Search

Due to the dominant position of deep learning (mostly deep neural networks) in various artificial intelligence applications, recently, ensemble learning based on deep neural networks (ensemble deep learning) has shown significant performances in improving the generalization of learning system. However, since modern deep neural networks usually have millions to billions of parameters, the time and space overheads for training multiple base deep learners and testing with the ensemble deep learner are far greater than that of traditional ensemble learning. Though several algorithms of fast ensemble deep learning have been proposed to promote the deployment of ensemble deep learning in some applications, further advances still need to be made for many applications in specific fields, where the developing time and computing resources are usually restricted or the data to be processed is of large dimensionality. An urgent problem needs to be solved is how to take the significant advantages of ensemble deep learning while reduce the required expenses so that many more applications in specific fields can benefit from it. For the alleviation of this problem, it is essential to know about how ensemble learning has developed under the era of deep learning. Thus, in this article, we present fundamental discussions focusing on data analyses of published works, methodologies, recent advances and unattainability of traditional ensemble learning and ensemble deep learning. We hope this article will be helpful to realize the intrinsic problems and technical challenges faced by future developments of ensemble learning under the era of deep learning., Comment: 47 pages, 8 figures, 15 tables

Cell localization is still facing two unresolved challenges: 1) the large variability in cell size and shape, coupled with the heterogeneous intensity distribution of lightly stained cells, presents the bottleneck limiting accurate cell localization; 2) existing cell location map is unreasonable, which loses cell scale information and significantly affects the accuracy of cell counting. To address the challenges from the cell shape and heterogeneous intensity distribution, we propose a novel gradient-aware and shape-adaptive Difference-Deformable Convolution (DDConv), which can focus on the gradient information of lightly stained cells for overcoming the challenge of heterogeneous intensity distribution and meanwhile adaptively adjust the shape of the convolutional kernel for overcoming the challenge of the large variability in cell shape. Moreover, to solve the challenge of unreasonable location map, we propose a new cell location map, called Pseudo-Scale Instance (PSI) map. Our PSI map enables adaptively computing the scale information and associating it with each cell’s annotation, which addresses the unreasonable challenges in existing cell location map and advanced makes the model sensitive to the size of cells. We analyze and evaluate DDConv and PSI on two challenging cell localization tasks. Compared with existing methods, our proposed method has significantly improved the localization performance, setting a new benchmark for cell localization tasks.

In the field of medical image analysis, especially in the study of pathological morphology images of tumor tissues, accurately locating cells in images is a crucial task. However, this task faces multiple challenges, such as differences in cell staining intensity and cell morphology, and variance in cell shape and scale caused by manual stretching. Currently, the solution is to use fixed-scale density maps to supervise the model’s training, which makes it difficult for the model to handle cells of different scales and cell localization scenarios under different magnifications. To address this problem, this paper proposes a new cell location map called Pseudo Scale Instance (PSI) map, which has instance-level scale information. Additionally, to address the problem of significant differences between complex cell shapes and masks in PSI maps, this paper introduces a gradient alignment module based on Difference Deformable Convolution (DDC). The module uses gradient information to calculate the offsets in the deformable convolution, making the model’s output closer to the ground truth. Experiments show that the PSI map proposed in this paper can effectively utilize scale information to improve the baseline performance of cell localization and counting. Moreover, the DDC module can effectively bridge the gap between cell images and ground truth, thereby alleviating the problem of diverse cell shapes. Compared with existing methods, the proposed method has significantly improved the localization performance on two public datasets, setting a new benchmark for cell localization tasks.

Cell localization is an important area of medical image analysis, which is dedicated to predicting the precise location of cells in an image. The existing localization paradigm is to predict the density map using a Convolutional Neural Networks (CNN) model based on vanilla convolution and then process the density map using a local maximum search strategy to obtain the cell location and number information. However, there are three main problems in this paradigm: 1) CNN models based on vanilla convolution have difficulty in handling large variations in cell color; 2) The density map is difficult to provide accurate cell location information and ideal gradient information, and the information loss is more obvious in dense regions; 3) The post-processing strategy of density maps is susceptible to background noise and mutual interference between negative and positive cells. To tackle the above issues, we have made a comprehensive update of the existing paradigm, which consists of three parts: 1) A multi-scale gradient aggregation module based on difference convolution to effectively mitigate the challenge of large variations in cell color; 2) A new exponential distance transform map that provides accurate cell location information along with ideal gradient information for model learning; 3) A post-processing strategy named cell center localization strategy based on location maps that can significantly improve the localization performance. Extensive experiments on multiple datasets show that our approach can substantially improve cell localization and counting performance, establishing a new baseline for the cell localization task, and thereby increasing the efficiency of computer-aided diagnosis.

Background: The prognostic nutritional index (PNI) and different glucose metabolisms have been separately reported to be correlated with long-term prognosis in patients with acute myocardial infarction (AMI) undergoing percutaneous coronary intervention (PCI). However, PNI application in patients with an impaired glucose metabolism has not been well validated, especially in pre-diabetic patients. This study evaluated whether PNI influences a long-term risk of mortality along different glucose metabolism statuses. Methods: A total of 17,697 patients with AMI and a history of PCI were enrolled in this retrospective observational cohort study from January 2007 to December 2020. Three subgroups with different glucose metabolism statuses, including normal glucose regulation (NGR), pre-diabetes mellitus (pre-DM) , and diabetes mellitus (DM) , were divided into three groups according to the tertiles of PNI, respectively. Results: All-cause mortality occurred in 2,613 (14.8%) patients within a median of 4.1 years of follow-up. Upon analyzing the Kaplan–Meier plots for the NGR, pre-DM, and DM groups, the incidence of all-cause or cardiovascular mortality in the low PNI (PNI-L, ≤ 42.7) subgroup was significantly higher than that in the median PNI(PNI-M, > 42.7 and ≤ 48.2) and high PNI (PNI-H, > 48.2) subgroups (all, P < 0.001). After adjusting for confounding factors, the hazard ratio (HR) for all-cause mortality in the PNI-L group significantly increased compared to that in the PNI-H subgroups of the NGR group (HR, 1.35; 95% CI, 1.14–1.66; P < 0.001), pre-DM group (HR, 1.29; 95% CI, 1.02–1.62; P < 0.001), and DM group (HR, 1.36; 95% CI, 1.13–1.63; P < 0.001). Given that there was evidence of interactions between PNI and different glucose statuses(P for interaction < 0.001), patients were divided into nine subgroups, and we found that DM patients with PNI-L statuses had the highest risk of all-cause mortality compared to NGR patients with PNI-H statuses (HR, 1.69; 95% CI, 1.42–2.01; P < 0.001). Conclusion: Lower PNI is a significant and independent risk factor for all-cause mortality in AMI patients undergoing PCI with different glucose metabolism statuses, and this risk further increases with DM compared to NGR or pre-DM statuses.

Neoadjuvant chemotherapy (NAC) is a standard treatment option for locally advanced breast cancer. However, not all patients benefit from NAC; some even get worse outcomes after therapy. Hence, predictors for treatment benefit are crucial for guiding clinical decision-making. Here, we investigated the predictive potentials of breast cancer stromal histology via a deep learning (DL)-based approach and proposed the tumor-associated stroma score (TS-score) for predicting pathological complete response (pCR) to NAC with a multi-center dataset. The TS-score is demonstrated to to be an independent predictor of pCR as it not only outperformed the baseline variables and stromal tumor-infiltrating lymphocytes (sTILs) but also significantly improved the prediction performance of the baseline variable-based model. Further, we discovered that unlike lymphocyte, collagen and fibroblasts in stroma were likely associated with poor response to NAC. The TS-score has potentials to be a candidate for better stratification of breast cancer patients in NAC settings.

BackgroundDifferent definitions of contrast-associated acute kidney injury (CA-AKI) have different predictive effects on prognosis. However, few studies explored the relationship between these definitions and long-term prognosis in patients with congestive heart failure (CHF). Thus, we aimed to evaluate this association and compared the population attributable risks (PAR) of different CA-AKI definitions.MethodsThis study enrolled 2,207 consecutive patients with CHF undergoing coronary angiography (CAG) in Guangdong Provincial People's Hospital. Two different definitions of CA-AKI were used: CA-AKIA was defined as an increase ≥.5 mg/dl or > 25% in serum creatinine (SCr) from baseline within 72 h after CAG, and CA-AKIB was defined as an increase of ≥.3 mg/dl or > 50% in SCr from baseline within 48 h after CAG. Kaplan-Meier methods and Cox regression were applied to evaluate the association between CA-AKI with long-term mortality. Population attributable risk (PAR) of different definitions for long-term prognosis was also calculated.ResultsDuring the 3.8-year median follow-up (interquartile range 2.1-6), the overall long-term mortality was 24.9%, and the long-term mortality in patients with the definitions of CA-AKIA and CA-AKIB were 30.4% and 34.3%, respectively. We found that CA-AKIA (HR: 1.44, 95% CI 1.19-1.74) and CA-AKIB (HR: 1.48, 95% CI 1.21-1.80) were associated with long-term mortality. The PAR was higher for CA-AKIA (9.6% vs. 8%).ConclusionsOur findings suggested that CA-AKI was associated with long-term mortality in patients with CHF irrespective of its definitions. The CA-AKIA was a much better definition of CA-AKI in patients with CHF due to its higher PAR. For these patients, cardiologists should pay more attention to the presence of CA-AKI, especially CA-AKIA.

Due to the difficulty to accurately define the morphologies of Helicobacter Pylori (H. pylori) and the complexity of dealing with the whole slide images (WSIs), no computer-aided solution has currently been presented for detecting H. pylori infection in WSIs. We present the first image semantic segmentation solution for the computer-aided detection of H. pylori in WSIs. The solution only requires polygon annotations as weak supervision, which roughly, instead of pixel-level accurately, label the H. pylori infected areas in WSIs. We propose a new weakly supervised multi-task learning framework (WSMLF) that aims to improve the segmentation performance by more effectively exploiting the weak supervision. To make more effective usage of the weak supervision, we extract multiple inaccurate targets representing different modes of the true target from the available weak annotations. For improvement of the segmentation performance, we design a weakly supervised multi-task learning algorithm that can automatically learn from the weighted summarization of the extracted multiple inaccurate targets. These two advances constitute the resulting technique WSMLF. Introducing the proposed WSMLF to several common deep image semantic segmentation approaches for the detection of H. pylori in WSIs, we observe that WSMLF can enable these approaches to achieve more reasonable segmentation results, which eventually improve the detection performance of H. pylori by at most 6%. WSMLF provides new thoughts for more effectively employing weak supervision to achieve more effective results for image semantic segmentation.

The phase relations in the Ce2Co17-Sm2Co17 system over the whole concentration range have been studied by means of Xray powder diffraction, differential thermal analysis and scanning electron microscopy equipped with energy dispersive Xray spectroscopy. The X-ray powder diffraction results reveal that all the alloys (Ce1-xSmx)2Co17 are similar to the end member of the investigated system, Sm2Co17. It is implied that continuous solid solutions are formed in this system. The lattice parameters and unit cell volumes of (Ce1-xSmx)2Co17 solid solutions increase linearly with x increasing from 0 to 1.0. The occurrence of the polymorphic transformation reaction α-(Ce, Sm)2Co17 = β-(Ce, Sm)2Co17 is confirmed in the Ce2Co17-Sm2Co17 system, but its transition temperature cannot be determined. The differential thermal analysis measurements show that both the decomposition temperature and the Curie temperature of the (Ce1-xSmx)2Co17 alloys increase gradually with increasing Sm content. Based on the X-ray powder diffraction results and differential thermal analysis data, the tentative vertical section of Ce2Co17-Sm2Co17 pseudobinary system has been constructed.

Multi-model ensemble is an important fundamental technique of practical value for many artificial intelligence applications. However, the usage for multi-model ensemble has been limited when it is combined with deep neural networks to construct ensemble of deep neural networks. Due to the big time and computing resources required to train and to integrate multiple deep neural networks for the achievement of multi-model ensemble, the engineering application field where developing time and computing resources are usually restricted, has not yet widespreadly benefited from ensemble of deep neural networks. To alleviate this situation, we present a new multi-model ensemble approach entitled feature transferring based multi-model ensemble (FTBME), for ensemble of deep neural networks. Primarily, we propose a feature transferring based multi-model training strategy to more affordably find multiple extra models based on a given previously optimized deep neural network model. Sequentially, to develop better ensemble solutions, we design a more effective random greedy based ensemble selection strategy to filter out models non-positive to ensemble generalization. Finally, inspired by the idea of averaging parameter points, we propose to fuse the obtained models in weight space which eventually reduces the expense of ensemble at the testing stage to a single deep neural network model while retaining the generalization. These three advances constitute the resulting technique FTBME. We conducted extensive experiments using deep neural networks, from light weight to complex, on ImageNet, CIFAR-10 and CIFAR-100. Results show that, given a deep neural network model which has been well-optimized and reaching its limit, FTBME can obtain better generalization with minor extra training expense while maintaining the expense to a single model at ensemble testing. This promising property of FTBME make us believe that it could be leveraged to broaden the usage for ensemble of deep neural networks, alleviating the situation that the engineering application field has not yet widespreadly benefited from ensemble of deep neural networks.

The object detection algorithm is a key component for the autonomous operation of unmanned surface vehicles (USVs). However, owing to complex marine conditions, it is difficult to obtain large-scale, fully labeled surface object datasets. Shipborne sensors are often susceptible to external interference and have unsatisfying performance, compromising the results of traditional object detection tasks. In this paper, a few-shot surface object detection method is proposed based on multimodal sensor systems for USVs. The multi-modal sensors were used for three-dimensional object detection, and the ability of USVs to detect moving objects was enhanced, realizing metric learning-based few-shot object detection for USVs. Compared with conventional methods, the proposed method enhanced the classification results of few-shot tasks. The proposed approach achieves relatively better performance in three sampled sets of well-known datasets, i.e., 2%, 10%, 5% on average precision (AP) and 28%, 24%, 24% on average orientation similarity (AOS). Therefore, this study can be potentially used for various applications where the number of labeled data is not enough to acquire a compromising result.

Non-high-density lipoprotein cholesterol (non-HDL-C) and low-density lipoprotein cholesterol (LDL-C) were established as the target for blood lipid management among patients with coronary artery disease (CAD). Previous study reported a negative relation between baseline LDL-C levels and long-term prognosis. However, the association between baseline non-HDL-C concentration and clinical outcomes is unknown.A total of 41,182 CAD patients admitted to Guangdong Provincial People's Hospital in China were included in this study from January 2007 to December 2018 and divided into two groups (non-HDL-C 2.2 mmol/L, n = 3236; non-HDL-C ≥ 2.2 mmol/L, n = 37,946). The Kaplan-Meier method, Cox regression analyses and restricted cubic splines were used to assess the association between non-HDL-C levels and long-term all-cause mortality.The overall mortality was 12.74% (n = 5247) over a median follow-up period of 5.20 years. Kaplan-Meier analysis showed that low non-HDL-C levels were paradoxically associated with a worse prognosis. After adjustment for baseline confounders (e.g., age, sex and comorbidities, etc.), multivariate Cox regression analysis revealed that low non-HDL-C levels (2.2 mmol/L) were not significantly associated with all-cause mortality (adjusted HR, 1.03; 95% CI, 0.93-1.14). After adjustment for nutritional status, the risk of all-cause mortality in patients with low non-HDL-C levels decreased (adjusted HR, 0.86; 95% CI, 0.78-0.95). In the final multivariate Cox model adjusting for full covariates, low non-HDL-C level was related to better prognosis (adjusted HR, 0.88; 95% CI, 0.80-0.98).This study found a paradoxical association between baseline non-HDL-C concentration and long-term all-cause mortality. Malnutrition mainly mediates to the non-HDL-C paradox. Elevated non-HDL-C concentration is still a risk factor of long-term all-cause mortality after considering nutritional status.

Aluminum ion-exchanged Ca-montmorillonite with different concentrations of aluminum ion were prepared and applied in the conversion of carbohydrates to 5-hydroxymethylfurfural (HMF). It was found that two types of acid sites were present in the catalyst, Lewis acid sites for isomerization of glucose and Bronsted acid sites for dehydration of fructose. With Al-mont-10 as catalyst, the yield of HMF converted from glucose was 80.4% in a biphase system (THF: NaCl solution) at 180 °C for 2.5 h. Further, the catalyst also shows excellent catalytic performance in the conversion of starch and inulin, and the yield of HMF reached 60.1% and 81.6% under identical conditions, respectively.

Background Pathological complete response (pCR) is considered a surrogate endpoint for favorable survival in breast cancer patients treated with neoadjuvant chemotherapy (NAC). Predictive biomarkers of treatment response are crucial for guiding treatment decisions. With the hypothesis that histological information on tumor biopsy images could predict NAC response in breast cancer, we proposed a novel deep learning (DL)-based biomarker that predicts pCR from images of hematoxylin and eosin (H&E)-stained tissue and evaluated its predictive performance. Methods In total, 540 breast cancer patients receiving standard NAC were enrolled. Based on H&E-stained images, DL methods were employed to automatically identify tumor epithelium and predict pCR by scoring the identified tumor epithelium to produce a histopathological biomarker, the pCR-score. The predictive performance of the pCR-score was assessed and compared with that of conventional biomarkers including stromal tumor-infiltrating lymphocytes (sTILs) and subtype. Results The pCR-score derived from H&E staining achieved an area under the curve (AUC) of 0.847 in predicting pCR directly, and achieved accuracy, F1 score, and AUC of 0.853, 0.503, and 0.822 processed by the logistic regression method, respectively, higher than either sTILs or subtype; a prediction model of pCR constructed by integrating sTILs, subtype and pCR-score yielded a mean AUC of 0.890, outperforming the baseline sTIL-subtype model by 0.051 (0.839, P = 0.001). Conclusion The DL-based pCR-score from histological images is predictive of pCR better than sTILs and subtype, and holds the great potentials for a more accurate stratification of patients for NAC.

Background Several studies have found that a low baseline low -density lipoprotein cholesterol (LDL-C) concentration was associated with poor prognosis in patients with acute coronary syndrome (ACS), which is called the “cholesterol paradox”. Low LDL-C concentration may reflect underlying malnutrition, which was strongly associated with increased mortality. The aim of this study was to investigate the cholesterol paradox in patients with CAD and the effects of malnutrition. Method A total of 41,229 CAD patients admitted to Guangdong Provincial People’s Hospital in China were included in this study from January 2007 to December 2018 and divided into two groups (LDL-C n = 4863; LDL-C ≥ 1.8 mmol/L, n = 36,366). The Kaplan-Meier method and Cox regression analyses were used to assess the association between LDL-C levels and long-term all-cause mortality and the effect of malnutrition. Result In this real-world cohort (mean age 62.9 years; 74.9% male), there were 5257 cases of all-cause death during a median follow-up of 5.20 years [interquartile range (IQR): 3.05–7.78 years]. Kaplan–Meier analysis showed that low LDL-C levels were associated with a worse prognosis. After adjusting for baseline confounders (e.g., age, sex and comorbidities, etc.), multivariate Cox regression analysis revealed that a low LDL-C level ( Conclusion This study demonstrated that the cholesterol paradox existed in CAD patients but disappeared after accounting for the effects of malnutrition.

The appearance of cloud computing has provided an efficient and economical way for users to receive computing resources with a pay-as-you-go model from cloud service provider. Cloud computing has attracted wide attention in both academia and industrial application. However, with the widely deploying of large-scale data center, high energy consumption becomes one of the key challenges and difficulties that restricting the development of cloud computing. Due to the limitations and dynamics of resources, how to map the cloud tasks to the resources aiming at decreasing the energy consumption has remained an essential issue. In this paper, we have made some improvements to the DeepJS and present DeepEnergyJS to minimize energy consumption for data center that receives enormous numbers of tasks arrive dynamically. We firstly introduce a power model to estimate the power of data center of the moment, then design a reward function that reflects the goal of optimizing power consumption and implement the simulation experiment on independent workloads to verity its availability of reducing the energy consumption. Finally, DeepJS considered only the independent workloads, we model inter-task dependencies in jobs and then conduct experiments in a similar way with independent workloads to prove that DeepEnergyJS can be extended for workloads that with dependencies which make it more functional. What’s more, our experiments are based on different configured servers while the severs used in DeepJS are all the same, which make our study closer to the real cloud computing environment. The experiments results show that DeepEnergyJS can significantly outperforms three baselines in both independent workloads and workloads with dependencies.

Objective: Previous studies have shown that the relationship between hypertension (HT) and contrast-associated acute kidney injury (CA-AKI) is not clear. We apply a systematic review and meta-analysis to assess the association between HT and CA-AKI. Methods: We searched for articles on the study of risk factors for CA-AKI in the Embase, Medline, and Cochrane Database of Systematic Reviews (by March 25, 2021). Two authors independently performed quality assessment and extracted data such as the studies’ clinical setting, the definition of CA-AKI, and the number of patients. The CA-AKI was defined as a serum creatinine (SCr) increase ≥25% or ≥0.5 mg/dL from baseline within 72 h. We used fixed or random models to pool adjusted OR (aOR) by STATA. Results: A total of 45 studies (2,830,338 patients) were identified, and the average incidence of CA-AKI was 6.48%. There was an increased risk of CA-AKI associated with HT (aOR: 1.378, 95% CI: 1.211–1.567, I2 = 67.9%). In CA-AKI with a SCr increase ≥50% or ≥0.3 mg/dL from baseline within 72 h, an increased risk of CA-AKI was associated with HT (aOR: 1.414, 95% CI: 1.152–1.736, I2 = 0%). In CA-AKI with a Scr increase ≥50% or ≥0.3 mg/dL from baseline within 7 days, HT increases the risk of CA-AKI (aOR: 1.317, 95% CI: 1.049–1.654, I2 = 51.5%). Conclusion: Our meta-analysis confirmed that HT is an independent risk factor for CA-AKI and can be used to identify risk stratification. Physicians should pay more attention toward prevention and treatment of patients with HT in clinical practice.

Background Contrast-associated acute kidney injury (CA-AKI) is a common complication after coronary angiography (CAG), which brings a poor prognosis. But up to now, there were fewer studies to discuss the incidence of CA-AKI comprehensively. We comprehensively explore the incidence of CA-AKI after coronary angiography. Methods We searched Medline, Embase, and Cochrane Database of Systematic Reviews (to 30th June 2019). We evaluated the world’s incidence of the CA-AKI, and associated mortality, and to described geographic variations according to countries, regions, and economies. CA-AKI was defined as an increase in serum creatinine ≥ 0.5 mg/dl or ≥ 25% within 72 h. Random effects model meta-analyses and meta-regressions was performed to derive the sources of heterogeneity. Results A total of 134 articles (1,211,106 participants) were included in our meta-analysis. Most studies originated from China, Japan, Turkey and United States, from upper middle income and high income countries. The pooled incidence of CA-AKI after coronary angiography was 12.8% (95% CI 11.7–13.9%), and the CA-AKI associated mortality was 20.2% (95% CI 10.7–29.7%). The incidence of CA-AKI and the CA-AKI associated mortality were not declined over time (Incidence rate change: 0.23% 95% CI − 0.050 to 0.510 p = 0.617; Mortality rate change: − 1.05% 95% CI − 3.070 to 0.970 p = 0.308, respectively). Conclusion CA-AKI was a universal complication in many regions, and the burden of CA-AKI remains severe. In clinical practice, physicians should pay more attention to the occurrence and active prevention and treatment of CA-AKI. Graphic abstract

Accurate estimation of food volume is critical in the medical field. However, estimating food volume is a challenging task due to the diverse nature of food, multi-scale and other characteristics. In this paper, we explore the relationship between the properties and volume of the object (food and reference) in the image. By combining Faster R-CNN, Grabcut, Median filtering, and CNN algorithm, we propose a framework for estimating food volume based on reference. The framework uses a front view which contains reference and food to estimate food volume and is applied to image datasets for 5 kinds of foods. The experimental results show the effective performance of this method for predicting volume, and the mean absolute error of each kind of food is less than 4.5%, which shows the model is robust to estimate volume for irregular food.

Additional file 1: Table S1. Baseline characteristics of non-PCI group and PCI group in total cohort. Table S2. Baseline characteristics of CA-AKI group and non CA-AKI group in developing cohort. Table S3. Multivariate Cox regression analysis of risk factors for long-term mortality. Table S4. Multivariate logistic regression of variates with statistically significant difference in univariate analysis.

In this paper, we novelly consider visual localization in active and passive two ways, with simple definition that active localization assists device to estimate location of its interest while passive localization aids device to estimate its own location in environment. Expecting to indicate some insights into visual localization, we specifically performed two explorations on active localization and more importantly explored to upgrade them from active to passive localization with extra geometry information available. In order to produce unconstrained and accurate 2D location estimation of interested object, we constructed an active localization system by fusing detection, tracking and recognition. Based on recognition, we proposed a collaborative strategy making mutual enhancement between detection and tracking possible to obtain better performance on 2D location estimation. Meanwhile, to actively estimate semantic location of interested visual region, we employed latest state-of-the-art light weight CNN models specifically designed for efficiency and trained two of them with large place dataset in perspective of scene recognition. What’s more, using depth information available from RGB-D camera, we improved the active system for 2D location of interested object to a passive system for relative 3D location of device to the interested object. Firstly estimated was the 3D location of the interested object in the coordinate system of device, then relative location of device to the interested object in world coordinate system was deduced with appropriate assumption. Evaluations both subjectively on a RGB-D sequence obtained in a lab environment and practically on a robotic platform in an office environment indicated that the improved system was suitable for autonomous following robot. As well, the active system for rough semantic location estimation of interested visual region was promoted to a passive system for fine location estimation of device, with available 3D map describing the visited environment. In perspective of place recognition, we first adopted one of the efficient CNN models previously trained for semantic location estimation as a base to generate CNN features for both retrieval of candidate loops in the map and geometrical consistency checking of retrieved loops, then true loops were used to deduce fine location of device itself in environment. Comparison with state-of-the-art results reflected that the promoted system was adequate for long-term robotic autonomy. Achieving favorable performances, the presented four explorations have implied adequacy for elaborating on some insights into visual localization.

Adaptively learning the difference between object and background, discriminative trackers are able to overcome the complex background problem in visual object tracking. However, they are not robust enough to handle the out-of-plane rotation of object, which reduces recall performance. Meanwhile, allowing individual parts certain criterion of freedom, part-based trackers can better handle the out-of-plane rotation problem. However, they are prone to be affected by complex background, leading to low precision performance. To simultaneously address both issues, we propose a collaborative strategy that makes mutual enhancement between a discriminative tracker and a part-based tracker possible to obtain better overall performance. On one hand, we use validated results from the part-based tracker to update the discriminative tracker for recall performance improvement. On the other hand, based on confident results from the discriminative tracker we adaptively update the part-based tracker for simultaneous precision performance improvement. Experiments on various challenge sequences show that our approach achieved the state-of-the-art performance, which demonstrated the effectiveness of mutual collaboration between the two trackers.

In order to correlate the physico-chemical properties of Ce x Zr 1-x O 2 catalysts with their catalytic activities for direct carboxylation of methanol to dimethyl carbonate (DMC), ceria-zirconia mixed oxides with the same nominal composition (Ce 0.5 Zr 0.5 O 2 ), which were prepared by different methods, were applied. The catalysts were characterized by means of XRD, N 2 adsorption, ICP-MS, NH 3 -TPD, CO 2 -TPD, Raman spectroscopy and XPS. Significant variations in their physico-chemical properties were observed by varying the preparation conditions. The results showed that the catalytic activity of the ceria-zirconia mixed oxides increased with improving the proportion of Ce 4+ , acidity and basicity of the catalyst surface. And the number of surface Ce 4+ content is the most important factor.

In view of the massive data transmission and storage of various types of ocean data, and the limitations of IPv4 transmission protocols on marine projects, we propose a digital ocean cloud platform architecture based on IPv6 intelligent gateway, which utilizes an unlimited number of address resource pools of IPv6. The characteristics of efficient and secure data dissemination combined with cloud computing service technology, based on distributed storage technology, provide a safe and reliable storage environment for marine sensor data, ocean images and text, and ocean data through cloud computing technology and ocean data can be aggregated, classified and processed and through cloud computing technology can deepen the value of ocean data.

Named entity recognition is a basic problem in natural language processing, and also an indispensable part of many natural language processing technologies such as information extraction and information retrieval. There are a large number of proprietary entities in the marine field. In this paper, we propose a method based on user dictionary and conditional random field, which recognize energy, organisms and drugs named entities in the marine filed. Analyzing the characteristics of marine named entities and successively adding different features to raise the recognition rate. Ultimately, the accuracy rate reaches over 96.95% and the recalling rate reaches over 96.41%. In addition, the method of combining neural network with statistical model (BiLSTM-CRF) is also used in this paper. In order to reduce the influence of inaccurate word segmentation on named entity recognition in the marine field, we use character vectors instead of word vectors as input of neural network layer. Finally, the accuracy rate reaches over 72.23% and the recalling rate reaches over 66.76%. The experimental results confirm that the method based on user dictionary and conditional random field is feasible for named entity recognition in the marine field.

Digital Ocean Construction has become a national strategy, the development of which is restricted by traditional IPv4 network. In this paper, IPv6 transition technology is researched to solve the current digital ocean network problems, and an IPv6 transition mechanism is proposed. The work of collecting data with IPv6 sensor network in the digital ocean by deploying border gateway devices which support 6LoWPAN and IPv6 network has been done. The border gateway devices can communicate with digital ocean IPv6 datacenters over an IPv4 network by using 6rd tunnel technology. With dual stack equipment, the digital ocean IPv6 datacenters can be accessed by both IPv4 and IPv6 networks. This transition mechanism satisfies the network demands for research and development of digital ocean.

Ensembles of deep convolutional neural networks (CNNs), which integrate multiple deep CNN models to achieve better generalization for an artificial intelligence application, now play an important role in ensemble learning due to the dominant position of deep learning. However, the usage of ensembles of deep CNNs is still not adequate because the increasing complexity of deep CNN architectures and the emerging data with large dimensionality have made the training stage and testing stage of ensembles of deep CNNs inevitably expensive. To alleviate this situation, we propose a new approach that finds multiple models converging to local minima in subparameter space for ensembles of deep CNNs. The subparameter space here refers to the space constructed by a partial selection of parameters, instead of the entire set of parameters, of a deep CNN architecture. We show that local minima found in the subparameter space of a deep CNN architecture can in fact be effective for ensembles of deep CNNs to achieve better generalization. Moreover, finding local minima in the subparameter space of a deep CNN architecture is more affordable at the training stage, and the multiple models at the found local minima can also be selectively fused to achieve better ensemble generalization while limiting the expense to a single deep CNN model at the testing stage. Demonstrations of MobilenetV2, Resnet50 and InceptionV4 (deep CNN architectures from lightweight to complex) on ImageNet, CIFAR-10 and CIFAR-100, respectively, lead us to believe that finding local minima in the subparameter space of a deep CNN architecture could be leveraged to broaden the usage of ensembles of deep CNNs.

Diverse products were obtained from transfer hydrogenation of methyl levulinate (ML) by adjusting different factors under solvothermal conditions. Excellent yields of gamma-valerolactone (GVL, 75%), 1,4-pentanediol (1,4-PDO, 39%), and 1-pentanol (1-PAO, 13.8%) could be obtained using zirconium-supported copper catalyst calcined at 550 °C and under optimal conditions.

We reported herein a strategy for combining CUS-based MOF (CUS = coordinatively unsaturated metal sites) with ionic liquid (IL) functional sites to form bifunctional heterogeneous catalysts with extra high activity for CO2 fixation. Based on this strategy, two quaternary ammonium salt and quaternary phosphorus salt ionic liquid functionalized CUS-containing MOF heterogeneous catalysts, MIL-101-N(n-Bu)3Br and MIL-101-P(n-Bu)3Br, have been prepared for the first time by a post-synthesis modification method. Due to the synergetic role of dual functional sites including Lewis acid sites in the MOF framwork and Br− ions in the IL functional sites, MIL-101-N(n-Bu)3Br and MIL-101-P(n-Bu)3Br exhibit high catalytic activity for the cycloaddition of CO2 and epoxide under mild and co-catalyst free conditions, which significantly outperforms other benchmark MOF catalysts. Moreover, such bifunctional catalysts can be easily recovered and recycled several times without leaching and loss of activity. Our work thus paves a way for the development of IL functionalized MOFs as heterogeneous catalysts for CO2 fixation.

Research on automatic text categorization of the oceanographic big data, has always been the core work of establishing a oceanographic information database based on oceanographic big data information platform. However, in face of extreme complex Internet data and the professional needs of the oceanographic field, the categorization accuracy of the traditional tf-idf algorithm is difficult to meet the demands. Based on traditional tf-idf algorithm and mutual information algorithm, I propose improved tf-idf-miow algorithm in order to meet the demands of text categorization of the oceanographic big data. Optimize the mutual information algorithm, calculate the correlation coefficient between the characteristic word and the oceanographic field. In this way, I set the weight of oceanographic field: miow by this correlation coefficient, and bring miow into traditional tf-idf algorithm. The results of automatic text categorization experiments show that the recall rate of tf-idf-miow in oceanographic field is 10.33% higher than of traditional tf-idf algorithm, and the f1-score is improved by 6.92%.

High dropout rate of MOOC is criticized while a dramatically increasing number of learners are appealed to these online learning platforms. Various works have been done on analysis and prediction of dropout. Machine learning techniques are widely applied to this field. However, a single classifier may not always perform reliable for predictions. In this work, we study dropout prediction for MOOC. A technique is proposed to predict dropouts using learning activity information of learners. We applied a two-layer cascading classifier with a combination of three different machine learning classifiers — Random Forest (RF), Support Vector Machine (SVM), and MultiNomial Logistic Regression (MLR) for prediction. Experimental results indicate that the technique is promising in predicting dropouts with achieving 97% precision.

In this work, we studied time management behavior, performance of students, and their association in online learning. We propose three novel time-related features, i.e., video-watching frequency, video-watching interval and learning efficiency. Leveraging these features, we train classifiers for identifying study habits of students and assessing the impact of behavior on performance. Experimental results on dataset from our online learning platform demonstrate that the proposed features are effective for learning behavior description. Comprehensive experiments show that our method can get promising performance.

The purpose of this study is to investigate an exergy analysis on embedded pipes of radiant ceiling cooling system. Based on the thermodynamic principles, the exergy analysis model of radiant ceiling cooling system is established. The exergy efficiency of radiant ceiling cooling system and its main equipments can be simulated and calculated by the model. Under the different pipe spacing, depth and diameter, the exergy efficiency of radiant ceiling, refrigerating unit, pump and radiant ceiling cooling system and the surface temperature of radiant ceiling panel are compared and analyzed and the exergy loss coefficients of radiant ceiling panel, refrigerating unit and pump are calculated. The conclusions show that pipe depth and spacing have primary influence on exergy efficiency and surface temperature distribution respectively and the refrigerating unit is the most important part for exergy saving in radiant ceiling cooling system.

Based on the understanding of the mechanism of energy dissipation, a new type of plunge pool is presented with the advantages of high rate of energy dissipation, low impact pressure, and small close-to-bed velocity on the soleplate of the plunge pool. The first advantage owes to enlarged shearing energy dissipation areas in the plunge pool, while the other two are caused by the jet that enters in a nearly horizontal level to keep the soleplate from being scoured directly. All of the above arrangements make this new type of energy dissipator distinct from the underflow energy dissipation. Through experiments on the physical model, the authors found that the water flow maintained stable and submerged in horizontal direction when the flow was narrow in horizontal but thick in vertical direction. However, the wide flat nappe was ready to submerge or float as the downstream water level rised or dropped. The entire flow fields of multi-horizontal submerged jets into plunge pool were also numerically simulated. The numerical results of water surface curve, close-to-bed velocity and floor pressure agree well with the experimental data. The flow pattern in the plunge pool was analyzed after combining the laboratory data and numerical simulation.

This paper solves the problem of the excessive memory footprint and the slow speed of parsing when using the DOM method to parse the XML data. As we know the current XML parsing technology is the key of XML research field and now there are a lot of XML parsers, their works plays a significant role in promoting XML parsing, but failed to achieve a good balance in the analysis of both time and memory usage. We propose a hash function based on the XML parser, for evaluating the memory consumption. We have demonstrated the text part of delay spread in terms of the time consumption, using a hash function. Our testing results show that, compared with the traditional method, the memory footprint reduced by 5% and the analytical efficiency increased by 3%.

Due to the characteristics of the current open-source Android system, and looser user rights management mechanism, information security issues become a problem of common concern to the user. At the present stage, there are many measures encrypting data for Android database, which are divided into several aspects of hardware layer, core layer, virtual machine layer and application framework layer. This paper presents a method which achieves the database encryption by modifying SQLite source code in Librariesi¾źandi¾źAndroidi¾źRuntime Layer. The encryption function can not only encrypt the data itself, but also be invoked by the upper application though JNI interface, so that the application can access more secure database, and improve the security of the instant messaging application. Experimental results show feasibility and effectiveness of the SQLite database encryption which is used in the instant messaging application.

With the enrichment of Marine information on the Internet, the users find that it is hard to find the knowledge they need when they face with a deluge of Marine information resources, they wish to acquire marine information data rapidly which is integrated and desirable. Recently, researchers have been committed to solve these problems with the Web Clustering Mining, according to the user's access action, looking for patterns of behavior or retrieval of users with similar interests. Aiming at the deficiency in the current problem of calculating similarity between users based on user sessions. This paper proposes an algorithm to use associated queries about query logs as the compensation characteristics to measure the similarity between users, combined with user-browsing path algorithm to clustering the user data. The experimental results of the algorithm have a good reference value for users of personalized service and promote the development of marine information. KeywordsMarine information; Web Clustering Mining; users query log; associated queries; user-browsing path; (key words)

The 3-D flow fields in the plunge pools of'the Xiaowan high arch dam and Laxiwa high arch dam in China are simulated by the turbulence mathematical model and measured by the live-hole Pitot sphere combined with the pressure sensors and automatic data sampling and processing system. The typical 3-D flow pattern and energy dissipation characteristics in plunge pool are obtained. The simulated results of energy dissipation show that the water body in plunge pool can be divided into three regions: shear, impact and mixing dissipation regions of energy.

Various food safety and health events have occurred over the recent years. Therefore, how to effectively guarantee food safety is increasingly becoming an urgent issue to study and resolve. The establishment of food safety trace ability system, which takes effective tracking and tracing for food is an important means to effective protection of food safety. In this paper, we propose a vegetable safety trace ability model integrated with two-dimensional barcode technology and Web Service technology by analysis of the vegetable supply chain. The model takes two-dimensional barcode as information carrier which identifies entity object closely related to vegetable cultivation, processing, transportation, storage, marketing, and adopts Web Service Technology for processing and exchanging information. Consumers first install two-dimensional barcode reading software on intelligent terminals, and then scan the two-dimensional barcode on the vegetable product packaging through the terminal camera. The reading software sends request to the Web Service interface provided by the trace ability system, and returns vegetable trace ability information to the terminal after program parses.

This paper introduces a real-time interactive application development model on the Android platform. This model uses persistent socket connection mechanism to establish the connection of clients and servers and uses Protocol Buffers as the transmission medium. In order to maintain the smooth connection with the server in the communication process, clients send keep alive packages to servers at regular time. In addition, this model adds retransmission mechanism and reconnection mechanism to ensure the correctness of the packages and the robustness of the system. This module has a good performance in network data consumption, power saving and the correctness of the packages etc.

With the rapid spread of the cell phones, more and more security threats against it appear. Nowadays people pay more attention to mobile phone security. Snort is a lightweight network intrusion detection system, and the multi-pattern matching algorithms it usually uses are AC algorithm and AC-BNFA algorithm. However, these algorithms are slow and take up too much memory. Now a new pattern matching algorithm is presented, which reduces the automaton states through transitions. The detection rate is ensured, the match rate is accelerated, the memory footprint is reduced, and the detection efficiency of the intrusion detection system is improved.

For the development and operating efficiency of Web applications based on the Model-View-Controller (MVC) framework, and, according to the actual business environment and needs in the project practice, the framework of Web application system is studied in this paper. Through the research of Spring MVC framework and Mybatis framework as well as some related core techniques, combined with JSP and JSTL technology, this paper realizes the design of a lightweight Web application framework based on Spring MVC and Mybatis.