Abstract:Due to the diversity of the distribution and shape of rain streaks, existing rain removal algorithms produce problems such as blurred image background and poor generalization performance while removing rain.A multi-scale feature fusion image rain removal approach based on attention mechanism is proposed to address these problems.The feature extraction consists of multiple residual groups containing two multi-scale attention residual blocks, which use the multi-scale feature extraction module to extract and aggregate feature information at different scales and further improve the feature extraction capability of the network through coordinate attention.Local feature fusion is performed within groups, and the global feature fusion attention module is used between groups to better fuse features at different levels and to focus the network on rain streak regions through pixel attention.The quantitative metrics of the proposed approach are significantly improved compared with other existing image rain removal algorithms on both simulated and real rain image datasets, and the rain removal images are greatly improved in both visual effects and generalization performance.

Abstract:Generative Adversarial Network (GAN) can generate images that are close to real images, thus plays an important role in image style transfer.However, the GAN-based image transfer is perplexed by problems of low quality of generated images and difficult training of models, herein a new style transfer approach based on BicycleGAN model is proposed.First, the residual module is introduced into the generator of GAN to solve the degradation of GAN in training, and the self-attention mechanism is employed to obtain more image features thus improve the generation quality of the generator.To solve the gradient explosion in the training of GAN, the spectral normalization is added behind each convolution layer of the discriminator.Then the perceptual loss is introduced to address the unstable training and low generated image quality.The experiments on Facades and AerialPhoto&Map datasets show that the proposed approach outperforms other image style transfer methods in the PSNR and SSIM values of the generated images.

Abstract:In order to solve the problems of color distortion, key information blur and detail loss perplexed underwater image, an underwater image enhancement method based on SK attention residual network is proposed.The generator structure in the generative adversarial network is improved, and a residual module is introduced to reduce the feature loss between encoder and decoder, thus enhance the image detail and color.To make the network adapt to different scale feature maps to extract key information of images, the SK attention mechanism is added after the residual module.Meanwhile, a parametric rectified linear unit is used to improve the fitting ability of the network.This method is verified on real and synthetic underwater image datasets, and traditional method and deep learning method are used for subjective and objective evaluations.In the subjective effect analysis, it is found that the color, key information and detail features have been greatly improved in enhanced images.In the objective evaluation, it is found that the indicator values of the proposed method are higher than those of existing underwater image enhancement algorithms, which verifies the effectiveness of this method.

Abstract:To address the issue that convolutional neural networks can hardly balance the local details and global semantic consistency of results in the process of image inpainting, a multi-scale semantic learning model for face image inpainting based on generative adversarial networks is proposed.First, the face image is decomposed into components with different perceptual fields and feature resolutions using gated convolution, and multi-scale features are extracted using convolution kernels of different sizes to enhance the detail of the restoration results by extracting appropriate local features.Second, the extracted multi-scale features are fed into the semantic learning module to learn the semantic relationships between features from both the channel and spatial perspectives, thus enhance the global consistency of the restoration results.Finally, skip connections are introduced to complement the features on the encoding side to the decoding side to reduce the loss of detail information caused by sampling and improve the texture details of the restoration results.Experiments on the CelebA-HQ face dataset show that the proposed model has significant improvements in three performance metrics:peak signal to noise ratio, structure similarity and l₁, and the inpainting results are visually more reasonable in terms of local details and global semantics.

Abstract:Micro-expression is the facial expression that people reveal involuntarily when they try to hide their true emotions, which is a hot spot in research of affective computing in recent years.Micro-expression is a subtle facial movement thus is difficult to recognize.Considering its excellent performance in image classification and ability to capture subtle feature information, the cross-attention multiscale ViT (CrossViT) is used as the backbone network to improve the cross-attention mechanism in the network, and the Dual Attention (DA) module is proposed to extend traditional cross-attention mechanism to determine the correlation between attention results, thus improve the micro-expression recognition accuracy.The proposed network learns from three optical flow features (optical strain, horizontal and vertical optical flow fields), which are calculated from the starting frame and peak frame of each micro-expression sequence, and classifies the micro-expression by Softmax.Experiments on the micro-expression fusion dataset show that the proposed network reaches 0.727 5 and 0.727 2 in UF1 and UAR, respectively, which is more accurate than the mainstream micro-expression recognition algorithms, verifying the effectiveness of the dual attention CrossViT based network.

Abstract:There are quite a few comment sentences without emotional words in aspect-level emotional texts, and the study of their emotions is called aspect-level implicit sentiment analysis.The existing models have the problems that the context information related to aspect words may be lost in the pre-training process, and the deep features in the context cannot be accurately extracted.Aiming at the first problem, this paper constructs an aspect-aware BERT pre-training model, and introduces aspect words into the input embedding structure of basic BERT to generate word vectors related to aspect words.Aiming at the second problem, this paper constructs a context-aware attention mechanism.For the deep hidden vectors obtained from the coding layer, the semantic and syntactic information is introduced into the attention weight calculation process, so that the attention mechanism can more accurately assign weights to the context related to aspect words.The results of comparative experiments show that the proposed model outperforms the baseline model.

Abstract:Virtual reality haptic rendering has high requirements for image texture feature extraction.However, a single texture extraction algorithm cannot accurately describe the characteristics of image texture due to the complex and irregular texture factors.Therefore, a texture material classification approach based on GLCM (Gray-Level Co-occurrence Matrix) and Tamura fusion features is proposed.Additionally, we optimize the GLCM and propose the T-GLCM operator, thus improve the rotation invariance of GLCM pair and reduce a lot of redundant information.In this approach, the Tamura texture features are used to quantify the image, and the feature regions are quantified and then cascaded into a set of feature vectors.The texture features of T-GLCM are fused, and the texture materials are classified by Support Vector Machine (SVM).The experimental results show that the proposed approach outperforms traditional texture feature extraction algorithms in classification accuracy and robustness.

Abstract:To improve the accuracy of short-term wind direction forecasting, a hybrid model, named EEMD-CNN-GRU, is proposed based on Ensemble Empirical Mode Decomposition (EEMD), Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU).The EEMD is used to decompose the data into multiple components to address the randomness and unsteadiness of wind direction series, then the local connection and weight sharing of CNN are employed to extract the potential features in each component, and the GRU is adopted to reconstruct the extracted features and superpose the predicted values of each component to obtain the final prediction results.The experimental results show that the proposed method outperforms models of BP neural network and long short-term memory (LSTM).

Abstract:Accurate prediction of wind speed in extreme weather can provide important guidance for distribution network to enhance disaster prevention and resilience.This paper proposes a method based on Temporal Convolutional Network (TCN), Bi-directional Long Short-Term Memory(BiLSTM) and error correction for wind speed prediction in extreme weather.First, the time series characteristics of multi-feature weather data are extracted by TCN, and then input into BiLSTM for wind speed prediction.To further improve the prediction accuracy, Variational Mode Decomposition (VMD) is introduced to decompose the error sequence, and BiLSTM models are constructed to perform error prediction for the decomposed error subsequences respectively.Then the error prediction value is used to correct the wind speed prediction value.Finally, simulations are carried out for a place of Henan province, and the results show that compared with measured weather data, the proposed method can effectively predict wind speed with high accuracy when extreme weather occurs.

Abstract:To meet the application needs of speaker recognition for short-duration speech in open scenarios, we herein optimize the speaker recognition model in aspects of accuracy and robustness.First, to realize the selection of important frequency features from the input acoustic data, a Reweighted-based Feature Enhancement Layer (RFEL) and a Reweighted-based Feature Enhancement Network (RFEN) are proposed to enhance the feature representation.Second, the loss function of Misclassified Vector guided Softmax loss (MVSoftmax) in face recognition is introduced into the speaker recognition to improve the mining ability towards hard samples.Third, a combined loss function of MVSoftmax and few-shot learning based Angular Prototypical loss (AP) is proposed, which solves the mismatch between the classification loss function and the actual evaluation requirements of speaker recognition, and relieve the strong dependence of the metric function on the sampling strategy.Finally, the experimental results show that the performance metric EER of the proposed model is reduced by 12.45% and the minDCF is decreased by 14.09% compared to the baseline model, achieving excellent performance in speaker recognition.

Abstract:To improve the access capacity of Distributed Photovoltaic (DPV) in the low-voltage distribution station area and promote photovoltaic consumption, a distributionally robust optimization method of DPV access in the low-voltage distribution station area is proposed considering the timing characteristics of source-load.First, a source-load joint timing scenario generation method based on optimized clustering is presented to handle the uncertainty of distributed photovoltaic output and load demand in the low-voltage distribution station area.Next, a distributionally robust optimization model of distributed photovoltaic access in the low-voltage distribution station area is constructed by taking into account the voltage constraints, line capacity constraints, reactive power compensation constraints of inverter and photovoltaic consumption constraints, etc.The proposed approach maximizes the access capacity of DPV while ensuring that the expected value of PV curtailment rate under the worst probability distribution of each typical scenario meets the requirements.Then, a mathematical model of distributed energy storage access in low-voltage distribution station area is established to study the influence of energy storage access and its charging-discharging mechanism on distributed photovoltaic access.Finally, the effectiveness of the proposed model is verified by taking simulation on actual low-voltage distribution station area.

Abstract:In view of the multi-disciplinary integration of intelligent manufacturing engineering, a robot sorting experiment system based on 3D vision are designed.The hardware experiment platform is built using Kinect camera, industrial robot, PC, and end-effector, while a Support Vector Machine (SVM) algorithm is designed to recognize target objects.Additionally, a method of cavity burr repair is proposed which combines median filter preprocessing with nearest neighbor interpolation.To check whether the target objects overlap or block each other, a strategy based on Hough transform is designed to calculate the object's center position and another strategy based on point cloud registration to estimate the object's pose.Then a series of robot sorting experiments are carried out under the guidance of upper computer interactive interface.The experiment results show that the system can accurately identify and stably sort the target objects of specific shape and color.The designed experiments involve knowledges and technologies related to robot, machine learning, image processing, software & hardware design, etc.In view of its strong comprehensiveness and openness, the proposed system can provide a comprehensive and innovative practice platform for intelligent manufacturing engineering laboratory.

Abstract:In view of the problem that traditional static parameter equivalent circuit model of a supercapacitor cannot effectively reflect its dynamic operating characteristics, a second-order ladder equivalent circuit model including dynamic charge/discharge resistance and capacitance parameters is proposed.The recursive least square method is used for preliminary offline parameter identification of the second-order ladder equivalent circuit model.Then the offline identified model parameters are used as initial values considering the dynamic change of the supercapacitor parameters, and the recursive least square method with forgetting factor is introduced to identify the dynamic parameters of supercapacitor in real-time.Simulation model and experimental test platform of the supercapacitor are established to verify the effectiveness and accuracy of the proposed equivalent circuit model and real-time identification method through comparison of simulation and experimental results.The results show that the second-order ladder equivalent circuit model with dynamic parameters can effectively reflect the dynamic charge/discharge operating characteristics of the supercapacitor.Compared with traditional three branch and second-order ladder equivalent circuit models with static parameters, the proposed model is improved in accuracy by over 2.08% and 3.56%, respectively.

Abstract:The severe haze pollution has become one of the major obstacles to the ecological integration development of the Yangtze River Delta (YRD).This paper constructs a haze control efficiency evaluation system taking labor, capital and technological innovation as input indicators, and the number of days with air quality at or better than level 2, PM_2.5 concentration and AQI as output indicators, and uses a three-stage DEA to measure the haze control efficiency of 27 cities in the YRD from 2014 to 2019.The results show that the three-stage DEA model can effectively remove the influence of environmental factors and random errors, and can truly reflect the level of haze control efficiency in the YRD.The change of return to scale is significantly influenced by environmental and random factors, and most cities are in the stage of increasing return to scale.The overall haze control efficiency of YRD cities from 2014 to 2019 is high and shows a W-shaped change trend, with significant inter-city differences.The GDP per capita and the share of secondary industry have a positive effect on haze control efficiency, while population size and urbanization rate have a negative effect on haze control efficiency.