W11 EECS 556 final project abstracts 1 1:40-2:05 Compressed sensing: Image reconstruction via l1 regularization Nibal Arzouni, Madison McGaffin, Matthew Prelee, Benjamin Schwartz In signal processing, the Nyquist-Shannon Sampling Theorem is the traditional rule of thumb for sampling with perfect reconstruction. In recent years, a new theory called compressed sensing has been developed that suggests new lower bounds on the number of samples necessary to reconstruct a signal. These lower bounds are achieved by taking advantage of the sparsity that some signals exhibit when represented in certain domains. In this project, we used sparsity-based regularization to reconstruct an image from a small number of samples taken in an appropriate "incoherent" domain. We focused on two classes of images: MR-images and natural images, are often empirically sparse in the Haar-wavelet basis. We achieved high-PSNR image reconstructions from data sampled at rates well below the conventional threshold. 2 2:05-2:30 Chromatic superresolution imaging Fred Buchanan, Mary Lin, Rob Moran This project attempts to enhance low resolution images to a higher resolution using superresolution algorithms. This allows us to obtain details from an image just like they do in TV shows like CSI where they would like to enhance the resolution and figure out the license plate of a car on a security camera. We implemented a robust and fast algorithm that was proposed by Peyman Milanfar. It uses a two stage cost minimization method where we 1) fuse the low resolution images together and 2) deblur and interpolate appropriately. Many previous algorithms of superresolution were unable to be extended to work effectively for chromatic (or color) images. However, Milanfar's method creates a very specific cost function that minimizes luminance, chrominance, and color channel orientation in addition the regular data fidelity penalty function. Our implementation is focused on the actual superresolution algorithm and analysis of dependencies on different motion estimation algorithms. 3 2:30-2:55 Investigation of the non local means algorithm Brittan Farmer, Arun Sundar Govindarajan, Raj Tejas Suryaprakash Image denoising is a fundamental processing step in image processing, since noise is inherent in image acquisition. It is also required as a necessary pre-processing step to improve performance of higher level operations such as edge detection. In this project, we investigate the Non Local Means (NL Means) algorithm for image denoising, which was proposed in [1]. The NL Means algorithm exploits self-similarity in images (e.g.: texture images). If the neighborhood of a pixel, say f[i,j], is similar to the neighborhood around the original pixel being considered, then pixel f[i,j] is assigned a higher weight for averaging. Since neighborhoods are compared instead of individual pixel values, the algorithm is sensitive to image features. We compare the performance of this algorithm to the performance of the algorithms considered in class, such as NPLS. 4 2:55-3:20 KLA challenge: detection of defects in integrated circuits Xiyu Duan, Chris Fink, Hao Sun, and Meng Wu KLA-Tencor has challenged students to develop automated algorithms to detect as many defects as possible in images of integrated circuits, with a zero false-positive rate. KLA-Tencor has supplied twelve image triplets, five with defects and seven without (each image triplet consists of two defect-free images and one image with potential defects). KLA-Tencor has also supplied a simple baseline algorithm which detected 154 out of 209 defects. We describe several of the approaches we employed in our attempt to improve upon KLA's algorithm. These included image interpolation and alignment, Fourier filtering, PCA analysis, Markov Random Field segmentation, morphological image processing, and wavelet analysis. In the end, two very simple approaches gave the best results, combining to detect a total of 187 defects with a zero false-positive rate. [5 minute break] 5 3:25-3:50 Playing card recognition system Ming-Kai Ko, Josh Mann, Kyle Polack, Yi-Hsuan Tsai Playing card recognition systems have several applications in the gaming industry, including the incorporation into casino security systems. This project demonstrates the recognition of both rank and suit of standard playing cards against a simulated casino tabletop by a birds-eye-view camera. Contemporary methods of playing card recognition rely on detection of card edges to locate cards, scan for information and determine rotation angles. This reliance on edge detection poses limitations on card orientation and tends to reduce accuracy when multiple cards are present. This project eliminates the requirement of visible card edges by using a combination of color, size and distance filters to locate the regions containing rank and suit information. Once identified, the angle of rotation of each information region is determined using the radon transform of the rank component. Each piece of information is then rotated and scaled and is finally compared to a library of high-resolution binary images of each rank and suit. Comparison is made by applying an erosion filter to the absolute difference of the region and its respective set of library images. This method requires that only a single information-carrying region be visible for proper card identification, which should increase identification rates over contemporary methods in situations with multiple, overlapping cards laying at different angles of rotation. 6 3:50-4:15 Scenery auto-creating based on deep learning Guanyu Zhou / Tianpei Xie / Yajing Chen / Jiaying Yu Modern machine learning algorithms have received increasing attention by research and industrial communities in the past two decades. However, when learning algorithms are applied to real problem, major task are often passive. Instead of passive learning datasets, we want to create new data that belongs to a specific category to send out meaningful information. Our project processes 2D images by developing the machine learning algorithms to construct a feature pattern from a series of images related to a user-specific theme. We focus on the integration of machine learning algorithm and the signal generation algorithm that means we not only establish a certain judgment on input images, but also create a new image based on accumulated knowledge. Main technologies of our project include stacked autoencoder, image segmentation, normalization and deblurring. We test our algorithm on both digit image and nature images. 7 4:15-4:40 Real-time simultaneous localization and mapping with loop closure Gaurav Pandey, Pat O'Keefe, Paul Ozog This project is a real-time implementation of Simultaneous Localization and Mapping (SLAM) with Loop Closure. SLAM is a technique used by robots and autonomous vehicles to construct a map of an unknown environment while simultaneously positioning the robot in the map. Loop closure is an event that takes place during the SLAM process where the robot recognizes it has returned to a position it has been before. Once a loop closure event has been identified, the accumulated error of the robot position is reduced not only at the current location, but throughout the entire map. The ultimate goal of SLAM is to allow a robot to robustly answer the questions "What does the world look like?" and "Where am I?" For our final implementation, we have used techniques and concepts such as SIFT and SURF feature extraction, hierarchical k-means, vocabulary trees, and general solutions to optimization problems. We present the results of our real-time C++ implementation by applying it to a set of data collected from a specially outfitted truck in an urban environment. 8 4:40-5:05 Using optical flow plane detection and depth maps for augmented reality Leng-Chun Chen, Yu-Hui Chen, Yi-Sing Hsiao, Srinath Sridhar A three-dimensional geometry of a given scene was reconstructed on pairs of pictures by the optical flow plane detection and depth map estimation. The reconstructed geometry then provided a 3-D space for implementing a augmented reality. Augmented Reality(AR) has been well known with a lot of applications including entertainment, take support, or navigation. The implementation of AR requires a known geometry of the scene as well as camera positions. Currently, most people achieved these requirements by putting specific `marker' patterns into the scene or by using extra sensors to measure the camera's position. However, these methods would not be feasible when applied to some already taken pictures. Thus, our method, providing the 3D geometry of a given scene by image processing, could build a spacial template for AR implementation without manipulation of a scene when pictures were taken. 9 5:05-5:30 The mystery of stereograms! - synthesis and analysis methods Bing Liao, Mahta Mousavi, Yun Xu, Yuanhao Zhai Stereogram is a single image, designed to create a visual illusion of a three dimensional scene, from a two dimensional image in the human brain. There are tricks for the naked eye to help make seen the hidden image behind a stereogram, however, in this project we tried to implement codes to do this for us. We first implemented simple synthesis and analysis methods for images with uniform depth level and tried to restore the shape as much as possible applying different filters and restoration methods. Then we implemented a more sophisticated algorithm for images with more depth levels and did the image enhancement after restoration to improve the minimum mean squared error. Then we added noise to the stereogram and investigated the effects and modified the scheme to achieve the best restoration result. Finally, we tried our method on animated autostereograms to restore the embedded animation.