Suchendra M. Bhandarkar

Image segmentation using evolutionary computation

Image segmentation denotes a process by which a raw input image is partitioned into nonoverlapping regions such that each region is homogeneous and the union of any two adjacent regions is heterogeneous. A segmented image is considered to be the highest domain-independent abstraction of an input image. The image segmentation problem is treated as one of combinatorial optimization. A cost function which incorporates both edge information and region gray-scale uniformity is defined. The cost function is shown to be multivariate with several local minima. The genetic algorithm, a stochastic optimization technique based on evolutionary computation, is explored in the context of image segmentation. A class of hybrid evolutionary optimization algorithms based on a combination of the genetic algorithm and stochastic annealing algorithms such as simulated annealing, microcanonical annealing, and the random cost algorithm is shown to exhibit superior performance as compared with the canonical genetic algorithm. Experimental results on gray-scale images are presented.

An edge detection technique using genetic algorithm-based optimization

Abstract In this paper we present a genetic algorithm-based optimization technique for edge detection. The problem of edge detection is formulated as one of choosing a minimum cost edge configuration. The edge configurations are viewed as two-dimensional chromosomes with fitness values inversely proportional to their costs. The design of the crossover and the mutation operators in the context of the two-dimensional chromosomal representation is described. The knowledge-augmented mutation operator which exploits knowledge of the local edge structure is shown to result in rapid convergence. The incorporation of meta-level operators and strategies such as the elitism strategy, the engineered conditioning operator and adaptation of mutation and crossover rates in the context of edge detection are discussed and are shown to improve the convergence rate. The genetic algorithm with various combinations of meta-level operators is tested on synthetic and natural images. The performance of the genetic algorithm-based cost minimization technique is compared both qualitatively and quantitatively with local search-based and simulated annealing-based cost minimization approaches. The genetic algorithm-based technique is shown to perform very well in terms of robustness to noise, rate of convergence and quality of the final edge image.

The REFINE multiprocessor—theoretical properties and algorithms

Abstract A reconfigurable interconnection network based on a multi-ring architecture called REFINE is described. REFINE embeds a single 1-factor of the Boolean hypercube in any given configuration. The mathematical properties of the REFINE topology and the hardware for the reconfiguration switch are described. The REFINE topology is scalable in the sense that the number of interprocessor communication links scales linearly with network size whereas the network diameter scales logarithmically with network size. Primitive parallel operations on the REFINE topology are described and analyzed. These primitive operations could be used as building blocks for more complex parallel algorithms. A large class of algorithms for the Boolean n-cube which includes the FFT and the Batchers bitonic sort is shown to map efficiently on the REFINE topology. The REFINE multiprocessor is shown to offer a cost-effective alternative to the Boolean n-cube multiprocessor architecture without substantial loss in performance.

Parallel stereocorrelation on a reconfigurable multi-ring network

A reconfigurable network termed as the reconfigurable multi-ring network (RMRN) is described. The RMRN is shown to be a truly scalable network in that each node in the network has a fixed degree of connectivity and the reconfiguration mechanism ensures a network diameter of O(log2N) for anN-processor network. Algorithms for the two-dimensional mesh and the SIMD or SPMD n-cube are shown to map very elegantly onto the RMRN. Basic message passing and reconfiguration primitives for the SIMD/SPMD RMRN are designed for use as building blocks for more complex parallel algorithms. The RMRN is shown to be a viable architecture for image processing and computer vision problems using the parallel computation of the stereocorrelation imaging operation as an example. Stereocorrelation is one of the most computationally intensive imaging tasks. It is used as a visualization tool in many applications, including remote sensing, geographic information systems and robot vision.

Eye Tracking for Everyone

From scientific research to commercial applications, eye tracking is an important tool across many domains. Despite its range of applications, eye tracking has yet to become a pervasive technology. We believe that we can put the power of eye tracking in everyones palm by building eye tracking software that works on commodity hardware such as mobile phones and tablets, without the need for additional sensors or devices. We tackle this problem by introducing GazeCapture, the first large-scale dataset for eye tracking, containing data from over 1450 people consisting of almost 2:5M frames. Using GazeCapture, we train iTracker, a convolutional neural network for eye tracking, which achieves a significant reduction in error over previous approaches while running in real time (10-15fps) on a modern mobile device. Our model achieves a prediction error of 1.71cm and 2.53cm without calibration on mobile phones and tablets respectively. With calibration, this is reduced to 1.34cm and 2.12cm. Further, we demonstrate that the features learned by iTracker generalize well to other datasets, achieving state-of-the-art results. The code, data, and models are available at http://gazecapture.csail.mit.edu.

The Hough transform on a reconfigurable multi-ring network

Abstract A novel reconfigurable network referred to as the Reconfigurable Multi-Ring Network (RMRN) is described. The RMRN is shown to be a truly scalable network, in that each node in the network has a fixed degree of connectivity and the reconfiguration mechanism ensures a network diameter of O (log 2 N ) for an N -processor network. Algorithms for the 2-D mesh and the SIMD n -cube are shown to map very elegantly onto the RMRN. Basic message passing and reconfiguration primitives for the SIMD RMRN are designed which could be used as building blocks for more complex parallel algorithms. The RMRN is shown to be a viable architecture for image processing and computer vision problems via the parallel computation of the Hough transform. The parallel implementation of the Y -angle Hough transform of an N × N image is showed to have a asymptotic complexity of O ( Y log 2 Y + log 2 N ) on the SIMD RMRN with O ( N 2 ) processors. This compares favorably with the O ( Y + log 2 N ) optimal algorithm for the same Hough transform on the MIMD n -cube with O ( N 2 ) processors.

CATALOG: a system for detection and rendering of internal log defects using computer tomography

Abstract. This paper describes the design and implementation of a machine vision system CATALOG for detection and classification of some important internal defects in hardwood logs via analysis of computer axial tomography (CT or CAT) images. The defect identification and classification in CATALOG consists of two phases. The first phase comprises of the segmentation of a single CT image slice, which results in the extraction of 2D defect-like regions from the CT image slice. The second phase comprises of the correlation of the 2D defect-like regions across CT image slices in order to establish 3D support. The segmentation algorithm for a single CT image is a complex form of multiple-value thresholding that exploits both, the prior knowledge of the wood structure within the log and the gray-level characteristics of the image. The algorithm for extraction of 2D defect-like regions in a single CT image first locates the pith of the log cross section, groups the pixels in the segmented image on the basis of their connectivity and classifies each 2D region as either a defect-like region or a defect-free region using shape, orientation and morphological features. Each 2D defect-like region is classified as a defect or non-defect via correlation across corresponding 2D defect-like regions in neighboring CT image slices. The 2D defect-like regions with adequate 3D support are labeled as true defects. The current version of CATALOG is capable of 3D reconstruction and rendering of the log and its internal defects from the individual CT image slices. CATALOG is also capable of simulation and rendering of key machining operations such as sawing and veneering on the 3D reconstructions of the logs. The current version of CATALOG is intended as a decision aid for sawyers and machinists in lumber mills and also as an interactive training tool for novice sawyers and machinists.

A multilayer self-organizing feature map for range image segmentation

Abstract This paper proposes and describes a hierarchical self-organizing neural network for range image segmentation. The multilayer self-organizing feature map (MLSOFM), which is an extension of the traditional (single-layer) self-organizing feature map (SOFM) is seen to alleviate the shortcomings of the latter in the context of range image segmentation. The problem of range image segmentation is formulated as one of vector quantization and is mapped onto the MLSOFM. The MLSOFM combines the ideas of self-organization and topographic mapping with those of multiscale image segmentation. Experimental results using real range images are presented.

Multiscale image segmentation using a hierarchical self-organizing map

Abstract Multiscale structures and algorithms that unify the treatment of local and global scene information are of particular importance in image segmentation. Vector quantization, owing to its versatility, has proved to be an effective means of image segmentation. Although vector quantization can be achieved using self-organizing maps with competitive learning, self-organizing maps in their original single-layer structure, are inadequate for image segmentation. A hierarchical self-organizing neural network for image segmentation is presented. The Hierarchical Self-Organizing Map (HSOM) is an extension of the conventional (single-layer) Self-Organizing Map (SOM). The problem of image segmentation is formulated as one of vector quantization and mapped onto the HSOM. By combining the concepts of self-organization and topographic mapping with those of multiscale image segmentation the HSOM alleviates the shortcomings of the conventional SOM in the context of image segmentation.

A reconfigurable architecture for image processing and computer vision

In this paper we describe a reconfigurable architecture for image processing and computer vision based on a multi-ring network which we call a Reconfigurable Multi-Ring System (RMRS). We describe the reconfiguration switch for the RMRS and also describe its VLSI implementation. The RMRS topology is shown to be regular and scalable and hence well-suited for VLSI implementation. We prove some important properties of the RMRS topology and show that a broad class of algorithms for the n-cube can be mapped to the RMRS in a simple and elegant manner. We design and analyze a class of procedural primitives for the SIMD RMRS and show how these primitives can be used as building blocks for more complex parallel operations. We demonstrate the usefulness of the RMRS for problems in image processing and computer vision by considering two important operations—the Fast Fourier Transform (FFT) and the Hough transform for detection of linear features in an image. Parallel algorithms for the FFT and the Hough transform on the SIMD RMRS are designed using the aforementioned procedural primitives. The analysis of the complexity of these algorithms shows that the SIMD RMRS is a viable architecture for problems in computer vision and image processing.