Chee-Hung Henry Chu

Multiresolution superpixels for visual saliency detection

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.

Volume compression of MRI data using zerotrees of wavelet coefficients

Volume data such as those acquired by magnetic resonance imaging techniques can be compressed efficiently using the wavelet transform. Wavelet compression methods need to retain both the value and the location of the significant coefficients. We present experimental results demonstrating the use of zerotree encoding methods in wavelet compression can enhance the ability to further compress volume data.

Detection of reflecting surfaces by a statistical model

Remote sensing is widely used assess the destruction from natural disasters and to plan relief and recovery operations. How to automatically extract useful features and segment interesting objects from digital images, including remote sensing imagery, becomes a critical task for image understanding. Unfortunately, current research on automated feature extraction is ignorant of contextual information. As a result, the fidelity of populating attributes corresponding to interesting features and objects cannot be satisfied. In this paper, we present an exploration on meaningful object extraction integrating reflecting surfaces. Detection of specular reflecting surfaces can be useful in target identification and then can be applied to environmental monitoring, disaster prediction and analysis, military, and counter-terrorism. Our method is based on a statistical model to capture the statistical properties of specular reflecting surfaces. And then the reflecting surfaces are detected through cluster analysis.

Shadow removal from textured images

Shadows and shadings are typical natural phenomena, which can often be found in images and videos acquired under strong directional lighting, such as those taken outdoors on a sunny day. Unfortunately, shadows can cause many difficulties in image processing and vision-related tasks, such like image segmentation and object recognition. Therefore, shadow removal is needed for improving the performance of these image understanding tasks. We present a new shadow removal algorithm for real textured color images. The algorithm is based on the statistical property of textures in images. The experimental results on real-world data are shown to demonstrate this algorithm.

Saliency Detection using Geometric Context Contrast Inferred from Natural Images

Image saliency detection using region contrast is often based on the premise that salient region has a contrast with the background which becomes a limiting factor if the color of the salient object background is similar. To overcome this problem associated with single image analysis, we propose to collect background regions from a collection of images where generative property of, say, natural images ensures that all the images are spun out of it hence negating any bias. Background regions are differentiated based on their geometric context where we use the ground and sky context as background. Finally, the aggregated map is generated using color contrast between the superpixels segments of the image and collection of background

Visual saliency approach to anomaly detection in an image ensemble

Visual saliency is a bottom-up process that identifies those regions in an image that stand out from their surroundings. We oversegment an image as a collection of “super pixels” (SPs). Each SP is salient if it is different in color from all other SPs and if its most similar SPs are nearby. We test our method on image sequences collected by a vehicle. We consider an SP in a frame as salient if it stands out from all frames in a collection that consists of an ensemble of images from different road segments and a sequence of immediate past frames.

Predicting Hospital Safety Measures using Patient Experience of Care Responses.

To make healthcare more cost effective, the current trend in the U.S. is towards a hospital value-based purchasing program. In this program, a hospital’s performance is measured in the safety, patient experience of care, clinical care, and efficiency and cost reduction domains. We investigate the efficacy of predicting the safety measures using the patient experience of care measures. We compare four classifiers in the prediction tasks and concluded that random forest and support vector machine provided the best performance.

Object detection and tracking under planar constraints

Automatic object detection and tracking has been widely applied in the video surveillance systems for homeland security and data fusion in the remote sensing and airborne imagery. The typical applications include human motion analysis, vehicle detection, and architectural building detection. Here we conduct object detection and tracking under planar constraints for interesting objects. Planar surface abounds in man-made environment. It provides much useful information for image understanding and then can be adopted to improve the performance of object detection and tracking. The experiments on real data show that object detection and tracking could be successfully implemented by incorporating planar information of interesting objects.

Fundamental matrix and planar homographies in stereo vision

We describe the epipolar constraint that specifies the geometry of stereo vision. We consider the 3D structure reconstruction from multiple views through the new perspective of basing the reconstruction from directly estimated planar homographies instead of using techniques that are based on matched point pairs. Planar homography parameters can more accurately extract scene planar surfaces and directly solve for the 3D structure and camera motion parameters. The new method has the advantage that it integrates larger amount of information because the homography parameters are estimated directly from the intensities and not from an abstracted descriptor of the neighborhood. Because it does not rely on a transformation of an entire image region, the method is efficient.

Compressive sampling approach to visual attention in image scene analysis

Many image scene analysis applications require a computational approach to visual attention. The foreground in these applications is typically sparse in spatial support. Compressive sampling enables an approach to reconstruct the sparse map of image regions that stand out from the background using fewer measurements. A convex optimization algorithm, for instance, can be used to recover the sparse map in the wavelet domain. Besides being sparse in the transform domain, the background of natural images has an interesting property that the amplitude of the averaged Fourier spectrum is approximately proportional to the inverse of the frequency. This further enables us to approximate an average background signal for extracting the out-of-ordinary foreground signal corresponding to objects of interest.