| CS 766 | Fall 2002 |
CS 766 Student Projects -- Fall 2002Traffic volume data is basic and crucial to the study of transportation engineering. Traditionally, field review must be done to count the volume data manually which is tedious and labor intensive. Recently, video cameras and camcorders have been used in the field to capture videos of traffic volumes. Thus, by analyzing the sequence of images (or video) from these devices, traffic volume data can be detected automatically. For this purpose, a detection window will be firstly defined and the average squared differences of current frame and base frame will be calculated based on this selected window. There are two options for the base frame: background image or previous image. Both of them are to be tested and advantages and disadvantages of them will be given together with some testing results. More information is available on my project web site.
Teleconferencing has become important nowadays. One of the techniques that can be applied to enhance the teleconferencing result is to use view morphing to generate realistic views from various camera angles. In this assignment, we are going to apply view morphing technique and develop a user interface to synthesize new views based two input images of person heads taken from different cameras around the person. Basically the program will track feature lines from two views of a face and transform the 2 images in different angles. This technique can be applied to generate 3D face model and create a head turning effect. More information is available on our project web site.
In this paper, I investigate the use of pictorial structures in recognizing patterns in electron density maps. Such maps are an intermediate step in protein structure determination via x-ray crystallography. By building pictorial structures corresponding to individual amino acids, or short peptide sequences, patterns in the density map can be located. This provides a good estimate of the atomic coordinates of atoms in the located pattern. This technique could prove an important step in the development of a fully automated routine for interpreting electron density maps. In implementing pictorial structure recognition for this problem, several non-trivial additions need to be made to the linear-time algorithm developed by Felzenszwalb and Huttenlocher. The most significant of these involves building a 3D rotational model in which the structure is free to rotate around a connection (bond), but cannot rotate in any other direction. The linear-time algorithm was implemented producing a program which runs quickly provided too many discretizations (buckets) are not used for variables in the six-dimensional conformation space. The memory cost is quite high, however. Results of the program are disappointing. While a low-cost solution is found to the specifications provided, the solution is often far from optimal. More information is available on my project web site.
I will experiment with making Single Image Random Dot Stereograms. I will use Thimbleby, Inglis, and Witten's symmetric algorithm to correct for depth distortion, hidden surfaces, and echo artifacts. Some experimental improvements may be possible, time permitting. More information is available on my project web site.
Texture synthesis is a method that takes in a small patch of texture and produces a new texture patch or block that looks as if it were created using the same methods as the example texture. In this paper, I explore two modern texture synthesis techniques presented in two recent Siggraph papers: "Fast Texture Synthesis Using Tree-Structured Vector Quantization" and "Image Quilting for Texture Synthesis and Transfer." The advantages and disadvantages of the two techniques are shown through clear examples generated with implementations of the two techniques. More information is available on my project web site.
This paper will describe several techniques for determining depth from defocus (DFD), particularly as it relates to histological data. DFD is a family of methods to determine the depth of scene points based on the relative blurring between a small number of images due to defocus. It will also describe an implementation that models the bluring of images due to defocus. Histological research can use DFD techniques to obtain volumes with high depth resolution without the artifacts of other methods.
A technique for automatic MRI image registration is based on the Fast Fourier Transform. This technique uses images in the spectral domain to compute the phase correlation. The phase correlation relates two images by the Fourier shift property, and can be used to register MRI images. More information is available on my project web site.
This paper describes the various issues to be considered while maintaining and querying a multimedia database consisting of images and videos. Different querying strategies and techniques that can return useful information in a short span of time are discussed. The paper also describes an image indexing and retrieval algorithm for a multimedia database. The algorithm characterizes the color variations over the spatial extent of the image in a manner that provides semantically meaningful image comparisons. The indexing algorithm applies a Daubechies wavelet transform for each of the three color components. The wavelet coefficients and variances in the lowest frequency band are stored as feature vectors. A two step method which involves a crude selection based on variance and then a refined search involving a feature vector match is performed between the selected image and query. For better results multi resolution matching may also be used. This algorithm is extended to videos by capturing the key frames which capture the essential information in a video and using an image search on the key frames. The key frames are gleaned from the video abstraction of the entire video, akin to trailers in movies. A querying scheme involving text and image querying (as discussed above) is used to extract better results. More information is available on my project web site.
Volumetric scene modeling constructs the volumes or surfaces in the world which are consistent with the multiple input images. These scene space methods allow widely-separated views but it generally depends on calibrated cameras to determine the absolute relationship between points in space and visual ray. The increase in the number of voxels in scene space could reduce the estimation gap and increase the resolution in the generation of the scene. As a result, we can enhance the image quality from other views. However the tradeoff of this improvement will be more awful calculations which may make this technique impractical. In this project, we hope that we can to develop a method to first construct voxels in coarse resolution and then explore the more voxels in fine resolution as needed. More information is available on my project web site.
A growing demand for traffic flow data and automatic vehicle identification leads researchers around the world to adopt advanced electronic and computer vision technologies to monitor and control traffic. These technologies were used for traffic surveying and monitoring, e.g. finding stolen cars, controlling access to car parks and gathering traffic flow statistics. Many commercial applications were already developed and applied to the real time based situations. In this project, the most effective algorithm was applied to recognize the vehicle license plate from video recorded images. More information is available on my project web site.
A 3D cortical surface mesh exists simply as a collection of connected polygons. The purpose of segmenting such a mesh is to impose a higher level structure which represents something about the underlying structure of the mesh itself. This segmentation should reduce the mesh into "meaningful," connected pieces. In this paper, segmentation using the watershed algorithm is implemented on brain cortical surface meshes. The height function used is a curvature measure inherent in the geometry of the mesh. Four different curvature measures are compared: mean, Gaussian, absolute, and root mean square. More information is available on my project web site.
Methods for identifying key features in complex images are improving. Vectorial tracking is a method used to trace blood vessels in complex high resolution vasculature images. The algorithm is described in a paper written at Rensselaer Polytechnic Institute [1]. Vectorial tracking does recursive tracing of vasculature based on a localized model. The method will be applied to work being done at the University of Wisconsin. Preliminary results are included. More information is available on my project web site.
Road extraction from remotely sensed image is a difficult task due to complex characteristics of roads along their extents in digital imagery. Due to imperfections in primary approximations of road regions and occlusions, true road regions may be deleted or undetected, or some falsely detected road may exist. Hence, the extracted road lines need to be tracked to construct a complete network according to their local conditions. The purpose of this project is to construct a search region to determine if a gap between disconnected road segments should be bridged or not based on slope variance within the search region. The specific goals of this project are as follows:
More information is available on my project web site.
An interactive tool for selecting objects from images using simple freehand sketches is described, and various extensions for improving the quality of the selection are introduced. The tool uses a two phase algorithm for segmentation. The first phase consists of global image segmentation and processing of user input to produce an initial foreground region. The centroids of the segments formed as a result of the global segmentation are triangulated to produce a representation of the image that allows for easy detection of boundary regions of the foreground. In the second phase, these boundary regions are processed along with a piece-wise linear approximation of the initial foreground region to produce a high quality final mask. This tool allows users to extract regions with complex boundaries from images and has immediate applications in image editing, manipulation and processing. More information is available on my project web site.
An algorithm based on OCRchie is reviewed and implemented to recognize scanned in files. The algorithm is carefully analyzed step by step. A few difficult situations are separated out to give special solutions. OCRchie is robust for scanned in files because of the invariant OCR vector for scaling and translation. However, this method does not apply to projective or affine transform because the invariance no long holds for such condition. More information is available on my project web site.
"Color by Numbers" is a popular children's coloring activity. A black and white segmented image has colors assigned to each region and it is then the job of the child to color in a region with its respective color that is denoted by an assigned number. This paper implements an algorithm for generating such "Color by Numbers" art from a color input image.
This paper augments Lucchese and Mitra's color segmentation algorithm described in "An Algorithm for Fast Segmentation of Color Images." Lucchese and Mitra's algorithm uses several parameters that affect the quality of the segmentation. The "Color by Numbers" algorithm presented extends the segmentation to labeled components to achieve the final image. More information is available on my project web site.
Recent advances in texture synthesis have provided effective algorithms for producing high quality textures a fraction of the time taken by older algorithms. One such algorithm is Image Quilting as proposed by Efros and Freeman. In image quilting, a new texture image is synthesized by selecting patches from the input texture and 'quilting' them together. Since the algorithm adds patches to the new image instead of pixels, a high level of efficiency is accomplished. The images produced are of high quality, but not quite as high as earlier pixel based approaches do to patch boundary artifacts. This paper outlines an implementation of the image quilting algorithm and then analyzes some simple extensions designed to remove latch boundary irregularities. More information is available on my project web site.
For this project I would like to do research on and implement a gesture recognition system. I will do a survey of the current techniques used for gesture recognition and implement one or more techniques. The data that I will be primarily using is of a lecturer in a classroom. From previous work that I have done, I am able to "tell" when and where writing occurs on the chalkboard. I would like to use this information in order to improve on the gesture recognition system that I implement. More information is available on my project web site.
Large DNA molecules bearing transcription complexes are stretched out on the optical mapping surface (chemically modified glass coverslip). YOYO-1 (fluorescence dye) staining makes DNA molecules visible under epifluorescence microscope, which can also collect the image by mounted CCD camera(see figure). In this project, image segmentation will be applied to extract interesting DNA molecule object. Secondly, the linear DNA molecule information will be extracted from each object. Thirdly, the further edge detection along the DNA molecule will be applied to filter out complex information. Specific cases, like overlapping DNA molecules, noisy spots will be detected, analyzed and selectively eliminated during this analysis.