Search

This paper reviews the challenge on constrained high dynamic range (HDR) imaging that was part of the New Trends in Image Restoration and Enhancement (NTIRE) workshop, held in conjunction with CVPR 2022. This manuscript focuses on the competition set-up, datasets, the proposed methods and their results. The challenge aims at estimating an HDR image from multiple respective low dynamic range (LDR) observations, which might suffer from under- or over-exposed regions and different sources of noise. The challenge is composed of two tracks with an emphasis on fidelity and complexity constraints: In Track 1, participants are asked to optimize objective fidelity scores while imposing a low-complexity constraint (i.e. solutions can not exceed a given number of operations). In Track 2, participants are asked to minimize the complexity of their solutions while imposing a constraint on fidelity scores (i.e. solutions are required to obtain a higher fidelity score than the prescribed baseline). Both tracks use the same data and metrics: Fidelity is measured by means of PSNR with respect to a ground-truth HDR image (computed both directly and with a canonical tonemapping operation), while complexity metrics include the number of Multiply-Accumulate (MAC) operations and runtime (in seconds)., Comment: CVPR Workshops 2022. 15 pages, 21 figures, 2 tables

This paper proposes a novel accent recognition system in the framework of a transformer-based end-to-end speech recognition system. To incorporate the pronunciation and linguistic knowledge into the network, we first pre-train an ASR model in a hybrid CTC/attention manner. Then, focusing on accent recognition, we extend the output token list by inserting accent labels to the transcripts and finetune the network parameters with an accented speech dataset. Our work is evaluated on the Interspeech 2020 Accented English Speech Recognition Challenge. Experiments show that our method achieves an accuracy of 72.39% on the test set and 80.98% on the development set, outperforming the baseline system by a very large margin. Our submitted system ranked second in the accent recognition task in the challenge.

Recently, a variety of acoustic tasks and related applications arised. For many acoustic tasks, the labeled data size may be limited. To handle this problem, we propose an unsupervised pre-training method using Transformer based encoder to learn a general and robust high-level representation for all acoustic tasks. Experiments have been conducted on three kinds of acoustic tasks: speech emotion recognition, sound event detection and speech translation. All the experiments have shown that pre-training using its own training data can significantly improve the performance. With a larger pre-training data combining MuST-C, Librispeech and ESC-US datasets, for speech emotion recognition, the UAR can further improve absolutely 4.3% on IEMOCAP dataset. For sound event detection, the F1 score can further improve absolutely 1.5% on DCASE2018 task5 development set and 2.1% on evaluation set. For speech translation, the BLEU score can further improve relatively 12.2% on En-De dataset and 8.4% on En-Fr dataset., Comment: Accepted by ICASSP 2021

With the continuous improvement of China’s scientific and technological level, computer network has become an indispensable part of people’s daily life. It can not only effectively improve the efficiency of production and life, and shorten the distance between people, but also further promote the speed of China’s social and economic development, which has a positive impact on the realization of China’s modernization. Under the new information security demand environment at present, we should pay attention to the related information security work and formulate effective security measures and strategies. In order to effectively prevent these information security problems, people should actively adopt firewall technology, encryption technology, network access control technology and network virus prevention technology for effective protection. This paper analyzes the security problems in the application of wireless sensor networks and explores the mechanism of defending information security, hoping to strengthen the security and stability of wireless sensor networks through effective measures, so that people can better enjoy the convenience brought by the network age.

Recent rapid development of information and communication technology boosts the advance of distributed management and control system, especially for power system. Massive data and information have been accumulated, however, the meaningful fault information hidden in these data is not fully utilized, as the existing fault detection technologies are usually based on monitoring and diagnosis rather than prediction. In this paper, we introduce the deep learning algorithm into the fault prediction of generators in power system, and explore the validity and feasibility of generator operation data in fault prediction application. Our method includes two parts, the first is a Partial Least Square (PLS)-based pre-process module which is used to reduce the feature dimension, the second is a deep linear regression model which is dedicated to regressing the generator operation data and predicting the fault behavior of generators. Experimental results demonstrate the effectiveness of our method.

With the employment of the necessary information and communication technologies (ICTs) infrastructures, the smart grid evolves into a complex and large-scale cyberphysical system (CPS). Massive interacting components challenge the traditional data exchange mode. In order to meet the interaction requirements between intelligent substations and control center, the data service architecture for wide-area transparent access is proposed based on the general service protocol for electric power system. Then the key issue related to location transparency is discussed in terms of cross-domain service discovery as well as the service status analysis for ensuring the capabilities of wide-area services. Finally, the prototype of transparent access is implemented.

The generation-grid-load-energy storage collaborative control lacks effective simulation and testing methods currently, scenario simulation and verification is an effective solution. This paper first describes the architecture of the generation-grid-load-energy storage collaborative control scenario, then focuses on the method of generating wind speed with chronological characteristics. The wake effect in large-scale wind farms and the differences between day and night should be taken into account. Finally, the large-scale wind farm output under the scene simulation is modeled. The timing output of the wind farm is also calculated and compared with the output power of an actual wind farm. The rationality and correctness of the probabilistic timing output simulated in this paper are verified.

In the fault recovery problem of distribution network with distribution generators (DG), the method of island partitioning is always adopted without considering network reconfiguration and tie switches. And in the traditional distribution network reconfiguration problem, the radial constraints of the network are necessary to ensure that all loads are powered. In this paper, a fault recovery strategy, which can rationally coordinate the distribution network reconfiguration with isolated island partitioning, is presented. Firstly, load cutting and islanding operation are allowed by modifying the relevant radial constraints in the reconfiguration problem. Then, the 0-1 state variables in the constraint are used to simplify the network model and improve the computing efficiency. Finally, the original nonconvex nonlinear problem is optimized and relaxed to the standard mixed integer second-order cone programming (MISOCP) problem which can be easily solved via commercial solver. The example results prove the effectiveness of the strategy and the superiority and accuracy of the algorithm.

Based on the “physical distribution and logical integration” architecture of new generation power grid dispatching and control system, with requirements of the crossterminal display, cross-center transparent access and multicenter business collaboration, it is necessary to develop a technical framework of data synchronization based on global analysis and decision. Through unified data synchronization, distributed transactions and data recovery and other functions to achieve multi-center data consistency, improve the centralized management of data in the control system. Finally, through the test and live demonstration, data synchronization function of multi-center was verified.

With the development of smart grid and new energy technologies, both load and source uncertainties need be considered in the calculation of total transfer capability (TTC) of power system transmission. Based on deep learning technology and the improved multi-point estimation method, a fast TTC estimation method considering uncertainties of load power, demand response and wind power generation is proposed. The uncertainties are represented by probability distribution of prediction error and Nataf transformation is introduced to deal with the non-normal probability distributions and their correlations. The stacked denoising autoencoder is employed to estimate TTCs of operating scenarios generated by Nataf transform and the improved multi-point estimation method is used to obtain their probability. Simulation results demonstrate that the fast estimation method is able to consider both load and source uncertainties effectively and calculate TTC fast and accurately.

Alcohol intoxication can affect people both physically and psychologically, and one's speech will also become different. However, detecting the intoxicated state from the speech is a challenging task. In this paper, we first implement the baseline model with ComParE feature and then explore the influence of the speaker information on the intoxication detection task. Besides, we apply a ResNet18 based model to this task. The model contains three parts: a representation learning subnetwork with Deep Residual Neural Network(ResNet) of 18-layer, a global average pooling(GAP) layer and a classifier of 2 fully connected layers. Since we cannot perform speaker z-normalization on the variant-length feature input, we employ the batch z-normalization to train the proposed model. It also achieves similar improvement like applying the speaker normalization to the baseline method. Experimental results show that speaker normalization on baseline model and batch z-normalization on ResNet18 based model provides 4.9% and 3.8% improvement respectively. The results show that speaker normalization can improve the performance of both the baseline model and the proposed model.

In this paper, we describe our submitted DKU- Tencent system for the oriental language recognition AP18- OLR Challenge. Our system pipeline consists of three main components, including data augmentation, frame-level feature extraction, and utterance-level modeling. First, we perform speed perturbation to increase the diversity and amount of training data. Second, we extract several kinds of frame-level features, including the hand-crafted acoustic features as well as the deep phonetic features. Third, we aggregate the frame-level features into fixed-dimensional utterance-level representation through i- vector and x-vector modelings. We also propose a deep residual network to obtain the utterance-level language posteriors in an end-to-end manner. Our submitted primary system achieves C avg of 0.0499, 0.0146, and 0.0135 for the corresponding short- utterance, confusing language and open-set tasks on the evaluation set.

Many applications rely on the local descriptors extracted around a collection of interest points. Recently, the security of local descriptors has been attracting increasing attention. In this paper, we study the possibility of image reconstruction from these descriptors, and propose a coarse-to-fine framework for the image reconstruction. By resorting to our gradually reconstructing network architecture, the novel multiscale feature map generation algorithm, and the strategically designed loss functions, our proposed algorithm can recover the images with very high perceptual quality, even partial descriptors are provided only. Extensive experimental results are reported to show its superiority over the existing algorithms. Our study implies that the local descriptors contain surprisingly rich information of the original image. Users should pay more attention to sensitive information leakage when using local descriptors.

Wind power ramp events had a significant impact on the power balance of power system and may lead to load shedding. A data driven method was proposed for wind power ramp events assessment in this paper. The K-means clustering algorithm was used to divide the samples to several classes. The stacked denoising autoencoder was used to extract layer features to train support vector machine. Historical and forecast data of wind power, load power, conventional unit and pumped storage station power were taken as inputs. The output was whether ramp event occurred. A severity function was constructed to assess the severity grade which was predicted to be a wind power ramp event based on effect theory. The credibility of the assessment result was represented by confidence interval. Simulation results of a provincial power grid showed that the prediction method in this paper was more accurate and credibility was high enough to help the dispatchers to take measures for the security of power grid.

For a given power system operating condition, sensitivity analysis is needed when it is necessary to analyze how the changes of some variables will affect other variables. With the increasing scale of the power grid, it is computationally expensive to perform sensitivity analysis under various operating conditions, such as changes in generation power and line outages. In order to accelerate sensitivity analysis of large-scale power grids, a batch computing method based on graphics processing unit-central processing unit (GPU-CPU) heterogeneous computing framework is proposed. Considering the coarse-grained parallelism of the calculation under different operating conditions and further mining fine-gain parallelism in calculation is an effective approach to improve computational efficiency. This paper proposes a fine-grained parallel batch calculation method to simultaneously obtain line outage distribution factors and generation shift distribution factors under various operating conditions. Finally, by comparing with the standard example, the effectiveness of the proposed method is verified by case study.

Faced with the rapid development of UHV AC-DC transmission technology and networking technology, the power grid pattern has undergone great changes, making the impact of faults on the system global, so transient security assessment is need. To accurately simulate the dynamic characteristics of the AC-DC power systems after fault in the evaluation process while reducing the simulation scale and calculation amount, it is necessary to use a dynamic equivalent to represent the systems. The determination of the equivalent boundary usually has a great influence on the dynamic performance of original system. Therefore, this paper analyzes the coupling between AC and DC systems by simulating the dynamic characteristics of AC-DC power systems with different equivalent boundaries and evaluating the DC response after faults occurring at different locations of the AC system, which provides guidance on the determination of dynamic equivalent boundary.

In the task of the unsupervised query by example spoken term detection (QbE-STD), we concatenate the features extracted by a Self-Organizing Map (SOM) and features learned by an unsupervised GMM based model at the feature level to enhance the performance. More specifically, The SOM features are represented by the distances between the current feature vector and the weight vectors of SOM neurons learned in an unsupervised manner. After fetching these features, we apply sub-sequence Dynamic Time Warping (S-DTW) to detect the occurrences of keywords in the test data. We evaluate the performance of these features on the TIMIT English database. After concatenating the SOM features and the GMM based features together, we achieve an improvement of 7.77% and 7.74% on Mean Average Precision (MAP) and P@10 on average.

Typical temperature control loads, like Heating, ventilating and air conditioning(HV ACs), are distributed comprehensively which have large potential to respond to system control. A set of simulation methods of HV ACs are proposed in this paper, including multi-agent system modeling method of HV ACs and a series of interfaces developing methods. With these interfaces.the automatic dispatching system, D5000, can be driven by HV ACs simulation system. At last, the HV ACs' response process is simulated, by participating day-ahead market, and the results are analyzed.

The distributed generations (DGs) make the establishment of switch function complex, and the faulty section location methods for traditional distribution networks are no longer applicable. In order to improve the rapidity and accuracy of faulty section location, a Multi-Objective Particle Swarm Optimization (MOPSO) based faulty section location method for distribution networks with photovoltaic (PV) generations is proposed. The influence of fault current characteristics of PV generations under different light intensities is taken into account. Switch function for dynamic switching of PV generations is proposed. Since the single objective optimization intelligence algorithms easily cause the premature convergence and the computational complexity of the NSGA-II algorithm is high, the MOPSO algorithm is used to solve the problem, which can avoid the determination of the weighting factors. The simulation results show that the proposed method improves the rapidity and accuracy of location effectively, and has superior fault-tolerance to distortion information.

With the extensive penetration of distributed generators (DGs), especially photovoltaic generators and other intermittent DGs, the intrinsic characteristic of three-phase unbalanced condition of distribution network becomes even more prominent. It brings a new challenge to the evaluation of power supply capability (PSC) for distribution network. The evaluation model of PSC for unbalanced three-phase distribution network with DGs is proposed. In this model with the objective of maximizing the power supply load for distribution network, the operation constraints of bus voltage magnitude, branch current, transformer capacity, packet switch capacitor and three-phase bus voltage unbalance degree are taken into account. An algorithm of quantum particle swarm optimization (QPSO) was put forward to solve this problem. The algorithm uses the advantages of superposition state and probability expression, which improves the ability for global optimization. The simulation results of the modified IEEE33-bus three-phase distribution system show that the proposed model and algorithm are effective.

Mid-long term load forecasting is the prerequisite of power system planning. This paper first introduces the sub-area partitioning of power grid in Nanjing core area. Then the influences of social-economic factors to mid-long term load are studied by two ways, which are correlation analysis method and multiple linear regression analysis method. Based on the analysis results of the two methods, a grey MGM (1, N) load forecasting model is constructed. Furthermore, by forecasting the maximal load of different districts, the multiple linear regression Grey MGM (1, N) model and ordinary multiple linear regression models are compared. The results show that the method proposed in this paper has higher accuracy and wider adaptability. Finally, this paper analyzes the load developing tendency in different districts of the city and proposes corresponding suggestions.

A non-ideal iris image enhancement algorithm based on local standard deviation was put forward. The algorithm can enhance gray and color iris images with poor contrast. Experiments show that the proposed method can increase the correct acceptance rate and reduce the false rejection rate of iris recognition system, so the algorithm can improve the accuracy of iris recognition system.

A novel iris image segmentation method using watershed and region merging is put forward. Total variation flow model is used to divide the eye image into structure part and texture part. Light spots in pupil are removed by closing. Watershed and region merging are both applied in structure part. Structure part is divided into separate parts. The regions may contain iris are labeled. Then, merging iris regions using average gray rule and total pixels rule. CASIA3.0-Interval iris image database is used to test the proposed method. Experiment results show that, the proposed method can segment iris images accurately and it can be used for different types of iris images.

In this paper, a new stochastic model based on cellular automata is established to simulate the occurrence and development of Schistosoma japonicum (S. japonicum) infection in an endemic population. We included the process of the pathogen invasion from exposure to worm development and till worm death when the infection is cleared in the model. We further utilized the model to predict the prevalence as the outcomes of the selected chemotherapy carried out in Jiahu village. Comparing model predicted prevalence and intensities with the observed parameters, it is anticipated that our cellular automaton transmission model can serve as a tool for studying schistosomiasis transmission dynamics in endemic areas.

A novel approach to achieve real-time protocol implementations is presented in this research. TWI, μC/OS-II, light sensor and our testing tools execute the protocol in the prediction and evaluation embedded systems of the light environment. The protocol is used by software simulation instead of TWI hardware. The method saves the hardware cost and brings the TWI protocol to a popular application. Furthermore this method is valid by validation test.