features

We present an image segmentation system specifically targeted for oil sand ore size estimation. The system learns spectral and shape characteristics of training images of oil sand ore samples for image quality enhancement followed by segmentation of ore image shapes. The proposed segmentation has achieved superior accuracy over the current state of the art systems.

Towards Automated Creation of Image Interpretation Systems

Automated image interpretation is an important task in numerous applications ranging from security systems to natural resource inventorization based on remote-sensing. Recently, a second generation of adaptive machine-learned image interpretation systems have shown expert-level performance in several challenging domains. While demonstrating an unprecedented improvement over hand-engineered and first generation machine-learned systems in terms of cross-domain portability, design-cycle time, and robustness, such systems are still severely limited. This paper inspects the anatomy of the state-of-the-art Multi resolution Adaptive Object Recognition framework (MR ADORE) and presents extensions that aim at removing the last vestiges of human intervention still present in the original design of ADORE. More specifically, feature selection is still a task performed by human domain experts and represents a major stumbling block in the creation process of fully autonomous image interpretation systems. This paper focuses on minimizing such need for human engineering. After discussing experimental results, showing the performance of the framework extensions in the domain of forestry, the paper concludes by outlining autonomous feature extraction methods that may completely remove the need for human expertise in the feature selection process.

Feature Extraction for Classification of Proteomic Mass Spectra: A Comparative Study

To satisfy the ever growing need for effective screening and diagnostic tests, medical practitioners have turned their attention to high resolution, high throughput methods. One approach is to use mass spectrometry based methods for disease diagnosis. Effective diagnosis is achieved by classifying the mass spectra as belonging to healthy or diseased individuals. Unfortunately, the high resolution mass spectrometry data contains a large degree of noisy, redundant and irrelevant information, making accurate classification difficult. To overcome these obstacles, feature extraction methods are used to select or create small sets of relevant features. This paper compares existing feature selection methods to a novel wrapper-based feature selection and centroid-based classification method. A key contribution is the exposition of different feature extraction techniques, which encompass dimensionality reduction and feature selection methods. The experiments, on two cancer data sets, indicate that feature selection algorithms tend to both reduce data dimensionality and increase classification accuracy, while the dimensionality reduction techniques sacrifice performance as a result of lowering the number of features. In order to evaluate the dimensionality reduction and feature selection techniques, we use a simple classifier, thereby making the approach tractable. In relation to previous research, the proposed algorithm is very competitive in terms of (i) classification accuracy, (ii) size of feature sets, (iii) usage of computational resources during both training and classification phases.

Automated selection of vision operator libraries with evolutionary algorithms

Adaptive image interpretation systems can learn optimal image interpretation policies for a given domain without human intervention. The policies are learned over an extensive generic image processing operator library. One of the principal weaknesses of the method lies with the large size of such libraries which can make machine learning process intractable. In this paper we demonstrate how evolutionary algorithms can be used to reduce the size of operator library thereby speeding up learning of the policy while still keeping human experts out of the development loop. Experiments in a challenging domain of forestry image interpretation exhibited a 93.3% reduction in the execution time, while maintaining the image interpretation accuracy within 5.5% of optimal.

Performance of Lookahead Control Policies in the Face of Abstractions and Approximations

This paper explores the formulation of image interpretation as a Markov Decision Process (MDP) problem, highlighting the important assumptions in the MDP formulation. Furthermore state abstraction, value function and action approximations as well as lookahead search are presented as necessary solution methodologies. We view the task of image interpretation as a dynamic control problem where the optimal vision operator is selected responsively based on the problem solvingstate at hand. The control policy, therefore, maps problem-solving states to operators in an attempt to minimize the total problem-solving time while reliably interpretingthe image. Real world domains, like that of image interpretation, usually have incredibly large state spaces which require methods of abstraction in order to be manageable by todays information processingsystems. In addition an optimal value function (V*) used to evaluate state quality is also generally unavailable nrequiring appro ximations to be used in conjunction with state abstraction. Therefore, the performance of the system is directly related to the types of abstractions and approximations present.

Heterogeneous stacking for classification-driven watershed segmentation

Marker-driven watershed segmentation attempts to extract seeds that indicate the presence of objects within an image. These markers are subsequently used to enforce regional minima within a topological surface used by the watershed algorithm. The classification-driven watershed segmentation (CDWS) algorithm improved the production of markers and topological surface by employing two machine-learned pixel classifiers. The probability maps produced by the two classifiers were utilized for creating markers, object boundaries, and the topological surface. This paper extends the CDWS algorithm by (i) enabling automated feature extraction via independent components analysis and (ii) improving the segmentation accuracy by introducing heterogeneous stacking. Heterogeneous stacking, an extension of stacked generalization for object delineation, improves pixel labeling and segmentation by training base classifiers on multiple target concepts extracted from the original ground truth, which are subsequently fused by the second set of classifiers. Experimental results demonstrate the effectiveness of the proposed system on real world images, and indicate significant improvement in segmentation quality over the base system.

Comparison of Machine Learned Image Interpretation Systems in the Domain of Forestry

Automated image interpretation is an important task with numerous applications. Until recently, designing such systems required extensive subject matter and computer vision expertise resulting in poor cross-domain portability and expensive maintenance. Recently, a machine-learned system ADORE (adaptive object recognition) was successfully applied in an aerial image interpretation domain. In this paper we evaluate an extended version of this system, applied for the first time to a natural image interpretation domain. Performance of MR ADORE system is compared to the hierarchical hidden Markov random field (HHRMF) algorithm for supervised image annotation. We show that a hybrid system, easily constructed by utilizing the HHMRF models as operators within MR ADORE, performs significantly better than either of the systems on their own. To the best of our knowledge this is the first successful case of learning both vision operators and an adaptive control policy guiding their application in a single system

Supervised image segmentation via ground truth decomposition

This paper proposes a data driven image segmentation algorithm, based on decomposing the target output (ground truth). Classical pixel labeling methods utilize machine learning algorithms that induce a mapping from pixel features to individual pixel labels. In contrast we propose to first extract features from both images and labels. Subsequently we induce a mapping from pixel features to label features and synthesize the final output by combining the newly derived label components. We demonstrate the effectiveness of the proposed approach by applying log-Gabor filters to both input and ground truth images of mineral ore. Subsequently we train perceptrons and regression trees to produce individual output components that are combined in frequency space to create the final segmentation. Experimental results show significant improvements over contextual pixel labeling and over ensemble methods.