Maricor Soriano

Adaptive skin color modeling using the skin locus for selecting training pixels

Techniques for color-based tracking of faces or hands often assume a static skin model yet skin color, as measured by a camera, can change when lighting changes. Therefore, for robust skin pixel detection, an adaptive skin color model must be employed. We demonstrate a chromaticity-based constraint to select training pixels in a scene for updating a dynamic skin color model under changing illumination conditions. The method makes use of the ‘skin locus’ of a camera, that is, the area in chromaticity space where skin chromaticity under various lighting and camera calibration conditions is observed. Skin color models derived from the technique are compared with that derived by a common spatial constraint and is shown to be more consistent with manually extracted ground truth skin model per frame even as localization errors increase. The technique is applied to color-based face tracking in indoor and outdoor videos and is shown to succeed more often than other color model adaptation techniques. ? 2002 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.

Behavior of skin color under varying illumination seen by different cameras at different color spaces

The appearance of skin colors in the images depends among other things, on the camera, the calibration of the camera, and the illumination under which the image was taken. In this study, we investigate how the skin colors appear in the chromaticity coordinates of different color spaces like HSV/HSL, normalized rgb, YES and TSL. For this purpose, we have taken images of faces under 16 different illumination/camera calibration conditions using simulated illuminants (Horizon, A, fluorescent TL84 and daylight) with different RGB cameras (1CCD web cameras and a 3CCD camera). In the making of this series of 16 images, first the selected camera was calibrated to one of the four light sources and an image was taken. After that the light source was changed to the other light sources and at each time the person was imaged. The process was repeated to the other two light sources. The same procedure was done for all four light sources and for each camera. The skin regions were extracted from these images and this skin data was then converted to different color spaces. We inspected how the chromaticities of different skin color groups in these color spaces overlap in images taken in all 16 different cases and only in those cases in which the selected camera was calibrated to the current illuminant. These investigations were also made between different cameras. In addition to this, we examined the overlapping of all skin chromaticities from the different skin color groups between cameras.

Classification of coral reef images from underwater video using neural networks

We use a feedforward backpropagation neural network to classify close-up images of coral reef components into three benthic categories: living coral, dead coral and sand. We have achieved a success rate of 86.5% (false positive = 6.7%) for test images that were not in the training set which is high considering that corals occur in an immense variety of appearance. Color and texture features derived from video stills of coral reef transects from the Great Barrier Reef were used as inputs to the network. We also developed a rule-based decision tree classifier according to how marine scientists classify corals from texture and color, and obtained a lower recognition rate of 79.7% for the same set of images.

Detection of skin color under changing illumination: a comparative study

Faces and hands recorded under natural environments are frequently subject to illumination variations which affect their color appearance. This is a problem when the color cue is used to detect skin candidates at pixel level. Traditionally, color constancy has been suggested for correction, but after a lot of effort no good solution suitable for machine vision has emerged. However, many approaches have been proposed for general skin detection, but they are typically tested under mild changes in illumination chromaticity or do not define the variation range. This makes it difficult to evaluate their applicability for objects under varying illumination. The paper compares four state-of-the-art skin detection schemes under realistic conditions with drastic chromaticity change.

Curve spreads: a biometric from front-view gait video

We introduce the curve spread as an efficient descriptor of front-view gait of humans walking towards a camera. The curve spread is a compact two-dimensional representation of the time-variations of a moving body outline. Most gait biometrics employ features derived from side-view videos because limb swings are more pronounced from the side than from the front. However, side-view observations are often impractical and incompatible with the ability of humans to recognize others from front-view gait. Identification tests using cross correlation of curve spreads of 12 videos from 4 walking subjects, yielded 100% recognition rate.

Image classification of coral reef components from underwater color video

The purpose of this study is to automate coral reef assessment, that is, to classify coral images into benthic categories from digitized underwater video using a computer-based classifier such that coral reef analysis becomes less subjective, less tedious and more precise. Corals exhibit a variety of color, texture and structure which are the visual cues used by marine scientists for their classification. In computer vision, color is a point property of a picture element while texture is a property of an area. Color and texture have been combined as color-texture which is a feature that describes the spatial organization of colors in an area. As inputs to a classifier, the authors extract color, texture and color-texture descriptors from coral images and measure recognition rates using each feature. Corals are 3D structures and, when imaged, are prone to varying resolutions, perspective projection and lighting conditions. Therefore, an additional objective of this study is to address the problem of illumination, rotation and scale invariance in pattern recognition of underwater images. Images were classified into one of five benthic categories: alive coral, dead coral, dead coral with algae, algae and abiotics. Overall, texture was found to be more discriminating than using color alone or color and texture combined. Dead coral was the most successfully recognized class using color features.

Automated benthic counting of living and non-living components in Ngedarrak Reef, Palau via subsurface underwater video

We introduce an automated benthic counting system in application for rapid reef assessment that utilizes computer vision on subsurface underwater reef video. Video acquisition was executed by lowering a submersible bullet-type camera from a motor boat while moving across the reef area. A GPS and echo sounder were linked to the video recorder to record bathymetry and location points. Analysis of living and non-living components was implemented through image color and texture feature extraction from the reef video frames and classification via Linear Discriminant Analysis. Compared to common rapid reef assessment protocols, our system can perform fine scale data acquisition and processing in one day. Reef video was acquired in Ngedarrak Reef, Koror, Republic of Palau. Overall success performance ranges from 60% to 77% for depths of 1 to 3 m. The development of an automated rapid reef classification system is most promising for reef studies that need fast and frequent data acquisition of percent cover of living and nonliving components.

Improved Classification Robustness for Noisy Cell Images Represented as Principal-Component Projections in a Hybrid Recognition System

Different types of cells are recognized from their noisy images by use of a hybrid recognition system that consists of a learning principal-component analyzer and an image-classifier network. The inputs to the feed-forward backpropagation classifier are the first 15 principal components of the 10 x 10 pixel image to be classified. The classifier was trained with clear images of cells in metaphase, unburst cells, and other erroneous patterns. Experimental results show that the recognition system is robust to image scaling and rotation, as well as to image noise. Cell recognition is demonstrated for images that are corrupted with additive Gaussian noise, impulse noise, and quantization errors. We compare the performance of the hybrid recognition system with that of a conventional three-layer feed-forward backpropagation network that uses the raw image directly as input.

Fluorescence spectrum estimation using multiple color images and minimum negativity constraint.

An inexpensive method to convert a microscope into an imaging spectrometer is presented. Unlike current microscope-based spectrometers which use specialized optics or scanning mechanisms, our system only requires at most two image captures with a 3-chip CCD camera and a lightly-tinted color filter to output the color signal of a sample at each pixel. Basis spectra are obtained by principal components analysis applied to an ensemble of color signals of commercially-available dyes observed with different dichroic mirrors. A transformation matrix from channel values to spectral coefficients is derived. Minimum negativity constraint is applied to eliminate negative parts of the reconstructed fluorescence spectrum. The technique is demonstrated on fluorescence microspheres (fluorospheres) and chlorophyll from plant leaf.

Compact time-independent pattern representation of entire human gait cycle for tracking of gait irregularities

We demonstrate a three-dimensional (for location, time, and magnitude of body part movement) pattern representation of entire time-dependent front-view gait cycle that simultaneously displays the coupled kinetics of different body parts thereby revealing possible irregularities in the gait characteristics of a moving human subject. The time-independent pattern is able to track attendant displacements of other parts (e.g., in the lower body) that result from a movement of a lead part (e.g., in the upper body). It is derived by applying a computationally simple silhouette feature extractor algorithm unto the video footage of the subject that is taken using one stationary CCD camera. The pattern illustrates in a single field-of-view, possible mutual interactions of all four limbs allowing us to identify the types and phases of the gait cycle, observe possible shifting of body weight and other nonlocal effects of the gait pathology. As a data representation, the pattern representation is more compact and easier to store, retrieve, transport and organize. The patterns are easily compared with each other via straightforward image cross-correlation technique. Front-view gait analysis permits an unambiguous and accurate description of the gait dynamics that is not possible with side- or top-view observation. Among the potential applications of our technique are improved diagnosis and treatment of gait pathologies in rehabilitation clinics and modelling schools as well as development of more robust surveillance systems.