Search

Crowd counting is a highly challenging problem in computer vision and machine learning. Most previous methods have focused on consistent density crowds, i.e., either a sparse or a dense crowd, meaning they performed well in global estimation while neglecting local accuracy. To make crowd counting more useful in the real world, we propose a new perspective, named pan-density crowd counting, which aims to count people in varying density crowds. Specifically, we propose the Pan-Density Network (PaDNet) which is composed of the following critical components. First, the Density-Aware Network (DAN) contains multiple subnetworks pretrained on scenarios with different densities. This module is capable of capturing pandensity information. Second, the Feature Enhancement Layer (FEL) effectively captures the global and local contextual features and generates a weight for each density-specific feature. Third, the Feature Fusion Network (FFN) embeds spatial context and fuses these density-specific features. Further, the metrics Patch MAE (PMAE) and Patch RMSE (PRMSE) are proposed to better evaluate the performance on the global and local estimations. Extensive experiments on four crowd counting benchmark datasets, the ShanghaiTech, the UCF-CC-50, the UCSD, and the UCFQNRF, indicate that PaDNet achieves state-of-the-art recognition performance and high robustness in pan-density crowd counting.

Large-scale three-dimensional (3D) seismic modeling is considered the foundation of imaging and inversion. Parallelization strategies are crucial to solving such a computationally intensive problem. In this paper, we focus on two factors, decomposition direction and decomposition dimension, that significantly affect the computational performance. The decomposition direction determines the cache hit ratio during register addressing, and the decomposition dimension influences the communication size. We thoroughly analyze these two factors by performing time-space domain staggered-grid finite-difference (SGFD) modeling with a set of decomposition strategies. Four metrics, including computation time, speedup ratio, strong scaling property, and memory usage, are introduced to evaluate the computational performance of each trial. After theoretical analysis and metrics testing, we conclude that the optimized domain decomposition strategy is: decomposing models at two dimensions, the decomposition directions are perpendicular to the fastest and the second fastest dimensions, here we refer the dimension in which data are continuously saved as the fastest dimension. Three examples further verify the feasibility and efficiency of the optimized parallel scheme. Considering that domain decomposition-based 3D seismic parallel simulation packages are seldom available in the public domain, we provide a program template for the optimized domain decomposition strategies as an open-source package.

Since its introduction in 2006, Hadoop technology has evolved dramatically and the Hadoop ecosystem has flourished. The Hadoop ecosystem is now composed of more than 60 components, ranging from HDFS and MapReduce. Hadoop’s architecture makes it far superior to other products in large-scale data processing and analysis, making it the best choice for all industries. This makes Hadoop the preferred framework for data analysis and processing in a variety of industries. With the widespread application of Hadoop, people are increasingly concerned about the data trustworthiness of HDFS such as the explicit storage of data and over-reliance on authentication mechanisms. In order to ensure the trustworthiness of Hadoop data and to take into account the performance factors of the big data framework, this study focuses on the HDFS-based data trustworthiness problem and the use of node classification, data encryption, trustworthiness measures and other measures to comprehensively enhance the trustworthiness of HDFS file system, improve the existing HDFS system, and ensure the data storage communication in a trustworthy environment.

A number of studies on lung nodule classification lack clinical/biological interpretations of the features extracted by convolutional neural network (CNN). The methods like class activation mapping (CAM) and gradient-based CAM (Grad-CAM) are tailored for interpreting localization and classification tasks while they ignored fine-grained features. Therefore, CAM and Grad-CAM cannot provide optimal interpretation for lung nodule categorization task in low-dose CT images, in that fine-grained pathological clues like discrete and irregular shape and margins of nodules are capable of enhancing sensitivity and specificity of nodule classification with regards to CNN. In this paper, we first develop a soft activation mapping (SAM) to enable fine-grained lung nodule shape \& margin (LNSM) feature analysis with a CNN so that it can access rich discrete features. Secondly, by combining high-level convolutional features with SAM, we further propose a high-level feature enhancement scheme (HESAM) to localize LNSM features. Experiments on the LIDC-IDRI dataset indicate that 1) SAM captures more fine-grained and discrete attention regions than existing methods, 2) HESAM localizes more accurately on LNSM features and obtains the state-of-the-art predictive performance, reducing the false positive rate, and 3) we design and conduct a visually matching experiment which incorporates radiologists study to increase the confidence level of applying our method to clinical diagnosis.

Seismic inversion is a highly ill-posed problem, due to many factors such as the limited seismic frequency bandwidth and inappropriate forward modeling. To obtain a unique solution, some smoothing constraints, e.g., the Tikhonov regularization are usually applied. The Tikhonov method can maintain a global smooth solution, but cause a fuzzy structure edge. In this paper we use Huber-Markov random-field edge protection method in the procedure of inverting three parameters, P-velocity, S-velocity and density. The method can avoid blurring the structure edge and resist noise. For the parameter to be inverted, the Huber-Markov random-field constructs a neighborhood system, which further acts as the vertical and lateral constraints. We use a quadratic Huber edge penalty function within the layer to suppress noise and a linear one on the edges to avoid a fuzzy result. The effectiveness of our method is proved by inverting the synthetic data without and with noises. The relationship between the adopted constraints and the inversion results is analyzed as well.

Visual prostheses based on micro-electronic technologies and biomedical engineering have been demonstrated to restore vision to blind individuals. It is necessary to determine the minimum requirements to achieve useful artificial vision for image recognition. To find the primary factors in common object and scene images recognition and optimize the recognition accuracy on low resolution images using image processing strategies, we investigate the effects of two kinds of image processing methods, two common shapes of pixels (square and circular) and six resolutions (8x8, 16x16, 24x24, 32x32, 48x48 and 64x64). The results showed that the mean recognition accuracy increased with the number of pixels. The recognition threshold for objects was within the interval of 16x16 to 24x24 pixels. For simple scenes, it was between 32x32 and 48x48 pixels. Near the threshold of recognition, different image modes had great impact on recognition accuracy. The images with ''threshold pixel number and binarization-circular points'' produced the best recognition results.

Visual prostheses for extreme vision impairment have come closer to reality during these few years. The task of this research has been to design exoteric devices and study image processing algorithms and methods for different complexity images. We have developed a real-time system capable of image capture and processing to obtain most available and important image features for recognition and simulation experiment based on DSP (Digital Signal Processor). Beyond developing hardware system, we introduce algorithms such as resolution reduction, information extraction, dilation and erosion, square (circular) pixelization and Gaussian pixelization. And we classify images with different stages according to different complexity such as simple images, medium complex images, complex images. As a result, this paper will get the needed signal for transmitting to electrode array and images for simulation experiment.

A study and results about pixelized images recognition in simulated prosthetic vision were presented. Twenty object and five scene images were chosen from the databases which were almost familiar by everyone. Two kinds of image processing methods (binarization directly and edge extraction from low resolution images), two common shapes of pixel (square and circular) and six pixels numbers(8×8, 16×16, 24×24, 32×32, 48×48 and 64×64) were used to form pixelized images and presented on head-mounted display (HMD) one by one. According to the trials, the mean recognition accuracy increased with the increase of pixel array number. The threshold of objects recognition was within the interval of 16×16 to 24×24. For simple scenes, it was between 32×32 and 48×48. The images with the threshold resolution, binarization method and circular pixel shape have shown the best results for recognition.

The goal of the optic nerve stimulation is to place multielectrode electrodes around the optic nerve and evoke multiplephosphenes covering the entire visual space by using complex patterns of electrical stimulation. This paper mainly studies the best image processing strategies applying to the optic nerve stimulation. Four image processing strategies that could be used in the optic nerve stimulation such as histogram equalization, edge detection, edge following, binary morphological operations. As we learn more about the neural coding of visual system, the better image processing strategies will be applied to the artificial vision.