Search

Traditional document summarization models cannot handle dialogue summarization tasks perfectly. In situations with multiple speakers and complex personal pronouns referential relationships in the conversation. The predicted summaries of these models are always full of personal pronoun confusion. In this paper, we propose a hierarchical transformer-based model for dialogue summarization. It encodes dialogues from words to utterances and distinguishes the relationships between speakers and their corresponding personal pronouns clearly. In such a from-coarse-to-fine procedure, our model can generate summaries more accurately and relieve the confusion of personal pronouns. Experiments are based on a dialogue summarization dataset SAMsum, and the results show that the proposed model achieved a comparable result against other strong baselines. Empirical experiments have shown that our method can relieve the confusion of personal pronouns in predicted summaries.

Visual Question Answering (VQA) is a challenging task which requires a fine-grained semantic understanding of visual and textual contents. Existing works focus on better modality representations. However, these methods give little consideration to the long-tailed data distribution in common VQA datasets. The extreme class imbalance causes training bias to behave well in head class, but fail in tail class. Therefore, we propose a unified Adaptive Re-balancing Network (ARN) to take care of classification in both head and tail classes, exhaustively improving performance for VQA. Specifically, two training branches are introduced to per-form their own duty iteratively, which learn the universal representations first and then emphasize the tail data progressively by the re-balancing branch with adaptive learning. Meanwhile, contextual information in the question is vital for guiding accurate visual attention. Thus our network is further equipped with a novel gate mechanism to give higher weight to contextual information. The Experimental results on common benchmarks such as VQA-v2 have demonstrated the superiority of our method compared with state of the art.

Visual Tracking plays a key role in computer vision application and artificial intelligent research. The main representation of tracking results come to rectangle boundingbox leading to inaccurate performance, which may not meet the requirements of computer vision application nowadays. We are more likely to obtain the pose estimation of tracking objects in order to achieve more complex computer vision mission including behavior detection and video analysis. While, rectangle boundingbox representation mostly dominates the output model when describing the target appearance in existing DNN-based trackers, leading to precision. We introduce the characteristic points(ChaPoints) to represent the both feature extraction and output model in order to accomplish pose estimation during training and tracking procedure. To this end, we build the mapping method between our proposed characteristic points and rectangle boundingbox. The multi-branches Siamese networks can further matching the candidates and the templates for final tracking results. We extensively prove the effectiveness of the proposed method through the ablation studies of the tracking benchmark, including OTB-2015 and UAV123.

Under most circumstances, it is very important to achieve fast and real-time low-cost micro/nano-satellite fault detection. Regarding of faults as dynamic modes which observe through the multi-sensors, with probabilistic data association based on multi-sensor, we obtain the fault detection results according to the association probability and the threshold values. Joint Probabilistic Data Association (JPDA) algorithm is one of the effective ways for multi-sensor and multi-target tracking. We improve the JPDA algorithm as follows: At first, we propose an approximation method for constructing the confirmation matrix by removing the small probability events using the right threshold, and then, we present the mathematical division of the confirmation matrix according to the intersection area of the association gate of fault targets to be tracked; Finally, we compute the association probability of fault targets through attenuating the value of the public measurement. The simulation results show preliminarily that our improved JPDA algorithm saves the computational time greatly, and meet the requirements of fast and real-time fault detection effectively.

Thermal infrared (TIR) tracking can be utilized to track the target in the images generated by thermal infrared sensors due to the weak influence by illumination changes. However, there are still some challenges to do thermal infrared tracking when suffering drastic appearance variation, heavy occlusion and background clutters. The absence of RGB patterns and low resolution also constrain the tracking performance in complex scenarios. The deep convolutional features are widely utilized to solve visual tracking problems which successfully extracted the spatial and semantic information though object representation. Motivated by these methods, we firstly propose to combine multi-stages cascaded Siamese networks to achieve deep features fusion in three stages, then achieve the tracking procedure by candidates matching strategy. The final results are obtained by non-maximum suppression and scale penalty. The proposed method can inherit the advantages by fusing multi-stages deep features and achieve end-to-end learning simultaneously. The experiments are evaluated with state-of-the-art methods on VOT-TIR2016 benchmark and attributes based comparison. The tracking results demonstrate that our proposed method outperforms the compared methods in terms of accuracy and robustness.

As the development of the computer vision, we could employ the facial expressions recognition technique in the human-computer interaction domain, such as chatting robot, psychological counselling robot and so on. In this paper, we propose a novel method to jointly learn the facial expression features based the combination of Siamese networks and spatial transformer networks. We posit that learning the features from both networks may make them more robust to the image variation especially illumination and viewpoint variation. Particularly, we construct the structure of the Spatial Transformer Siamese Networks(STSN) and learn the features robust to the facial expression image variation. The experimental results show that the proposed method outperforms the-state-of-art method on the MDLTI-PIE dataset.

Recent advances in visual tracking showed that deep Convolutional Neural Networks(CNN) trained for image classification can be strong feature extractors for discriminative trackers. However, due to the long-tail of categories, occlusions, deformations and some other attributes are so rare that they will hardly happen. Yet we want to learn a model invariant to such occurrences for fear that we could handle complex attributes in tracking procedure. In this paper, we propose an alternative solution. We propose to learn an Adversarial Fusion Networks(AFN) that generates examples with occlusions and deformation based on the internal structure of Region Proposal Network (RPN). We discovered that the internal structure of Adversarial Fusion Networks(AFN)'s top layer feature can be utilized for robust visual tracking. We illustrated that such networks can be more robust when the tracking object suffering from occlusion and deformation. Without ensemble and any extra treatment on feature maps, our proposed method achieved state-of-the-art results on several large scale benchmarks including OTB50, OTB100 and VOT2016.

As the development of Artificial Intelligent, visual object tracking plays a key role in computer vision area with numerous real-world applications. This paper proposes a novel approach with the combination of deep spatial transformer features and discriminative correlation filters-based tracking frameworks for the visual tracking problem. The deep spatial transformer features have several advantages compared to the standard deep features (fully connected layers). Firstly, they are more robust when object is suffering affine transformation because of the specific characteristic. Secondly, they have low dimensionality with deeper layer. Lastly, some structural information is contained during tracking procedure. We perform comprehensive experiments on two benchmark datasets: OTB and VOT2015. Surprisingly, compared with some traditional approach based on hand-crafted features, our results further show the affine robustness. We also maintain the accuracy and tracking speed compared with several discriminative correlation filter-based tracker at the same time. Finally, results comparable to state-of-the-art trackers are obtained on all two benchmark datasets.

As the development of Artificial Intelligent, computer vision has became one of the most important elements of all the technologies which composed the AI system, especially robot. Object tracking plays a key role in computer vision. While, there still remain some unsolved problems when the target suffering occlusion, illumination, scale change and rotation. The proposed tracking algorithm obtain the appearance model using the theory of compressive sensing, A LS-SVM classifier if used to separate the positive templates from negative samples. Then, we design a hypergraph propagation method to capture the contextual information on samples in order to improve the tracking accuracy. Updating scheme makes the algorithm more adaptive. Experimental results have proved the effectiveness and robustness of the proposed tracker.

The leakage current is a main characteristic quantity of contamination level of transmission lines insulator surfaces, measuring insulator leakage current is currently the principal means to monitor the contamination level, but because the environment of transmission lines is very complex, electromagnetic interference will must produce much noise at the time of collecting and testing leakage current, sometime we also get the full wave or half cycle, that means some ways to realize selective measurements, to improve the current sensor's abilities of anti-jamming, selectivity and accuracy when the sensors measure insulator leakage current, an optical fiber sensor was employed to measure leakage current of the insulator strings, which used LED as the photoelectric conversion and light signal stability to detect the current. Some tests were operated for three-piece XP-160 insulator strings with ESDD 0.05~0.2mg/cm2 and RH 85%. Results approved that the optical fiber sensor could achieve selective measurement of leakage current waves, the positive half cycle or negative half cycle, and of cause the full cycle current could be monitored after the differential circuit was improved. It had a great degree of linearity fitting, its sensitivity reached 40mV/mA, their pertinence coefficients of linear fitting are all more than 99.94%. its frequency response was 1Hz ~ 10 kHz ，which can meet the basic requirements of the leakage current measurement.

The investigations started from the assumption that perennial plants in the foreland of river oases in the Taklimakan desert are able to take up water from soils of low water content and low water potential during long drought periods. In order to check this assumption, P-V curves were established and selected water relation parameters were derived. Alhagi sparsifolia, Calligonum caput-medusae, Tamarix ramosissima, and Populus euphratica, growing in the foreland of Qira oasis were investigated. Contrary to the expectation, the osmotic potentials both at plant water saturation and at incipient plasmolysis are high in comparison with other desert plants and reveal that the investigated plants do not need to take up water against a low soil water potential. The osmotic potential at incipient plasmolysis is moderately decreasing during the growing season in three of the species, but this seems to be an endogenous process that is not caused by drought stress. Small differences between osmotic potentials at incipient plasmolysis and the minimum plant water potentials that are reached during daily courses point to a transpiration control that accepts very low turgor potentials. A plasmolysis risk is possibly diminished by an increase of cell wall elasticity during the growing season. The results support the conclusion that all investigated species have permanent access to groundwater and live as phreatophytes in the foreland of Qira oasis.

The application of computer in industrial design has become very important with the development of information technology and network technology. The computerization of design tools will change the ideology and management of industrial design inevitably.

In this paper, we consider the problem of designing nonlinear robust formation controllers on a team of Unmanned Aerial Vehicles (UAV) using off-the-shelf autopilots. Two nonlinear robust formation control algorithms are presented. The first controller requires knowledge of the leader’s velocity and acceleration. The second controller, on the other hand, does not have such a requirement. Additionally, the desired trajectory for the follower UAV is generated by a perturbation-based extremum-seeking (ES) controller. With these algorithms, two formation flight objectives have been achieved: (i) a reduction of communication overhead, and (ii) a maximum energy saving mission flight. Numerical simulations validate the performance of the proposed control methodology.