Improving VQA via Dual-Level Feature Embedding Network

Visual Question Answering(VQA)has sparked widespread interest as a crucial task in integrating vision and language.VQA primarily uses attention mechanisms to effectively answer questions to associate relevant visual regions with input questions.The detection-based features extracted by the object detection network aim to acquire the visual attention distribution on a predetermined detection frame and provide object-level insights to answer questions about foreground objects more effectively.However,it cannot answer the question about the background forms without detection boxes due to the lack of fine-grained details,which is the advantage of grid-based features.In this paper,we propose a Dual-Level Feature Embedding(DLFE)network,which effectively integrates grid-based and detection-based image features in a unified architecture to realize the complementary advantages of both features.Specifically,in DLFE,In DLFE,firstly,a novel Dual-Level Self-Attention(DLSA)modular is proposed to mine the intrinsic properties of the two features,where Positional Relation Attention(PRA)is designed to model the position information.Then,we propose a Feature Fusion Attention(FFA)to address the semantic noise caused by the fusion of two features and construct an alignment graph to enhance and align the grid and detection features.Finally,we use co-attention to learn the interactive features of the image and question and answer questions more accurately.Our method has significantly improved compared to the baseline,increasing accuracy from 66.01%to 70.63%on the test-std dataset of VQA 1.0 and from 66.24%to 70.91%for the test-std dataset of VQA 2.0.

Compared Insights on Machine-Learning Anomaly Detection for Process Control Feature

Anomaly detection is becoming increasingly significant in industrial cyber security,and different machine-learning algorithms have been generally acknowledged as various effective intrusion detection engines to successfully identify cyber attacks.However,different machine-learning algorithms may exhibit their own detection effects even if they analyze the same feature samples.As a sequence,after developing one feature generation approach,the most effective and applicable detection engines should be desperately selected by comparing distinct properties of each machine-learning algorithm.Based on process control features generated by directed function transition diagrams,this paper introduces five different machine-learning algorithms as alternative detection engines to discuss their matching abilities.Furthermore,this paper not only describes some qualitative properties to compare their advantages and disadvantages,but also gives an in-depth and meticulous research on their detection accuracies and consuming time.In the verified experiments,two attack models and four different attack intensities are defined to facilitate all quantitative comparisons,and the impacts of detection accuracy caused by the feature parameter are also comparatively analyzed.All experimental results can clearly explain that SVM(Support Vector Machine)and WNN(Wavelet Neural Network)are suggested as two applicable detection engines under differing cases.

Experimental Study of Monitoring and Controlling of Composite Cure Process in Autoclave Featured with Fiber Optic Sensor

With the aid of the latest fiber optic sensing technology parameters in the cure process of ther- mosetting resin-matrix composite, such as temperature, viscosity,void and residual stress, can be monitored entirely and efficiently.In this paper, experiment results of viscosity measurement in composite cure process in autoclave using fiber optic sensors are presented. Based on the sensed information, a computer program is utilized to control the cure process. With this technology, the cure process becomes more apparent and controllable, which will greatly improve the cured products and reduce the cost.

Improved preprocessed Yaroslavsky filter based on shearlet features

An improved preprocessed Yaroslavsky filter（IPYF）is proposed to avoid the nick effects and obtain a better denoising result when the noise variance is unknown.Different from its predecessors,the similarity between two pixels is calculated by shearlet features.The feature vector consists of initial denoised results by the non-subsampled shearlet transform hard thresholding（NSST-HT）and NSST coefficients,which can help allocate the averaging weights more reasonably.With the correct estimated noise variance,the NSST-HT can provide good denoised results as the initial estimation and high-frequency coefficients contribute large weights to preserve textures.In case of the incorrect estimated noise variance,the low-frequency coefficients will mitigate the nick effect in cartoon regions greatly,making the IPYF more robust than the original PYF.Detailed experimental results show that the IPYF is a very competitive method based on a comprehensive consideration involving peak signal to noise ratio（PSNR）,computing time,visual quality and method noise.

Comparison of River Terraces in the Middle Reach Valleys of the Yellow River and Analysis on the Multi-Gradational Features of Tectonism in the Formation of Terrace Series

Where the Yellow River flows through the Haiyuan-Tongxin arc-form tectonic region on the northeastern side of the Qinghai-Xizang (Tibet) Plateau, as many as 10～21 basis and erosion terraces have been produced, among which the biggest altitude above river level is 401m and the formation age of the highest terrace is 1.57 Ma B.P. Based on comparative analysis of the Yellow River terraces located separately in the Mijiashan mountain, the Chemuxia gorge, the Heishanxia gorge and the other river terraces in the vast extent of the northern part of China, it has been found that the tectonic processes resulting in the formation of the terrace series is one of multi-gradational features, i.e., a terrace series can include the various terraces produced by tectonic uplifts of different scopes or scales and different ranks. The Yellow River terrace series in the study region can be divided into three grades. Among them, in the first grade there are 6 terraces which were formed separately at the same time in the vast extent of the northern part of China and represent the number and magnitude of uplift of the Qinghai-Xizang Plateau since 1.6 Ma B. P.; in the second grade there are 5 terraces which were separately and simultaneously developed within the Haiyuan-Tianjingshan tectonic region and represent the number and magnitude of uplift of this tectonic region itself since 1.6Ma B.P.; in the third grade there are 10 terraces which developed on the eastern slope of the Mijiashan mountain and represent the number and amplitude of uplift of the Haiyuan tectonic belt itself since 1.6Ma B.P. Comparison of the terrace ages with loess-paleosoil sequence has also showed that the first grade terraces reflecting the vast scope uplifts of the Qinghai-Xizang Plateau are very comparable with climatic changes and their formation ages all correspond to the interglacial epochs during which paleosoils were formed. This implies that the vast extent tectonic uplifts resulting in river down-cutting are closely related to the warm-humid climatic periods which can also result in river downward erosion after strong dry and cold climatic periods, and they have jointly formed the tectonic-climatic cycles. There exists no unanimous and specific relationship between the formation ages of the second and third grade terraces and climatic changes and it is shown that the formation of those terraces was most mainly controlled by tectonic uplifts of the Tianjingshan block and the Haiyuan belt. The river terraces in the study region, therefore, may belong to 2 kinds of formation cause. One is a tectonic-climatic cyclical terrace produced jointly by vast extent tectonic uplifts and climatic changes, and the terraces of this kind are extensively distributed and can be well compared with each other among regions. Another is a pulse-tectonic cyclical terrace produced by local tectonic uplifts as dominant elements, and their distribution is restricted within an active belt and can not be compared with among regions.

Improved Blending Attention Mechanism in Visual Question Answering

Visual question answering(VQA)has attracted more and more attention in computer vision and natural language processing.Scholars are committed to studying how to better integrate image features and text features to achieve better results in VQA tasks.Analysis of all features may cause information redundancy and heavy computational burden.Attention mechanism is a wise way to solve this problem.However,using single attention mechanism may cause incomplete concern of features.This paper improves the attention mechanism method and proposes a hybrid attention mechanism that combines the spatial attention mechanism method and the channel attention mechanism method.In the case that the attention mechanism will cause the loss of the original features,a small portion of image features were added as compensation.For the attention mechanism of text features,a selfattention mechanism was introduced,and the internal structural features of sentences were strengthened to improve the overall model.The results show that attention mechanism and feature compensation add 6.1%accuracy to multimodal low-rank bilinear pooling network.

Feature Setup Determination in Integrated CAD/CAM System For Concurrent Engineering

This paper presents a feature-based method for machining process planning in integrated product designing and manufacturing system for CE(Concurrent Engineering) application. The feature setup generation and machining sequence can be determined automatically in this system. The set of knowledge-based rules for process planning and manufacturability evaluation is provided and can be shared by all stages of full product life-cycle. An approach for MTAD (Multiple Tool Axis Direction) feature setup generation is presented and the appropriate Tool Axis Direction(TAD) is chosen to minimize the total setup numbers of a part. The classification and process planning of interacting feature are discussed and the knowledge-based rules are used to solve the feature interaction problem.

Safety assessment methodology for a German high-level waste repository in clay formations

In the ANSICHT project that was jointly carried out by DBE TECHNOLOGY Gmb H,BGR,and GRS g Gmb H,two generic geological site models were used to develop a first draft of a methodology to demonstrate the safety of a high-level waste(HLW) repository in argillaceous formations in Germany,taking into account the regulatory requirements.The main results of the project are characterised by the developed repository concepts adapted to the geological conditions.The specific quantifications of the integrity criteria and their exemplary application with calculational proofs were used to demonstrate the integrity of the host rocks.The development of site-specific FEP(features,events,and processes) cataloges provided a complete system description for evaluation of the repository evolution.The developed work flow of the demonstration concept illustrated the complete sequence of the safety proof in a transparent way.It shows that various steps have to be performed,possibly iteratively,to provide a successful safety proof.The results form a useful tool in the pending search for a HLW repository site,especially when providing a basis for comparing safety analyses of different sites in Germany.

A Multiple Feature Approach for Disorder Normalization in Clinical Notes

In this paper we propose a multiple feature approach for the normalization task which can map each disorder mention in the text to a unique unified medical language system（UMLS）concept unique identifier（CUI）. We develop a two-step method to acquire a list of candidate CUIs and their associated preferred names using UMLS API and to choose the closest CUI by calculating the similarity between the input disorder mention and each candidate. The similarity calculation step is formulated as a classification problem and multiple features（string features,ranking features,similarity features,and contextual features） are used to normalize the disorder mentions. The results show that the multiple feature approach improves the accuracy of the normalization task from 32.99% to 67.08% compared with the Meta Map baseline.

Intelligent learning model-based skill learning and strategy optimization in robot grinding and polishing

With the rapid advancement of manufacturing in China,robot machining technology has become a popular research subject.An increasing number of robots are currently being used to perform complex tasks during manual operation,e.g.,the grinding of large components using multi-robot systems and robot teleoperation in dangerous environments,and machining conditions have evolved from a single open mode to a multisystem closed mode.Because the environment is constantly changing with multiple systems interacting with each other,traditional methods,such as mechanism modeling and programming are no longer applicable.Intelligent learning models,such as deep learning,transfer learning,reinforcement learning,and imitation learning,have been widely used;thus,skill learning and strategy optimization have become the focus of research on robot machining.Skill learning in robot machining can use robotic flexibility to learn skills under unknown working conditions,and machining strategy research can optimize processing quality under complex working conditions.Additionally,skill learning and strategy optimization combined with an intelligent learning model demonstrate excellent performance for data characteristics learning,multisystem transformation,and environment perception,thus compensating for the shortcomings of the traditional research field.This paper summarizes the state-of-the-art in skill learning and strategy optimization research from the perspectives of feature processing,skill learning,strategy,and model optimization of robot grinding and polishing,in which deep learning,transfer learning,reinforcement learning,and imitation learning models are integrated into skill learning and strategy optimization during robot grinding and polishing.Finally,this paper describes future development trends in skill learning and strategy optimization based on an intelligent learning model in the system knowledge transfer and nonstructural environment autonomous processing.

Visual Person Identification Using a Distance-dependent Appearance Model for a Person Following Robot

This paper describes a person identifcation method for a mobile robot which performs specifc person following under dynamic complicated environments like a school canteen where many persons exist.We propose a distance-dependent appearance model which is based on scale-invariant feature transform（SIFT） feature.SIFT is a powerful image feature that is invariant to scale and rotation in the image plane and also robust to changes of lighting condition.However,the feature is weak against afne transformations and the identifcation power will thus be degraded when the pose of a person changes largely.We therefore use a set of images taken from various directions to cope with pose changes.Moreover,the number of SIFT feature matches between the model and an input image will decrease as the person becomes farther away from the camera.Therefore,we also use a distance-dependent threshold.The person following experiment was conducted using an actual mobile robot,and the quality assessment of person identifcation was performed.

An approach for automatic sleep stage scoring and apnea-hypopnea detection

In this article we present an application of data mining to the medical domain sleep research, an approach for automatic sleep stage scoring and apnea-hypopnea detec- tion. By several combined techniques （Fourier and wavelet transform, derivative dynamic time warping, and waveform recognition）, our approach extracts meaningful features （fre- quencies and special patterns like k-complexes and sleep spindles） from physiological recordings containing EEG, ECG, EOG and EMG data. Based on these pieces of in- formation, an ensemble of decision trees is constructed us- ing the principle of bagging, which classifies sleep epochs in their sleep stages according to the rules by Rechtschaf- fen and Kales and annotates occurrences of apnea-hypopnea （total or partial cessation of respiration）. After that, case- based reasoning is applied in order to improve quality. We tested and evaluated our approach on several large public databases from PhysioBank, which showed an overall accu- racy of 95.2% for sleep stage scoring and 94.5% for classify- ing minutes as apneic or non-apneic.