Hyon Soo Lee

Mean shift-based SIFT keypoint filtering for region-of-interest determination

This paper presents an improved keypoint filtering method for region-of-interest (ROI) determination. Mean shift-based clustering was employed to group the scale invariant feature transform (SIFT) keypoints that appeared in the nearest region to get more locality. The proposed method uses the location of the extracted SIFT keypoints for grouping, and an average SIFT descriptor is calculated on the clustered keypoints. The support vector machine (SVM) classifies the average SIFT descriptor as an artificial or a natural keypoint. After the keypoint classification, only the keypoints classified as artificial keypoints by the binary SVM are used in near-duplicate detection (NDD). Finally, we determine the ROI using the adaptive selection of orientation histogram and the elimination of isolated keypoints. According to the result of experiments on keypoint classification, NDD and ROI determination, the proposed method obtained improved results compared to the previous methods.

The implementation of 2D FFT using multiple topology on 4×4 Torus

In this paper, we proposed 2D FFT for 8×8 matrix without transpose of data by using multiple topology on 4×4 Torus. The proposed 2D FFT used parallel operation on 1D FFT and applied an effective calculation by executing a pipeline operation. We implement the proposed architecture on Xilinx Virtex-IV device and a detailed evaluation has been reported based on maximum system frequency, chip area and image size. The implementation results show that the core speed of the proposed FFT architecture is around 157.3MHz and it occupies 11733 slices. The average SQNR for various images is 61.9dB. To compare the proposed 2D FFT with other methods, we can see that frame per second is improved 8 times.

An efficient score function generation algorithm with information maximization

In this study, we propose this new algorithm that generates score function in ICA (Independent Component Analysis) using entropy theory. To generate score function, estimation of probability density function about original signals are certainly necessary and density function should be differentiated. Therefore, we used kernel density estimation method in order to derive differential equation of score function by original signals. After changing the formula to convolution form to increase speed of density estimation, we used FFT algorithm which calculates convolution faster. Proposed score function generation method reduces estimation error, it is density difference of recovered signals and original signals. Also, we insert constraint which is able to information maximization using smoothing parameters. In the result of computer simulation, we estimate density function more similar to original signals compared with Extended Infomax algorithm and Fixed Point ICA in blind source separation problem and get improved performance at the SNR (Signal to Noise Ratio) between recovered signals and original signals.

Golf swing analysis system with a dual band and motion analysis algorithm

In this paper, we developed a golf position analysis system with an inertial sensor and swing motion analysis algorithm that can be used in golf courses. Specifically, we developed a band for both wrists that analyzes the golf swing of whoever is wearing it. By analyzing the movements of both the right and left wrists of the golfer, their golf swing can be judged more accurately, which can be used to advise the golfer regarding their golf swing. Experimental results show that this portable instrument along with its swing motion detection algorithm is a promising training tool for golfers.

Implementation of sin/cos Processor for Descriptor on SIFT

The SIFT algorithm is being actively researched for various image processing applications including video surveillance and autonomous vehicle navigation. The computation of sin/cos function is the most cost part needed in whole computational complexity and time for SIFT descriptor. In this paper, we implement a hardware to sin/cos function of descriptor on sift feature detection algorithm. The proposed Sin/Cosine processor is coded in Verilog and synthesized and simulated using Xilinx ISE 9.2i. The processor is mapped onto the device Spartan 2E (XC2S200E-PQ208-6). It consumes 149 slices, 233 LUTs and attains a maximum operation frequency of 60.01 MHz. As compared with the software realization, our FPGA circuit can achieve the speed improvement by 40 times in average.

Multi-level independent component analysis

This paper presents a new method which uses multi-level density estimation technique to generate score function in ICA (independent Component Analysis). Score function is very closely related with density function in information theoretic ICA. We tried to solve mismatch of marginal densities by controlling the number of kernels. Also, we insert a constraint that can satisfy sufficient condition to guarantee asymptotic stability. Multi-level ICA uses kernel density estimation method in order to derive differential equation of source adaptively score function by original signals. To increase speed of kernel density estimation, we used FFT algorithm after changing density formula to convolution form. Proposed multi-level score function generation method reduces estimate error which is density difference between recovered signals and original signals. We estimate density function more similar to original signals compared with existent other algorithms in blind source separation problem and get improved performance in the SNR measurement.

Design of new kernel density estimator for entropy maximization in independent component analysis

This paper proposes a new algorithm for estimating the score function using maximum entropy theory and kernel density estimation. The main idea is to control smoothing parameters for maximizing entropy in kernel density estimation. To generate score function, directly partial derivative equation from kernel density estimator is derived. To find suitable smoothing parameter, we adopted gradient descent method. Finally, the new kernel density estimator is experimented in blind separation and discuss on properties of the proposed learning algorithm.

Efficient strategies for error updating to improve performance backpropagation learning

There exists a neuron oscillation generated among neurons of the output layer and pattern oscillation generated due to correlation between patterns in error backpropagation learning. Because such oscillations have different features and originate in a mutually correlative situation, there exists the phenomenon that learning time lengthens considerably and convergency is fallen in the existing method that solves two oscillations by means of one learning strategy. In this paper, the authors propose learning strategies that correspond to the feature of each oscillation and apply a learning strategy that is suitable for the problem adaptively when learning a given problem. In order to show the effectiveness of the proposed learning strategies, the authors compared them with existing methods by applying them to 4-6 parity problems, seven segment display and pattern recognition. With the result that, learning time decreased considerably and convergence increased remarkably from the existing methods.