Learning Nonlinear Visual Processing from Natural Images

The paradigm of computational vision hypothesizes that any visual function -- such as the recognition of your grandparent -- can be replicated by computational processing of the visual input. What are these computations that the brain performs? What should or could they be? Working on the latter question, this dissertation takes the statistical approach, where the suitable computations are attempted to be learned from the natural visual data itself. In particular, we empirically study the computational processing that emerges from the statistical properties of the visual world and the constraints and objectives specified for the learning process. This thesis consists of an introduction and 7 peer-reviewed publications, where the purpose of the introduction is to illustrate the area of study to a reader who is not familiar with computational vision research. In the scope of the introduction, we will briefly overview the primary challenges to visual processing, as well as recall some of the current opinions on visual processing in the early visual systems of animals. Next, we describe the methodology we have used in our research, and discuss the presented results. We have included some additional remarks, speculations and conclusions to this discussion that were not featured in the original publications. We present the following results in the publications of this thesis. First, we empirically demonstrate that luminance and contrast are strongly dependent in natural images, contradicting previous theories suggesting that luminance and contrast were processed separately in natural systems due to their independence in the visual data. Second, we show that simple cell -like receptive fields of the primary visual cortex can be learned in the nonlinear contrast domain by maximization of independence. Further, we provide first-time reports of the emergence of conjunctive (corner-detecting) and subtractive (opponent orientation) processing due to nonlinear projection pursuit with simple objective functions related to sparseness and response energy optimization. Then, we show that attempting to extract independent components of nonlinear histogram statistics of a biologically plausible representation leads to projection directions that appear to differentiate between visual contexts. Such processing might be applicable for priming, \ie the selection and tuning of later visual processing. We continue by showing that a different kind of thresholded low-frequency priming can be learned and used to make object detection faster with little loss in accuracy. Finally, we show that in a computational object detection setting, nonlinearly gain-controlled visual features of medium complexity can be acquired sequentially as images are encountered and discarded. We present two online algorithms to perform this feature selection, and propose the idea that for artificial systems, some processing mechanisms could be selectable from the environment without optimizing the mechanisms themselves. In summary, this thesis explores learning visual processing on several levels. The learning can be understood as interplay of input data, model structures, learning objectives, and estimation algorithms. The presented work adds to the growing body of evidence showing that statistical methods can be used to acquire intuitively meaningful visual processing mechanisms. The work also presents some predictions and ideas regarding biological visual processing.

An algebraic approach to the restoration of images

Two different matrix algorithms are described for the restoration of blurred pictures. These are illustrated by numerical examples.

Decision fusion for reliable flood mapping using remote sensing images

Flood extent mapping is a basic tool for flood damage assessment, which can be done by digital classification techniques using satellite imageries, including the data recorded by radar and optical sensors. However, converting the data into the information we need is not a straightforward task. One of the great challenges involved in the data interpretation is to separate the permanent water bodies and flooding regions, including both the fully inundated areas and the wet areas where trees and houses are partly covered with water. This paper adopts the decision fusion technique to combine the mapping results from radar data and the NDVI data derived from optical data. An improved capacity in terms of identifying the permanent or semi-permanent water bodies from flood inundated areas has been achieved. Computer software tools Multispec and Matlab were used.

Block-based feature detection and matching for mosaicing of camera-captured document images

In this paper, we present a new feature-based approach for mosaicing of camera-captured document images. A novel block-based scheme is employed to ensure that corners can be reliably detected over a wide range of images. 2-D discrete cosine transform is computed for image blocks defined around each of the detected corners and a small subset of the coefficients is used as a feature vector A 2-pass feature matching is performed to establish point correspondences from which the homography relating the input images could be computed. The algorithm is tested on a number of complex document images casually taken from a hand-held camera yielding convincing results.

Edge-based connected component approach for skew correction of complex document images

Skew correction of complex document images is a difficult task. We propose an edge-based connected component approach for robust skew correction of documents with complex layout and content. The algorithm essentially consists of two steps - an 'initialization' step to determine the image orientation from the centroids of the connected components and a 'search' step to find the actual skew of the image. During initialization, we choose two different sets of points regularly spaced across the the image, one from the left to right and the other from top to bottom. The image orientation is determined from the slope between the two succesive nearest neighbors of each of the points in the chosen set. The search step finds succesive nearest neighbors that satisfy the parameters obtained in the initialization step. The final skew is determined from the slopes obtained in the 'search' step. Unlike other connected component based methods, the proposed method does not require any binarization step that generally precedes connected component analysis. The method works well for scanned documents with complex layout of any skew with a precision of 0.5 degrees.

Entropy based Skew correction of document images

The document images that are fed into an Optical Character Recognition system, might be skewed. This could be due to improper feeding of the document into the scanner or may be due to a faulty scanner. In this paper, we propose a skew detection and correction method for document images. We make use of the inherent randomness in the Horizontal Projection profiles of a text block image, as the skew of the image varies. The proposed algorithm has proved to be very robust and time efficient. The entire process takes less than a second on a 2.4 GHz Pentium IV PC.

Stability of scheduled message communication over degraded broadcast channels

We develop a multi-class discrete-time processor-sharing queueing model for scheduled message communication over a discrete memoryless degraded broadcast channel. The framework we consider here models both the random message arrivals and the subsequent reliable communication by suitably combining techniques from queueing theory and information theory. Requests for message transmissions are assumed to arrive according to i.i.d. arrival processes. Then, (i) we derive an outer bound to the stability region of message arrival rate vectors achievable by the class of stationary scheduling policies, (ii) we show for any message arrival rate vector that satisfies the outer bound, that there exists a stationary "state-independent" policy that results in a stable system for the corresponding message arrival processes, and (iii) under an asymptotic regime, we show that the stability region of information arrival rate vectors is the information-theoretic capacity region of a degraded broadcast channel.

GAP-RBF based NR image quality measurement for JPEG coded images

In this paper, we present a growing and pruning radial basis function based no-reference (NR) image quality model for JPEG-coded images. The quality of the images are estimated without referring to their original images. The features for predicting the perceived image quality are extracted by considering key human visual sensitivity factors such as edge amplitude, edge length, background activity and background luminance. Image quality estimation involves computation of functional relationship between HVS features and subjective test scores. Here, the problem of quality estimation is transformed to a function approximation problem and solved using GAP-RBF network. GAP-RBF network uses sequential learning algorithm to approximate the functional relationship. The computational complexity and memory requirement are less in GAP-RBF algorithm compared to other batch learning algorithms. Also, the GAP-RBF algorithm finds a compact image quality model and does not require retraining when the new image samples are presented. Experimental results prove that the GAP-RBF image quality model does emulate the mean opinion score (MOS). The subjective test results of the proposed metric are compared with JPEG no-reference image quality index as well as full-reference structural similarity image quality index and it is observed to outperform both.

Text localization and extraction from complex gray images

We propose two texture-based approaches, one involving Gabor filters and the other employing log-polar wavelets, for separating text from non-text elements in a document image. Both the proposed algorithms compute local energy at some information-rich points, which are marked by Harris' corner detector. The advantage of this approach is that the algorithm calculates the local energy at selected points and not throughout the image, thus saving a lot of computational time. The algorithm has been tested on a large set of scanned text pages and the results have been seen to be better than the results from the existing algorithms. Among the proposed schemes, the Gabor filter based scheme marginally outperforms the wavelet based scheme.

Automatic text block separation in document images

Separation of printed text blocks from the non-text areas, containing signatures, handwritten text, logos and other such symbols, is a necessary first step for an OCR involving printed text recognition. In the present work, we compare the efficacy of some feature-classifier combinations to carry out this separation task. We have selected length-nomalized horizontal projection profile (HPP) as the starting point of such a separation task. This is with the assumption that the printed text blocks contain lines of text which generate HPP's with some regularity. Such an assumption is demonstrated to be valid. Our features are the HPP and its two transformed versions, namely, eigen and Fisher profiles. Four well known classifiers, namely, Nearest neighbor, Linear discriminant function, SVM's and artificial neural networks have been considered and efficiency of the combination of these classifiers with the above features is compared. A sequential floating feature selection technique has been adopted to enhance the efficiency of this separation task. The results give an average accuracy of about 96.

Binarization and localization of text images captured on a mobile phone camera

This paper proposes and compares four methods of binarzing text images captured using a camera mounted on a cell phone. The advantages and disadvantages(image clarity and computational complexity) of each method over the others are demonstrated through binarized results. The images are of VGA or lower resolution.

Classification of magnetic resonance brain images using wavelets as input to support vector machine and neural network

In this paper. we propose a novel method using wavelets as input to neural network self-organizing maps and support vector machine for classification of magnetic resonance (MR) images of the human brain. The proposed method classifies MR brain images as either normal or abnormal. We have tested the proposed approach using a dataset of 52 MR brain images. Good classification percentage of more than 94% was achieved using the neural network self-organizing maps (SOM) and 98% front support vector machine. We observed that the classification rate is high for a Support vector machine classifier compared to self-organizing map-based approach.

Memory-based reasoning approach for pattern recognition of binary images

Template matching is concerned with measuring the similarity between patterns of two objects. This paper proposes a memory-based reasoning approach for pattern recognition of binary images with a large template set. It seems that memory-based reasoning intrinsically requires a large database. Moreover, some binary image recognition problems inherently need large template sets, such as the recognition of Chinese characters which needs thousands of templates. The proposed algorithm is based on the Connection Machine, which is the most massively parallel machine to date, using a multiresolution method to search for the matching template. The approach uses the pyramid data structure for the multiresolution representation of templates and the input image pattern. For a given binary image it scans the template pyramid searching the match. A binary image of N × N pixels can be matched in O(log N) time complexity by our algorithm and is independent of the number of templates. Implementation of the proposed scheme is described in detail.

Multiple contour extraction from graylevel images using an artificial neural network

For active contour modeling (ACM), we propose a novel self-organizing map (SOM)-based approach, called the batch-SOM (BSOM), that attempts to integrate the advantages of SOM- and snake-based ACMs in order to extract the desired contours from images. We employ feature points, in the form of ail edge-map (as obtained from a standard edge-detection operation), to guide the contour (as in the case of SOM-based ACMs) along with the gradient and intensity variations in a local region to ensure that the contour does not "leak" into the object boundary in case of faulty feature points (weak or broken edges). In contrast with the snake-based ACMs, however, we do not use an explicit energy functional (based on gradient or intensity) for controlling the contour movement. We extend the BSOM to handle extraction of contours of multiple objects, by splitting a single contour into as many subcontours as the objects in the image. The BSOM and its extended version are tested on synthetic binary and gray-level images with both single and multiple objects. We also demonstrate the efficacy of the BSOM on images of objects having both convex and nonconvex boundaries. The results demonstrate the superiority of the BSOM over others. Finally, we analyze the limitations of the BSOM.