Doo-Hyun Choi

A new reinforcement learning vehicle control architecture for vision-based road following

A new dynamic control architecture based on reinforcement learning (RL) has been developed and applied to the problem of high-speed road following of high-curvature roads. Through RL, the control system indirectly learns the vehicle-road interaction dynamics, knowledge which is essential to stay on the road in high-speed road tracking. First, computer simulation has been carried out in order to test stability and performance of the proposed RL controller before actual use. The proposed controller exhibited a good road tracking performance, especially on high-curvature roads. Then, the actual autonomous driving experiments successfully verified the control performance on campus roads in which there were shadows from the trees, noisy and/or broken lane markings, different road curvatures, and also different times of the day reflecting a range of lighting conditions. The proposed three-stage image processing algorithm and the use of all six strips of edges have been capable of handling most of the uncertainties arising from the nonideal road conditions.

A new evolutionary programming approach based on simulated annealing with local cooling schedule

The NPOSA (New Population-Oriented Simulated Annealing) technique is introduced as an efficient global search tool to solve optimization problems. Unlike the conventional simulated annealing or its hybrid algorithms, each individual in the population can intelligently plan its own annealing schedule in an adaptive fashion to the given problem at hand. This not only enhances the search speed but furthermore yields a solution near the global optimum. This technique has been applied to solve the traveling salesman problem (TSP) for combinatorial optimization, as well as a continuous function optimization problem, to demonstrate its validity and effectiveness.

Multiscale Detection of Defect in Thin Film Transistor Liquid Crystal Display Panel

A new automated inspection algorithm is proposed for detecting blob-Mura defects based on multiscale in a thin film transistor liquid crystal display (TFT-LCD) panel. As such, new kernels with different sizes are defined and then used to detect blob-Mura defects with varying sizes and brightness levels. To extract a seed point, an adaptive multilevel-threshold method is employed. Initially, smaller kernels are used to detect smaller defects, then gradually larger kernels are applied to detect larger defects. Through simulation it was verified that the proposed algorithm has a superior capability for detecting blob-Mura defects.

Iris feature extraction and matching based on multiscale and directional image representation

This paper presents a new filterbank-based iris recognition method that effectively extracts the spatial and directional features of iris patterns on multiple scales, then performs matching. First, the proposed method localizes the iris area from an input image and establishes a region of interest (ROI) for feature extraction. Second, the iris features are extracted on multiple scales from the ROI and a feature vector generated using a band pass filter and directional filter bank (DFB), which decomposes the image into several directional subband outputs. Finally, iris pattern matching robust to various rotations of the input is performed based on finding the Hamming distance between the corresponding feature vectors. Experimental results demonstrate that the proposed method is both effective in extracting directional and multiresolutional features from iris patterns and robust to input image rotation due to head tilt.

Wavelet-Based Image Enhancement for Defect Detection in Thin Film Transistor Liquid Crystal Display Panel

This paper proposes a wavelet-based prepossessing method to improve the detecting capacity of a blob-Mura-defect-detecting algorithm. The non-uniformity of the background region is eliminated by replacing the approximation coefficients with a constant value, and the brightness difference between the background region and defect regions is increased by multiplying the detail coefficients and a weighting factor. The proposed method can perfectly control the detectable defect level by properly selecting the defect detecting level. Experimental results demonstrate that the proposed method can effectively enhance blob-Mura defects in thin film transistor liquid crystal display panels.

Morphological Blob-Mura Defect Detection Method for TFT-LCD Panel Inspection

The current paper proposes a blob-Mura defect detection method for TFT-LCD panel inspection. A new constraint function that can grow and shrink is defined. Specially, a morphology-based preprocessing method is proposed to improve the detecting capacity of a blob-Mura-defect-detecting algorithm, whereby a test image with blob-Mura defects is expanded to facilitate the defect detection. Plus, in the case of defects with a diameter over 49 pixels, which are hard to detect due to the non-uniformity of their interior, the proposed method changes the image size instead of the constraint function size. The proposed method enables superior defect detection and the use of a simple detecting algorithm.

TD based reinforcement learning using neural networks in control problems with continuous action space

While most of the research on reinforcement learning assumed a discrete control space, many of the real world control problems need to have continuous output. This can be achieved by using continuous mapping functions for the value and action functions of the reinforcement learning architecture. Two questions arise here however. One is what sort of function representation to use and the other is how to determine the amount of noise for search in action space. The ubiquitous back-propagation neural network is used here to learn the value and action functions. Next, the reinforcement predictor that is intended to predict the next reinforcement is introduced that also determines the amount of noise to add to the controller output. This proposed reinforcement learning architecture is found to have a sound online learning control performance through a computer simulation of the ball and beam system as an example plant.

A Flashover Prediction Method for Contaminated Insulators using a Stochastic Analysis of Leakage Current

In this paper, we present the results of a new stochastic analysis of the leakage current on contaminated insulators under salt fog conditions. The stochastic analysis of the leakage current was conducted using a Hilbert transform and the level crossing rate. The Hilbert transform is used for the leakage current envelope, while the level crossing rate is used to convert the upper envelope into a cdf. Using a new defined weighting factor, the resulting cdfs show a probability that the stages until flashover can be discriminated from each other. Therefore, the proposed method is effective for flashover prediction and for monitoring the contamination conditions of outdoor insulators.

Scale-based image enhancement using modified anisotropic diffusion filter

We propose a scale-based image enhancement method using a modified anisotropic diffusion filter that employs sensor noise estimation and scale space methods based on a minimum reliable scale. Based on the relationship between the local gradient and the critical value function, an image is classified into various scales representing the regions complexity. The degree of enhancement is then adjusted according to each regions complexity. When compared to the existing linear unsharp masking, adaptive unsharp masking, and warping-based methods, the proposed algorithm produces negligible noise amplification in homogeneous regions and superior edge enhancement.

Fast evolutionary programming through search momentum and multiple offspring strategy

A new algorithm that helps to accelerate convergence as well as to enhance the diversity of the evolutionary programming (EP) search technique is proposed, based on an individual structure concept. The major components of the algorithm that lie behind its good performance includes scaling, selection strategy, the use of age and the search direction (or momentum) vector, and multiple offspring per parent. Not only are the multiple offspring approach and the search direction vector concept novel but the combination of these features used for EP is also new. Through a benchmark test, its search performance has been found to be superior to the conventional EP and one of its acceleration methods.