Search

Although deep neural networks (DNNs) achieve rather high performance in many cutting-edge applications (e.g., autonomous driving, medical diagnose), their trustworthiness on real-world scenarios still posts concerns, where some specific failure examples are often encountered during the real-world operational environment. With the limited failure examples collected during the practical operation, how to effectively leverage such failure cases to repair and enhance DNN so as to generalize to more potentially suspicious samples is challenging, but of great importance. In this paper, we formulate the failure-data-driven DNN repairing as a data augmentation problem, and design a novel augmentation-based repairing method, which to the best extent leverages limited failure cases. To realize the DNN repairing effects that generalize to specific failure examples, we originally propose few-shot guided mix (FSGMix) that augments training data with the guidance of failure examples. As a result, our method is able to achieve high generalization to the collected failure examples and other similar suspicious data. The preliminary evaluation on CIFAR-10 dataset demonstrates the potential of our proposed technique, which automatically learns to resolve the potential failure patterns in the DNN operational environment.

As deep learning (DL) opens the way to many technological innovations in a wild range of fields, more and more researchers and developers from diverse domains start to take advantage of DLs. In many circumstances, a developer leverages a DL framework and programs the training software in the form of source code (e.g., Python, Java). However, not all of the developers across domains are skilled at programming. It is highly desirable to provide a way so that a developer could focus on how to design and optimize their DL systems instead of spending too much time on programming. To simplify the programming process towards saving time and effort especially for beginners, we propose and implement DeepVisual, a visual programming tool for the design and development of DL systems. DeepVisual represents each layer of a neural network as a component. A user can drag-and-drop components to design and build a DL model, after which the training code is automatically generated. Moreover, DeepVisual supports to extract the neural network architecture on the given source code as input. We implement DeepVisual as a PyCharm plugin and demonstrate its usefulness on two typical use cases.

This paper describes the necessity of developing ecological monitoring in the Baisha Lake of Nanji Wetland National Natural Reserve. Considering the geographical and environmental factors, the urgent need of ecological monitoring is discussed in the study area. According to the framework of wireless sensor network, a scheme to solve some of the problems regarding the system installation is designed. And the scheme also describes the details in selection of monitoring locations, technical methods and the main contents of monitoring.

IBC is emerging as an important contender to traditional public key cryptography for its attractive properties. But the Wireless Transport Layer Security (WTLS) protocol does not support it at present. To resolve this problem, an improved WTLS protocol based on IBC (IB_WTLS) is proposed. Firstly, new cipher suits are designed for WTLS protocol, which give support to IBC scheme. Secondly, the IBC-based authentication is implemented by using IBC public-keys instead of X.509 public-keys. At last, IB_WTLS protocol message architecture is improved by defining a set of entity interaction operation sequences. The results show that IB_WTLS have advantages of lightweight and efficiency of IBC, as well as easy to deploy.

To investigate B cell epitopes on the VP3 of infectious bursal disease virus (IBDV), the VP3 gene of IBDV were subcloned into the pET28a vector to construct expression vector of pETVP3. The recombinant VP3 (rVP3) was expressed in E. coli BL21(DE3) by inducing with IPTG. The expressed VP3 protein was recognized specifically by anti-IBDV antisera in western blot assays. According to antigenic structure of IBDV VP3, overlapping peptides which cover the full sequence of VP3 were synthesized and Conjugated to carrier Protein BSA to prepare conjugated antigens. Their reactivities with anti-IBDV monoclonal antibodies (MAb) or polyclonal antibody were detected by peptide-ELISA and Dot-ELISA The results showed that there were 2 linear epitopes recognized by neutralizing Mabs on the VP3, 728PRDWDRLPYLNL739 and 982PKPKPKPNAPTQ993, which can also react with IBDV antiserum. Dot-ELISA confirmed that there were still 4 linear epitopes recognized by polyclonal antibody on the VP3, 818LANAPQAGSKSQRA831, 851QREKDTRISKK METMGIYFATP872, 876ALNGHRGPSPGQLKYWQN TREI897 and 961QMKDLLLTAMEMK973. It is helpful for the developmentof IBD epitope vaccine.

Aiming to increase the proportion of the samples that has been determinate classified in Naive Credal Classifier, this paper improves conservative inference rule and proposes an incomplete data classification model based on relaxed conservative inference rule. Simulation results of comparative experiment with Naive Bayesian Classifier and Naive Credal Classifier verify the effectiveness of this classification model.

Twelve species of marine organisms were collected from the Xing Hua Bay（XHB）, South China and were analyzed for 19 OCPs, including hexachlorocyclohexane isomers(HCHs), and DDT and its metabolites(DDTs), by GC-ECD. Date on mass fraction of persistent toxic substances in the marine organisms, were investigated and assembled . In XHB, DDTs mass fraction was greater than other toxic substances, the ratio of DDTs/OCPs were in the range of 57% ~ 91%. The ratios of (p,p′-DDD+p,p′-DDE)/ p,p′-DDT were in the range of 0.34~4.56, with an average ratio of 1.75, indicating that most of the DDTs were transformed to DDD and DDE, and DDTs in the marine organisms were mainly from the use of OCPs in previous periods.

Xinfeng, is the typical region of the vegetation restoration of serious soil and water loss zone in Jiangxi Province, in which dynamics of species diversity and achievement of 6 restoration communities were analyzed using three diversity indices, Detrended correspondence analysis (DCA) and Detrended canonical correspondence analysis (DCCA). Investigations were carried out in 2004yr and 2008yr. The results were as follows: (1) In comparison with 2004yr, flora and dominant families (with much genus and species) of this region increased in 2008yr, especially for Compositae. (2) Total richness and all diversity indices of shrub layer in 6 communities in 2008yr were higher than those in 2004yr. All diversity indices of tree and herb layer in communities IV, V and VI in 2008yr were higher than those in 2004yr. All diversity indices of tree, shrub and herb layer in types of closing hill for afforestation (II, IV and V) and types of mixed tree species (II, V and VI) were higher than those in pure forest without afforestation conservation (I and III), and type of closing hill for afforestation-mixed tree species (V) and type of mixed tree species (VI) showed highest species diversity. (3) The second DCA axis showed gradients of restoration modes and species construction. The first DCCA axis showed gradients of soil status, communities of restoration vegetation; the second DCCA axis showed gradients of surface runoff and community coverage. To summarize, variation in species richness and species diversity is related to without afforestation conservation and mono-tree species, and the need for conservation in the process of vegetation restoration is emphasized. High diverse mixed species in restoration and closing hill for afforestation (avoiding disturbance) in this region can maintain high species richness and diversity.

To improve the actual ability of the sonar in the war and solve the control tactic of the ocean, we have put our attention in the sonar emergency system. To overcome the complicated environment of the ocean and the invalidation of the sonar system, we bring forward the sonar emergency system. Under the help of the Genetic Algorithm, a usual method was advanced. The global search and adaptability of the Genetic Algorithm can meet the problem. The computer simulation result has proved the method can meet the actual need.

This paper is a new attempt to introduce a “Multi-resolution and Multi-cue” Particle Filter tracking method. It combines multiple features distribution and multiple resolutions to facilitate 2D video tracking from a monocular view. The method’s benefits lie in its speed, its robustness. Speed is improved by multiple resolutions which reduce number of particles dramatically while keeping its quality; robustness is implemented by multiple cues. The proposed method is demonstrated on man tracking with a non-stationary camera in unconstrained outdoor environments. Result shows it is speedier and robust again background noise such as cluttered background, occlusion, color distracters and illumination change than conventional PF and PF with multi-cue.

in order to improve the accuracy and efficiency of directly use of the K-Means clustering for semi-supervised learning, the paper proposes a new semi-supervised learning algorithm based on Radius-Distance. In the algorithm, according to Radius, farthest distance of sample to the cluster center of unlabelled samples using K-Means, and Distance, from cluster center of unlabelled samples to center of labeled samples, a small amount of unlabeled data are selected to aid training learning. Experimental results on the Kddcup'99 demonstrate that the advantages of proposed algorithm over the K-Means method and S3VM.

A new fabric defects segmentation approach based on texture primitive is put forward in the paper, which consists of four steps: 1) calculating texture primitive template using auto-correlation function; 2) enhancing defect image through calculating difference between each texture primitive and primitive template; 3) constructing mean gradation image to attenuate the high frequent background information; 4) segmenting defect images according to threshold values acquired automatically through Otsu's approach, etc. Validity and robustness of the approach were proved by different fabric defect images segmentation experiment.

Gas/liquid two-phase flow system is a complex, nonlinear and dynamic system, therefore is difficult to measure. Basing on the application of Electrical Resistance Tomography (ERT) in gas/liquid two-phase flow of vertical pipe, wavelets transform is used to analyses the measured data from ERT system. Furthermore, according to the Multi-resolution Analysis (MRA), the feature vector of multiple scales wavelet energy that can express the essential information of gas/liquid two-phase flow is constructed. At last, the Support Vector Machine (SVM) method in statistics theories is adopted to validate the feasibility of the feature extraction method, and the higher recognizing rate of flow regime is obtained.

Flow regime is one of the fundamental parameters with important engineering significance in two-phase flow systems, and has always been an important aspect of two-phase flow research. In this paper, the support vector machine (SVM) method in statistics theories was adopted to analyze the measured data of electrical resistance tomography (ERT) system so that the flow regime identification of gas/liquid two-phase in horizontal pipe was carried out. The analysis result of experiment data indicates that this method can obtain an inspiring result.