Detecting salient regions in an image or video frame is an important step in early vision and image understanding. We present a visual saliency detection method by measuring the difference in color content of an image segment with that of its neighbors. We represent each segment with richer color descriptors in the form of a regional dominant color descriptor. The color difference between a pair of neighbors is found using the Earth Mover’s Distance. The cost of moving color descriptors between neighboring segments robustly captures the difference between neighboring segments. We evaluate our method on standard datasets and compare it with other state-of-the-art methods to demonstrate that it has better true positive rate at a fixed false positive rate in detecting salient pixels relative to the ground truth. The proposed method uses local cues without being an edge highlighter, a common problem of local contrast-based methods.

Salient regions are those that stand out from others in an image. We present an algorithm to detect salient regions in an image that is represented as superpixels at a number of resolutions. Superpixels are segments generated by oversegmenting an image and they form a perceptually meaningful representation that preserves the underlying image structure. The novelty of our method is the ranking of a superpixel by its dissimilarities with respect to other superpixels and highlighting the statistically salient region proportional to their rank. This is based on the premise that salient region group together and they stand out. We tested our method using standard data sets containing images of varied complexity and compared the results to ground truth data. Our results show that our saliency detection algorithm is robust to changes in color, object size, object location in image and background type.

In this paper, the 1-D mapping is studied for estimating a general projective transformation between images. First, a ray model of image patches is proposed to convert a complex projective deformation of a patch in the Cartesian space into two simple rigid shifts of radial lines in the log-polar/inverse-polar space. To capture these shifts via direct search, two types of radial image line matching, viz. 1-D log-polar mapping and 1-D inverse-polar mapping, are formally defined. Also, the unique features of these shifts are analyzed. Then, a novel two-step mathematical framework is devised for transformation estimation. In the first step, the 1-D log-polar mapping is applied to estimating four affine parameters via two alternative methods for simulated/real images. In the second step, the 1-D inverse-polar mapping is applied to estimating two projective parameters. By a direct search combined with the two-step line matching, two translational parameters are also determined. Finally, the performance of the framework is evaluated first in estimating a variety of projective transformations by comparing the estimates with their ground truths and between the two proposed methods, and second in registering a wide range of challenging image pairs by comparing with a state-of-the-art feature-based method.

Visualization can facilitate the understanding of the structures of a collection of documents that are related to each other by links, such as citations in formal publications. We present results of visualizing minimum spanning trees based on document co-citation counts and on document citation correlations.

We propose a new methodology for visualizing association mining results. Inter-item distances are computed from combinations of itemset supports. The new distances retain a simple pairwise structure, and are consistent with important frequently occurring itemsets. Thus standard tools of visualization, e.g. hierarchical clustering dendrograms can still be applied, while the distance information upon which they are based is richer. Our approach is applicable to general association mining applications, as well as applications involving information spaces modeled by directed graphs, e.g. the Web. In the context of collections of hypertext documents, the inter-document distances capture the information inherent in a collection's link structure, a form of link mining. We demonstrate our methodology with document sets extracted from the Science Citation Index, applying a metric that measures consistency between clusters and frequent itemsets.

In many applications, the color bit-depth of an image has to be modified, typically lowered, to match the rendering device. Halftoning is typically used to reduce the effects of quantization, which can produce objectionable contouring effects. Until now, color-indexed images whose palette was a subset of the colors of the image could only be dithered using error diffusion. We present an algorithm for color halftoning when the color palette is image-dependent. Experimental results demonstrate that our new algorithm does not introduce color shifts, color spikes, or wormy textures, and has greater preservation of detail.

The standard wavelet transform can be considered to be a subtree of a complete multiresolution binary tree, in which each branch corresponds to a decomposition step. Any subtree of this multiresolution tree which covers all leaves is a valid wavelet packet processing tree, which specifies the processing steps for a wavelet decomposition of the original data. The present paper describes an adaptation scheme for configuring a wavelet packet processing tree using a genetic algorithm search of the complete multiresolution tree. Data compression is used as the target signal processing task. Trees are configured for different compression rates, and performance is measured based on the reconstruction distortion. Performances are validated using testing as well as training data; the performance of the configured trees are compared to that of the standard wavelet transform.

The representation of the data used in a computational algorithm should be coupled to the computing machine on which the algorithm is implemented. A model for representing image data for gray-level image processing by artificial neural networks is presented. An input layer converts an image with real-valued or quantized intensities to a binary representation. The input layer uses threshold logic as the activation rule. An image processing task is decomposed into a collection of massively parallel decision tasks, each of which is performed locally by a hidden processing unit. The output of the hidden layer, in binary form, can be converted to a resultant image with real-valued intensities by the output layer. An advantage of this approach is that it is well-matched to the decision and classification capabilities of neural networks. A noise-suppressing algorithm with robust performance is implemented to illustrate the efficacy of this model

The performances of three classes of edge operators are evaluated. The three are conventional 3×3 operators with a connectivity test, sequential edge detectors using both edge strength and direction information, and edge detectors using large masks. Implementations of the sequential edge operators using both the edge strength and direction information to detect edge segments with smoothly varying edge direction and physical length are discussed. Results of applying these edge detectors to real-world images are also presented.