The research works of Chinese font style transfer were classified according to different stages of research development. The traditional methods were briefly reviewed and the deep learning-based methods were combed and analyzed. The commonly used open data sets and evaluation criteria were introduced. The future research trends were expected from four aspects, which were to improve the generation quality, enhance personalized differences, reduce the number of training samples, and learn calligraphy font style.

An underwater image enhancement algorithm was proposed based on generative adversarial networks (GAN) and improved convolutional neural networks (CNN) in order to solve the problems of haze blurring and color distortion of underwater image. Generative adversarial network was used to synthesize underwater images to effectively expand the paired underwater data set. The underwater image was decomposed by multi-scale wavelet transform without losing the feature resolution. Then, combined with CNN, the compact learning method was used to extract features from multi-scale images, and skip connection was used to prevent gradient dispersion. Finally, the fog blur effect of the underwater image was resolved. In order to improve the color correction ability of the model and overcome the problem of color distortion of underwater images, the correlation between different channels of color images was learned by using the style cost function. Experimental results show that, in subjective visual and objective indicators, the proposed algorithm is superior to the contrast algorithm in comprehensive performance and robustness.

In order to promote the transformation of industrial cyber security defense mode from static passive defense to active defense, and alleviate the contradiction between the serious shortage of security experts and the sharp increase of cyber security demands, a cyber security active defense system framework of digital twin system was built from the perspective of bionics, and then five kinds of key technologies focusing on active defense were proposed based on the digital twin security brain (DTSB), including security data interaction and systems collaborative defense based on cloud-edge collaboration, cyber security active defense model of parallel digital twin system, situation awareness of parallel digital twin systems based on digital twin security brain, active defense and control technical framework for digital twin system based on immune system, and anti-attack intelligent recognition of digital twin system based on artificial intelligence. A case study of a digital twin workshop was given to demonstrate the successful application of digital twin cyber security in smart manufacturing.

In order to achieve the parallel optimization of route planning, and solve the problem of high response time of location-based services (LBS) caused by extensive concurrent queries during peak hours, a dual-level grid index (DLG-index) was firstly introduced, and then, based on DLG-index, a parallel optimization algorithm of route planning (PORP) was introduced. The top layer of DLG-index is a skeleton graph consisting of border nodes of the entire graph, and the bottom layer is composed of all grids partitioned by the entire graph. For a given query, the first step is to compute a global path based on the skeleton graph. Then the route planning task is divided into multiple local optimizations in grids passed by the global path. At the same time, each local optimization is maintained independently by different processors. The algorithm can optimize the planned route in real time based on varying traffic conditions. The entire optimization is implemented in several segments, which can be handled by multi-processors and achieve rapid response to massive concurrent queries. Experiments results showed that compared with CANDS algorithm, the response time of PORP was reduced by an average of 49.6% and the processing time was saved by an average of 28.5%.

A new knowledge-enhanced graph convolutional neural network (KEGCN) classification model was proposed aiming at the problem of text classification. In the KEGCN model, firstly a text graph containing word nodes, document nodes, and external entity nodes was constructed on the entire text set. Different similarity calculation methods were used between different types of nodes. After the text graph was constructed, it was input into the two-layer graph convolutional network to learn the representation of the node and classified. The KEGCN model introduced external knowledge to compose the graph, and captured the long-distance discontinuous global semantic information, and was the first work to introduce knowledge information into the graph convolution network for classification tasks. Text classification experiments were conducted on four large-scale real data sets, 20NG, OHSUMED, R52 and R8, and results showed that the classification accuracy of the KEGCN network model was better than that of all baseline models. Results show that integrating knowledge information into the graph convolutional neural network is conducive to learning more accurate text representations and improving the accuracy of text classification.

The research progress of direct air capture CO₂ adsorbents was reviewed. The advantages and disadvantages of alkali/alkaline metal based adsorbents, metal organic framework adsorbents, amine loaded adsorbents and moisture swing adsorbents were compared. Meanwhile, the properties of adsorbents from the aspects of adsorption capacity and amine efficiency, kinetics and supporters, regeneration mode and energy consumption, thermal stability and resistance to degradation were evaluated. Additionally, the related engineering demonstration projects and economic evaluation were briefly discussed. Finally, the problems existing in the current research were summarized, and the future research direction was prospected.

A publicly verifiable secret sharing technology was proposed based on the threshold secret sharing technology, in order to study the security of the user’s private secret in the blockchain. The secret fragments can be verified after the participating nodes receiving them, which can effectively preventing the master splitting node from doing evil when splitting the key. The secret fragments of the nodes participating in the secret splicing are publicly verified through the secret recovery phase, to prevent the nodes in the secret recovery phase from doing evil. Identity IDs are added to the participating nodes during the secret distribution phase, thus malicious nodes can be tracked and the node status can be updated in real time. The dynamic threshold mechanism was designed so that after the node holding the secret fragment is offline, the owner of the secret fragment and the master node can redistribute the secret fragment to the new participating nodes to ensure the integrity of the private secret fragment. Experimental results show that the private secret recovery rate of this scheme can reach 80%, and it has threshold characteristics, traceability, unforgeability and recoverability.

Traditional and new geotechnical engineering problems such as compressed air energy storage, intercepting water with compressed air, carbon dioxide sequestration and oil and gas underground reserve project are all involving air-water two-phase flow and stress coupling problems. For this engineering reality, based on the weak coupling theory of gas-water two-phase seepage and stress in unsaturated soil, a air-water two-phase percolation-stress coupling calculation program based on coupled TOUGH2 and FLAC3D was developed. The calculation program can simulate real air-water two phase flow, and can investigate the gas-water interaction of seepage process. The calculation program considers the direct interaction between gas-water two-phase seepage and soil skeleton deformation, reflects the process of porosity, permeability, capillary pressure and the change of soil physical and mechanical parameters, and achieve a more perfect gas-water two-phase seepage-stress coupling analysis. Furthermore, by comparing with classical drainage test and model test, it is verified that the program can accurately simulate the gas-water two-phase flow-stress interaction.

A join query optimization algorithm in a multiple blockchain environment was proposed, in order to improve the efficiency of join query processing on multi-blockchain. In this method, semantic information is added to the traditional multi-blockchain model, and a semantic multi-blockchain model is constructed to provide a basis for join query on multi-blockchain. Based on this model, referring to the index structure of the distributed database, a join index structure was proposed, which realizes attribute connection of multiple blockchains, improves the efficiency of connection calculation, and reduces the communication cost of data transmission. On these basis, a optimization algorithm about multi-blockchain join query was proposed to improve the efficiency of multi-blockchain connection query. The empirical study of the proposed method was conducted on two real public data sets. Results show that the connection index structure between multiple blockchains is stable. Compared with the traditional join query operation, multiple blockchain connection query optimization method simplifies the query processing process. Query results can be directly obtained by accessing the join index, which reduces local computing load and network overhead, and improves query efficiency.

A novel EEGNet variation based on the fusion of the Inception and attention mechanism modules was proposed, called IncepA-EEGNet, in order to achieve more efficient P300 signal feature extraction. Convolutional layers with different receptive fields were connected in parallel. The network’s ability to extract and express EEG signals were enhanced. Then the attention mechanism was introduced to assign weights to the features of different filters, and important information was extracted from the P300 signal. The validation experiment was conducted on two subjects of BCI Competition III dataset II. Results showed that the IncepA-EEGNet recognition accuracy reached 75.5% after just 5 epochs compared with other deep learning models. The information transmission rate was up to 33.44 bits/min on subject B after 3 epochs. These experimental results demonstrate that the IncepA-EEGNet effectively improves the recognition accuracy of the P300 signal, reduces the time of repeated trials, and enhances the applicability of the P300 speller.

Difference between non-liveness attack types is neglected, and adverse impact of category imbalance between liveness and non-liveness samples on model training is not considered in existing studies of silent liveness detection. In this paper, non-liveness attacks were subdivided into two categories, print attack and display attack, which transformed silent liveness detection from traditional two-classification problem into multi-classification problem. And the cross-entropy was used as the loss function to train network model. Thus, the disadvantage of binary classification and category imbalance can be eliminated, common features of the non-liveness face samples were likely to be identified more accurately through model training, and the accuracy of the network model was improved for non-liveness recognition. Moreover, a two-stream feature fusion the network model was constructed to further improve the feature representation capacity of the network model, which adopted the attention mechanism to adaptively fuse the feature vectors extracted from RGB and YCrCb. Abundant comparative experiments were performed on four public datasets, CASIA-FASD, Replay-Attack, MSU-MFSD and OULU-NPU. Experimental results indicate that silent liveness detection model adopting multi-classification strategy and feature fusion can effectively reduce the classification error and improve over-generalization ability.

A fault-tolerant strategy based on cache queue was proposed, aiming at the problem of low efficiency of stream processing job recovery due to global rollback after Flink task fails. In the job, the operator with the longest recovery time is taken as the key operator, the processed data are stored in the buffer queue, and active backup is performed for it. The backup operator will also accept the data from the upstream to reach the the effect that the job can be restored instantaneously after a failure. In order to solve the additional consumption caused by active backup, a data filtering algorithm was proposed. The backup operator will retrieve the current data from the cache component before processing the data each time to determine whether to continue processing. When the Flink operator itself fails, it will use the buffer queue in the strategy and Flink’s JobManager to send the data information at the time of the failure to the backup operator. When the backup operator receives the data, it will realize the effect of instant recovery. The strategy was evaluated on four evaluation indicators. Compared with the failure recovery mode of Flink1.8, the proposed strategy had a significant improvement in Flink task failure recovery. The recovery efficiency was increased by 56.3%, 51.3%, 46.2% and 45.8% under failure times of 1, 2, 3 and 4 separately. And the proposed strategy brings only a very small price in terms of processing delay, CPU utilization and memory usage.

An graph based intent identification embedding (G2IE) method was proposed, in order to solve the problems of behavior uncertainty and data sparsity faced by collaborative recommendation and sequence representation methods in user behavior prediction. In G2IE method, firstly the theory of planned behavior (TPB) is used to mine the controlled behavior patterns in the user behavior sequence, then the transfer intention intensity of the uncertain behavior list between adjacent controlled behaviors is calculated based on information entropy, and finally the behavior relationship is strengthened by integrating the behavior transfer intention to make up for the lack of behavior intention. In G2IE method, the uncertainty of behavior is identified and it is measured with a model, in order to solve the problem of behavior randomness. The problem of data sparsity can be alleviated to some extent by discovering more behavior relationships through the fusion of transfer intention. G2IE method has more accurate and rich expression ability compared with other methods that use behavior direct relation. Experimental results on three public user behavior datasets demonstrate the effectiveness of the proposed method.

At present, there is no public dataset suitable for the research work of material entity relationship extraction technology in the field of materials. Aiming at the above problem, the construction method of aluminum-silicon alloy entity relationship extraction dataset was proposed through the literature of high-silicon aluminum alloy spray deposition. The construction standards of the aluminum-silicon alloy entity relationship extraction dataset were formulated under the guidance of experts in the material field, and the collected data were marked with entities and relationships according to the construction standards. After the annotation was completed, the aluminum-silicon alloy entity relationship extraction dataset was generated through data preprocessing. Experiments were conducted through the entity-relationship joint extraction model to verify that the dataset can be applied to entity-relationship extraction tasks. Compared with the public dataset, the semantics and grammar of the sentence in the material dataset were more complicated, and there were more long sentences, which led to a slightly worse performance of the entity relationship joint extraction model on the material dataset. Therefore, a self-attention mechanism was added to the entity relationship joint extraction model, which increased the overall F1 value by about 5.8%. The method of constructing the dataset is universal, and the material dataset can be constructed by the construction method.

A new method named automatic segmentation (AUCSN), which combines edge detection algorithm with the object-oriented method, was proposed in order to solve the problems such as noise interference, spectral confusion, segmentation scale error, and water index unavailable which happen in water extraction from unmanned aerial vehicle (UAV) images. An anisotropic diffusion filtering algorithm was used to denoise the image. The Canny edge detection operator was used to extract the edge of the denoised image, and the extraction results were reconstructed with the denoised image. Then an improved absolute mean difference variance ratio method was used to select the optimal segmentation scale for the reconstructed image to conduct multi-scale segmentation. A model combined with the spectral, morphological, and texture feature of the water object was established in order to coarsely extract water objects from the segmented image. The morphological closed operation was used to fill the holes of the coarse extraction results, realizing water extraction. Results show that the AUCSN method can improve the extraction efficiency and the extraction accuracy can reach 96%.

A spatial global context information network (NC-Net) was proposed based on the U-Net baseline network in order to solve the problem that the convolutional neural network (CNN) easily lost spatial information and the boundary information was unclear in the feature extraction stage of semantic segmentation. A re-encoding stage was added (ReEncoder) in order to enhance the ability of spatial information recognition. Multi-scale features were output in the Decoder stage, which was combined with the ReEncoder stage to obtain global context information. The boundary loss function was retained, and a multi-scale loss function cascade method was designed to optimize the overall network. The experimental results on the GID and WHDLD data sets show that the overall accuracy of the method achieves the best results, significantly outperforming other baseline models.

A dehazing algorithm combined with atmospheric scattering model based on generative adversarial network was proposed in order to improve the performance of image dehazing. The algorithm was improved base on pix2pix GAN. Firstly, the generator is improved to a double decoder structure. The double decoder generates the fog-free image and the transmittance image, separately, then the fog-free image and the transmittance image are combined to restore the fog image by the atmospheric scattering model. The purpose is to improve the quality of decomposition. Secondly, in the Markov discriminator structure, the reverse learning mechanism is used to replace the random cropping mechanism, which aims to reduce the probability of inaccurate judgment caused by the random cropping algorithm. Finally, the haze loss function is added to the original loss function to improve the quality of image translation. The ablation experiments and contrast experiments were applied on STOS and NYU datasets. Experimental results showed that the proposed method was better than the original algorithm pix2pix GAN in terms of PSNR and SSIM, and both were better than the existing dehazing algorithms. The restored images have the advantages of high-resolution, low noise and rich texture.

In order to meet the demand for deformation monitoring of rockfill dams of 200 m to 300 m, and to compensate for the deficiencies of using single-track synthetic aperture radar (SAR) data that can only measure two-dimensional deformation in the line of sight and azimuth directions, interferometric synthetic aperture radar (InSAR) measurements and conventional point monitoring data were fused across scales based on ensemble Kalman filtering to improve the monitoring accuracy of rockfill dam exterior deformation and multi-dimensional monitoring data were used to reconstruct exterior deformation field. Shuibuya concrete-face rockfill dam was taken as an example, and results show that the new monitoring technology with characteristic of “large range, low accuracy and high efficiency” and the conventional monitoring technology with characteristic of “discrete points, high accuracy and low efficiency” can complement each other through multi-source data fusion. By reconstructing the exterior deformation field based on multi-dimensional monitoring data, the overall deformation state of the rockfill dam can be comprehensively grasped and the possibility of missing or misjudging the actual deformation by using single-dimensional monitoring results can be reduced. The method can also be used for deformation monitoring and exterior deformation field reconstruction of the reservoir bank slopes.

A pedestrian following method based on adaptive follower mechanism was proposed focusing on the problem that robots lose targets easily with fixed sensors. Based on task requirements, the field of view evaluation metrics of the follower perception mechanism were designed. On the basis of traditional planning strategies, an improved planning strategy derived from chassis direction and a depth weighting based adaptive angle planning strategy were proposed to improve the moving target following performance of the follower mechanism. To improve pedestrian position tracking with follower RGB-D sensor, the YOLOv3 algorithm was used for target detection, combined with 3D coordinate solving and position matching to achieve real-time tracking of multiple targets. Gazebo simulation platform and RoboMaster robot were used to implement robot's pedestrian following function. The proposed planning strategy is shown to achieve comprehensive optimal metrics and stable trajectory following to moving pedestrian targets, which proves the effectiveness of the target following method.

An API recommendation method based on natural nearest neighbors and collaborative filtering named N-APIRec was proposed in order to solve the problem of recommendation performance degradation caused by improper neighbor selection. In this model, BM25 algorithm was used to transform the projects into vectors. Then the natural neighbor algorithm was used to filter the similar projects in the dataset to reduce the search scope, and the similar method declarations were filtered from the similar projects. Finally, the APIs were recommended through collaborative filtering. N-APIRec was compared with the state-of-the-art approach on MV and SH data sets. The results were verified the effectiveness of N-APIRec, the success rate of MV and SH data sets recommendation was 77.38%and 30.00% respectively, which was better than the existing methods.