In this study, novel deep learning models based on high-speed flame images are proposed to diagnose the combustion instability of a gas turbine. Two different network layers that can be combined with any existing backbone network are established—(1) An early-fusion layer that can learn to extract the power spectral density of subsequent image frames, which is time-invariant under certain conditions. (2) A late-fusion layer which combines the outputs of a backbone network at different time steps to predict the current combustion state. The performance of the proposed models is validated by the dataset of high speed flame images, which have been obtained in a gas turbine combustor during the transient process from stable condition to unstable condition and vice versa. Excellent performance is achieved for all test cases with high accuracy of 95.1–98.6% and a short processing time of 5.2–12.2 ms. Interestingly, simply increasing the number of input images is as competitive as combining the proposed early-fusion layer to a backbone network. In addition, using handcrafted weights for the late-fusion layer is shown to be more effective than using learned weights. From the results, the best combination is selected as the ResNet-18 model combined with our proposed fusion layers over 16 time-steps. The proposed deep learning method is proven as a potential tool for combustion instability identification and expected to be a promising tool for combustion instability prediction as well.

We present two algorithms for aligning two colored point clouds. The two algorithms are designed to minimize a probabilistic cost based on the color-supported soft matching of points in a point cloud to their K-closest points in the other point cloud. The first algorithm, like prior iterative closest point algorithms, refines the pose parameters to minimize the cost. Assuming that the point clouds are obtained from RGB-depth images, our second algorithm regards the measured depth values as variables and minimizes the cost to obtain refined depth values. Experiments with our synthetic dataset show that our pose refinement algorithm gives better results compared to the existing algorithms. Our depth refinement algorithm is shown to achieve more accurate alignments from the outputs of the pose refinement step. Our algorithms are applied to a real-world dataset, providing accurate and visually improved results.

RGB-Depth (RGB-D) cameras are widely used in computer vision and robotics applications such as 3D modeling and human–computer interaction. To capture 3D information of an object from different viewpoints simultaneously, we need to use multiple RGB-D cameras. To minimize costs, the cameras are often sparsely distributed without shared scene features. Due to the advantage of being visible from different viewpoints, spherical objects have been used for extrinsic calibration of widely-separated cameras. Assuming that the projected shape of the spherical object is circular, this paper presents a multi-cue-based method for detecting circular regions in a single color image. Experimental comparisons with existing methods show that our proposed method accurately detects spherical objects with cluttered backgrounds under different illumination conditions. The circle detection method is then applied to extrinsic calibration of multiple RGB-D cameras, for which we propose to use robust cost functions to reduce errors due to misdetected sphere centers. Through experiments, we show that the proposed method provides accurate calibration results in the presence of outliers and performs better than a least-squares-based method.

In the last stage of colored point cloud registration, depth measurement errors hinder the achievement of accurate and visually plausible alignments. Recently, an algorithm has been proposed to extend the Iterative Closest Point (ICP) algorithm to refine the measured depth values instead of the pose between point clouds. However, the algorithm suffers from numerical instability, so a postprocessing step is needed to restrict erroneous output depth values. In this paper, we present a new algorithm with improved numerical stability. Unlike the previous algorithm heavily relying on point-to-plane distances, our algorithm constructs a cost function based on an adaptive combination of two different projected distances to prevent numerical instability. We address the problem of registering a source point cloud to the union of the source and reference point clouds. This extension allows all source points to be processed in a unified filtering framework, irrespective of the existence of their corresponding points in the reference point cloud. The extension also improves the numerical stability of using the point-to-plane distances. The experiments show that the proposed algorithm improves the registration accuracy and provides high-quality alignments of colored point clouds.

We present a depth-based fingertip recognition method for interactive projectors. We use a depth camera attached to a projector, so it is possible to change the relative pose between the projector and the projection surface without manual recalibration. For detection and classification of fingertips, we propose to use cascaded random forests boosted by our 3D pose-normalized pixel-difference features. The ensemble probabilities from the cascaded random forests are used to define a score function of a subset of detected fingertips. By finding the subset maximizing the score function, the fingertips in the subset are correctly classified, and the remaining incorrectly detected fingertips are rejected. Experiments show that the proposed method outperforms conventional random forest and convolutional neural network classifiers. In addition, our developed applications show the advantage of the proposed method in assigning different roles to different fingers.

Nonlocal means filtering is an edge-preserving denoising method whose filter weights are determined by Gaussian weighted patch similarities. The nonlocal means filter shows superior performance in removing additive Gaussian noise at the expense of high computational complexity. In this paper, we propose an efficient and effective denoising method by introducing a separable implementation of the nonlocal means filter and adopting a bilateral kernel for computing patch similarities. Experimental results demonstrate that the proposed method provides comparable performance to the original nonlocal means, with lower computational complexity.