In this study, a singular points detection method based on low-resolution image processing technique and Poincare index algorithm was introduced. First, 2D discrete wavelet transform (DWT) was used to conduct fingerprint image operation, then a LL-band image that has 1∕4 resolution of the original image was obtained to reduce the operation needed in the subsequent processing. After finishing the detection procedure of singular point, each possible singular point was taken as the center to calculate the local binary pattern (LBP) features at its peripheral as reference standard for selecting correct singular point. Through this way, fake singular point is removed. The experimental result shows that the method in this study could indeed reduce the chance to detect fake singular points due to noise effect. Meanwhile, the processing speed of the traditional Poincare index method can be speeded up.

A compounded capacitive micromachined ultrasound transducer (CMUT) using two SOI wafers is proposed. This transducer has the potential in multi-frequency transducer applications and medical application for both imaging and therapy. The compounded CMUT (CCMUT) has a composite design with both low- and high-frequency structures built into one compartment of CCMUT. The effective parameters on collapse voltage including air gap, thickness and radius of the membrane in CMUT are simulated. The capacity in CMUT and electromechanical efficiency are also simulated. The critical parameters related including collapse voltage, capacity and electromechanical efficiency are analyzed and discussed. This CCMUT can be utilized in medical imaging, therapy or multi-frequency application.

Construction managers periodically take photographs of on-site resources such as materials and facilities to document various aspects of construction activities. Although digital images can be effectively used for monitoring construction activities, they are not used at all but to show the situation of the site by being attached on the report. In this regard, this study proposes a system and a smart helmet that help site managers identify changes in the conditions of facilities and materials. The smart helmet is equipped with a small camera to record videos around the site, and a small GPS to collect the position data of the site manager wearing the smart helmet. The system includes two separate frameworks—one for fixed resources and the other for mobile resources. The system automatically detects changes in appearance of fixed resources and in locations of mobile resources. The frameworks involve image matching methods which play critical roles in detecting appearance changes of fixed resources as well as cross-checking the identities of mobile resources. Experimental results signify the system’s potential uses for effectively monitoring the conditions of on-site facilities and materials.

Concerning the uncertainty problem of different environment meteorological models and improving the accuracy of PM_2.5 concentration forecast, in this article, an improved wavelet neural network ensemble algorithm with additional momentum item is proposed and the forecast products of the three environment meteorological models including China Meteorological Administration Unified Atmospheric Chemistry Environment, Beijing Regional Environmental Meteorology Prediction System and Weather Research Forecasting/Chemistry are integrated. The multi-model ensemble rolling forecasting model of PM_2.5 concentration is established. The experiment is carried out by data of Beijing station, and the forecast results are compared with seven other machine learning algorithm models (Wavelet, BP, RBF, Elman, T-S fuzzy, SVR and CNN). The results show that PM_2.5 concentration forecasted by the improved wavelet neural network is better than the other models, and the new method reduces the prediction deviation effectively.

As a basic task of multi-camera surveillance system, person re-identification aims to re-identify a query pedestrian observed from non-overlapping multiple cameras or across different time with a single camera. Recently, deep learning-based person re-identification models have achieved great success in many benchmarks. However, these supervised models require a large amount of labeled image data, and the process of manual labeling spends much manpower and time. In this study, we introduce a method to automatically synthesize labeled person images and adopt them to increase the sample number per identity for person re-identification datasets. To be specific, we use block rectangles to randomly occlude pedestrian images. Then, a generative adversarial network (GAN) model is proposed to use paired occluded and original images to synthesize the de-occluded images that are similar but not identical to the original image. Afterward, we annotate the de-occluded images with the same labels of their corresponding raw images and use them to augment the number of samples per identity. Finally, we use the augmented datasets to train baseline model. The experimental results on CUHK03, Market-1501 and DukeMTMC-reID datasets show the effectiveness of the proposed method.

An efficient video coding scheme is significant for video storage and transmission, especially for the massive data of three-dimensional (3D) video. For 3D video coding, three-dimensional high efficiency video coding (3D-HEVC) supports the multi-view video with depth (MVD) format. Both texture and depth sequences for each view should be coded, and the coding process of depth map follows the principle of texture sequence, which increases the computational complexity of 3D-HEVC. Therefore, two fast algorithms for 3D-HEVC depth map coding are proposed to reduce the coding complexity of 3D-HEVC based on the gray-level co-occurrence matrix (GLCM). With GLCM, the texture complexity of the depth map in 3D-HEVC can be effectively described, which enables the prejudgment of CU segmentation depth and candidate intra-prediction modes. Experimental results show that the proposed algorithms can save about 19.1% and 20.2% coding time compared with the 3D-HEVC standard, and outperform the state-of-the-art fast algorithm by 5.7% and 6.8% time-saving, respectively, while keeping almost the same coding efficiency and quality of synthesized views.

In this article, a new floor estimation algorithm based on multiple deep learning image segmentation and conventional texture segmentations using fuzzy integrals theory is proposed. The proposed algorithm combines an FCN-8s, a DeepLabv2, and Canny Edge Detection with superpixel segmentation, two deep learning networks, and one texture classifier to recognize a walkable floor area for UGV robots. The authors intersect three results with an Improved Fuzzy Integrals (IFI) method. The experimental results show that the combination algorithm accuracy can reach up to 97.63% on average without any other sensor assistance. In order to achieve real-time performance, the proposed algorithm has been implemented on an NVIDIA Jetson TX2 embedded platform with ROS compatible environment supporting.

Object tracking is a hot spot in computer vision and has been developed rapidly in recent years. However, there are still some problems that must be solved in estimation of the state trajectory of the targets (position, orientation, scaling, etc.). In this paper, we focus on the problem of object scale change in tracking and propose scale-adaptive tracking method by combining a kernelized correlation filter with geometric estimation. We combine geometric estimation with low-complexity target scales, so the target scale can be definite during the tracking. Both normal and fault-tolerant target scales are established. Normal target scales are established to seek the optimal solution of the target scale and fault-tolerant target scales are used to correct the results of geometric estimation. Geometric estimation can estimate the target initial size roughly in the early stages of tracking and reduce the scope of scales during the tracking process. Experimental results show that the proposed algorithm is sensitive to the change of target scale and the tracking result is remarkably accurate when target scale changes in practice.

Although wavelet shrinkage is an effective image denoising method, it tends to over-discard the image signal energy and, thus, blur the edges of the produced image. Shrinking the wavelet coefficients of all subbands indifferently is inappropriate because denoising involves not only removing high-frequency signals but also preserving image information to the greatest possible extent. To fulfill these requirements, this study presents an intelligent fuzzy inference system (FIS) learning-based thresholding strategy. First, we propose a principal directional components analysis (PDCA) method for capturing the dominant contours of an image. Along with the principal directions, the directional wavelet transform is used to provide efficient representation of the image. In addition, adaptive directional wavelet packet (WP) decomposition is used to generate the optimal WP tree. Each subband of the WP tree is denoised separately by one of the following methods: total variation denoising, soft shrinkage, and linear interpolation shrinkage. Based on the subband level and diagonality, FIS learning is used to appropriately adjust the subband threshold. Finally, individual estimates are weighted averaged to produce the denoised image. Experimental results show that compared with other denoising methods, our method not only significantly removes heavy noise, preserving more structural edge information, but also provides better peak signal-to-noise ratio and structural similarity index performances.