Search

As an important research topic, source code vulnerability detection has many problems in its traditional methods, such as high manual participation, weak detection ability of unknown vulnerabilities. Aiming at the above problems, based on the syntactic and semantic information of open source code and improved differential evolution algorithm, this paper proposed a cold start optimization method for source code vulnerability detection model. This paper used source code slicing technology, heuristic optimization algorithms and neural network models, which solved the problem that the hyperparameters couldn’t be correctly selected at the beginning of the vulnerability detection model. For the case of sample information redundancy and mixture of positive and negative sample distinctive features in the experiment, it proposed an idea of cross-exclusion of positive and negative sample distinctive features to reduce the false negative rate and false positive rate of the model. Experiments show that this method can effectively accelerate the convergence of the model, and making the model stable within 10 epochs. While improving the accuracy of the source code vulnerability detection model, its convergence speed is 2～3 times higher than other models. In the subsequent improvement experiments, the source code vulnerability detection model has improved the accuracy of each type of vulnerability by 1～3 percentage points, which fully proves the effectiveness of the improvement measures. The optimization strategies and improvement measures of this method are also applicable to other neural network classification models, and it can provide ideas for exploring new methods and models in the field of vulnerability detection. [ABSTRACT FROM AUTHOR]

The existing imaging enhancement examination process faces issues such as suboptimal contrast agent injection schemes, lack of intelligent control over injectors, and disconnected operation steps, this paper proposed a Java-based integrated intelligent recommendation and control system for injection schemes to address these problems. Through the intelligent injection scheme recommendation, it established a model for recommending optimal injection schemes to realize the automatic recommendation of the most suitable injection scheme, thereby improving the effectiveness and safety of injections. Additionally, by integrating the injection equipment with image processing algorithms, the system achieved automated control over the injection scheme, thereby enhancing accuracy and consistency during injections. Furthermore, by monitoring the image data and physiological parameters during the injection process in real time, the system integrated the existing examination process, saved and displayed the process data to improve treatment effectiveness and enhance the patient experience. Ultimately, through the coordinated application of rules, case-based reasoning and deep neural networks, the system achieved intelligent generation of contrast agent schemes. And it also realized feedback and control over the high-pressure injector through a persistent connection, which provided a good comprehensive solution to improve the intelligence level of imaging enhancement, enhance the quality of medical processes and outcomes, and facilitate mutual recognition of medical results. The system has been successfully applied at West China Hospital, which can simplify operations in imaging enhancement examinations, improve image quality, reduce patient harm and medical resource waste. [ABSTRACT FROM AUTHOR]

Developers in large software development environments rely on bug reports to complete fixes. Since reporters may use different representations to describe the same bug due to different expression habits, automated detection of duplicate bug reports can reduce development redundancy as well as maintenance costs. Recent detection of repetitive bug reports tends to use deep neural networks and considers both structured and unstructured information to generate hybrid representation features. In order to obtain the features of unstructured information of bug reports more effectively, this paper proposes a D＿BBAS(Doc2 Vec and BERT BiLSTM-Attention Similarity) method, which trains a feature extraction model based on a large-scale bug report library to generate a bug summary text representation set and a bug description text representation set that can reflect deep semantic information. Then, these two distributed representation sets compute the similarity of bug report pairs, resulting in two new similarity features. These two new features will participate in the detection of duplicate bug reports when combined with the traditional features generated based on structured information. In this paper, the effectiveness of the proposed approach is verified on the bug report repositories of well-known open-source projects Eclipse, NetBeans and Open Office, which contain more than 500, 000 bug reports. The experimental results show that compared with the representative methods, the method in this paper improves the F1 value by 1.7% on average, which proves the effectiveness of the method in this paper. [ABSTRACT FROM AUTHOR]

Aiming at the traditional ciphertext retrieval scheme in cloud environment, which generates file vectors and retrieval vectors based on statistical model, and does not consider the deep-seated semantic information of files and requests, this paper proposes a deep-seated semantic ciphertext retrieval model based on hybrid Cloud Architecture. The vector generation model is constructed through the private cloud federated learning neural network model, and the ciphertext data is stored through the public cloud. In addition, this paper proposes a secret inverted index table to store file vectors, so as to improve the efficiency of retrieval without ensuring that the retrieval information is not leaked in the retrieval process of public cloud. The analysis and experiments on real data sets show that this scheme is better than the current ciphertext retrieval schemes of the same type in terms of security and search efficiency. [ABSTRACT FROM AUTHOR]

This paper proposed a new network framework for RGB-D salient object detection (SR-Net). In order to effectively integrate the complementarity of multi-model features, this paper took the depth feature extraction as an independent branch, use the Convolutional Block Attention Module (CBAM) to enhance the depth feature, and integrate the complementary information of the enhanced depth feature and RGB feature. Then, in order to remove feature redundancy and reduce the interference of background noise on the prediction results, it proposed a sequential refining network in the up-sampling network, that is, first, the primary global features are obtained by integrating the complementarity of multi-level and multiscale features, and used the Primary Global Feature Weight Matrix Acquisition Module (PFW) which based on the channel weight to obtains the weight matrix of the primary global feature, and then uses the obtained weight matrix to refine the features of each level to suppress the interference which caused by background noise. Finally, in order to better optimize the whole network, it proposed a new loss function. The experimental results on four public datasets show that the model is superior to nine advanced methods in different model evaluation indexes, and achieves more advanced performance. [ABSTRACT FROM AUTHOR]

Recent years, the motion deblurring algorithm based on deep learning has been widely concerned, while single defocus image deblurring is rarely studied. In order to specifically solve the defocus blur problem of single image, this paper proposed a defocus deblurring algorithm based on deep recurrent neural network. Firstly, the algorithm used two cascaded residual networks to estimate the defocus map and image deblurring, respectively. After that, to ensure that the depth features of defocus map and clear images could better propagate across stages and interact within stages, the algorithm introduced LSTM ( long short-term memory) as a recurrent layer in the residual network. Finally, the whole residual network underwent several iterations and reused the network parameters during the iterative stages. To train the network, this paper produced a synthetic defocus blur image dataset, where each defocus blurred image contained a corresponding clear image and defocus map. The experimental results show that, compared with existing defocus deblurring methods, the proposed algorithm has significant advantages in both the subjective and objective image quality evaluation, and can produce sharper edges and clear details in the restoration results. On the real defocus blur dual-pixel image dataset DPD, the proposed algorithm improves the peak signal-to-noise ratio ( PSNR) and structural similarity( SSIM ) by 0. 77 dB and 5. 6%, respectively, compared with DPDNet-Single. Therefore, the proposed method can effectively deal with defocus blur in real scenes. [ABSTRACT FROM AUTHOR]

Aiming at the traditional deep neural network is difficult to extract effective features from vessel wall image segmentation, this paper proposed a vascular wall image segmentation method combining dense connection and adaptive weighted loss . Firstly, it constructed a densely connected segmentation network to learn more boundary and contour representations to promote feature reuse and fusion, and then designed an improved adaptive weight loss and boundary compactness loss constraint training network. It used the adaptive weighted loss to automatically adjust the loss ratio of different regions to guide the network to learn in the best direction. At the same time, it introduced the boundary compactness loss constraint to make full use of the boundary information and improve the segmentation accuracy of the blood vessel wall image. Finally, this paper performed validation experiments using the MERGE blood vessel wall dataset containing 2 544 MRI. The results show that the proposed improved method can effectively extract the feature information of the vessel wall image, segmentation Dice reaches 93 . 65 % and 95. 81 % in the segmentation of the lumen contour and the outer wall contour, respectively. The ablation experiment also fully proves the effectiveness of the various module, which can better achieve high-precision image segmentation. [ABSTRACT FROM AUTHOR]

This paper characterized the dietary characteristics of the catering consumer group from sales data of Chinese optional restaurants and predicted the dish sales of the restaurant. Due to the huge changes in the daily dishes served in Chinese optional restaurants, traditional time series forecasting methods can hardly solve the problem. This paper proposed a forecas ting model based on decision theory. It used the MACBETII method and the expected utility theory to transform the restaurant 's dishes into a unified utility value, which served as the basis for the dining group to choose dishes, used the decision weight to represent the dietary characteristics of the consumer group, and then designed a decision model for catering consumers based on the behavioral decision theory. This paper used the running data of Chinese optional restaurants for training. The crossentropy of the prediction model on the training set is 0. 082 and the cross -entropy on the test set is 0 . 086. Compared with the prediction model based on neural network, the model is more accurate and has good interpretability. Therefore, the model can accurately predict the sales volume of Chinese optional restaurants under a given dish supply, thereby supporting the efficient operation of Chinese optional restaurants . [ABSTRACT FROM AUTHOR]

In the absence of data on user interaction complementary items, this paper integrated user preferences for items into complementary item recommendation that only considered item relationships to improve the performance of the recommendation model. This paper proposed a method for recommending complementary items based on a knowledge graph, inferring the complementary items of user interaction in the user's historical interaction items, extracting the user's preference for complementary items based on the knowledge graph, and using the difference between image and text learning items complementary relationship, and finally realized the common learning based on neural network. Compared with the sub-optimal baseline method on the Amazon data set, the proposed method has a 7% increase in ACC and a 3% increase in precision, which shows that the performance of the proposed method is excellent. The method in this paper jointly learns the complementary relationship between the user's preference for items and the items, and improves the recommendation performance. [ABSTRACT FROM AUTHOR]

Recently, graph neural networks have received widespread attention due to their rich representation and reasoning capabilities. To best knowledge, current research mainly focuses on amending the convolutional strategy and network structure for higher performance, so the performance will be inevitably constrained by the limitations of the single model. Inspired by the idea of ​​ensemble learning, this paper innovatively proposed an ensemble learning framework for graph neural network (EL-GNN).Unlike regular text and images, graph data not only had features but also had rich topology information. Therefore, EL -GNN additionally supplemented the structure information during the ensemble stage rather than merely integrating the prediction results of independent classifiers. Besides, this paper further revised the ensemble strategy through reconstructing a feature-level similarity graph for subsequent assembling, which balances the feature and structure information on the basis of the assumptions of those nodes with the similar feature or easy reachability of high probability to share the same labels .The comprehensive experiments indicate that the proposed ensemble strategy achieves an impressive performance and EL-GNN is superior to other off-the-shelf models on the node classification task. The comprehensive experiments indicate that the proposed ensemble strategy achieves an impressive performance and EL-GNN is superior to other off-the-shelf models on the node classification task. The comprehensive experiments indicate that the proposed ensemble strategy achieves an impressive performance and EL-GNN is superior to other off-the-shelf models on the node classification task. [ABSTRACT FROM AUTHOR]

Most of the event detection methods based on dependency parsing only focus on the single-hop connection of dependency syntax structure. It ignores the multi-hop connection between words, resulting in the lack of semantic information between trigger words and some related entities, which affects the efficiency of event detection. Therefore, this paper proposed an event detection model integrating multi-hop relation labels and dependency syntactic structure information to improve the recognition performance of event trigger words. This paper designed a new dependency syntax "multi hop tree" and multi hop relation label search algorithm to enhance the event-representation ability of core vocabulary. Furthermore, it used graph attention network to aggregate the multi-level word representation, which improved the event detection performance. Results on ACE2005 dataset show that this event detection method improves the performance of the benchmark model by nearly 2%. [ABSTRACT FROM AUTHOR]

The method of predicting the properties of compounds based on machine learning plays an important role in the virtual screening of materials discovery. Existing methods of traditional machine learning models have to manually extract features, which have problems such as difficulty in extract features and difficulty in processing Simplified Molecular Input Line Entry System (SMILES) codes. In order to solve these problems, this paper proposes an end-to-end graph neural network to predict the high temperature performance of resin materials. First, the SMILES code of the resin material is represented as a graph, where the vertices are atoms and the edges are chemical bonds. Then, the vector representation of the molecule is obtained by constructing the graph neural network of the molecular graph. Finally, a full connection neural network contains information such as molecular vector, environment, and conditions to predict the maximum temperature by regression model. Experiments of the resin material data set show that the end-to-end graph neural network model proposed in this paper has more than doubled the accuracy of model prediction compared with the traditional machine learning models. [ABSTRACT FROM AUTHOR]

Due to factors such as illumination changes，camera jitter and dynamic background，existing foreground subtraction algorithms cannot achieve good segmentation results in complex scenes. To solve this kind of problems，this paper proposes a subtraction algorithm based on a fully convolutional siamese neural network，which can accurately segment the foreground with only two arbitrary frames. Specifically，the input two images are divided into the base image and the image to be segmented. The algorithm uses the fully- convolutional siamese network to get the similarity metric map of input frames. The similarity metric map contains information about changes in pixels of the image to be segmented relative to the base image. Then， the similarity metric map is fused with the image to be segmented，and the encoder-decoder network is used to achieve end-to-end foreground subtraction results. The paper evaluates the proposed algorithm on the CDnet2014 dataset to prove its effectiveness. [ABSTRACT FROM AUTHOR]

Single object tracking( SOT) is a technique to analyze the motion status and provide localization of the single target by using target appearance and context information in video. It has promising applications in intelligent surveillance, intelligent interaction, navigation and guidance, etc. However, problems such as occlusion, background interference and appearance variation led to slow progress in practical applications. With the rapid development of deep learning in recent years, the study of using deep learning techniques to optimize SOT algorithm has become one of the hot spots in computer vision. Around the SOT algorithm based on deep learning, this paper respectively provided outline and analysis of each of the six aspects of correlation filter, Siamese networks, meta-learning, attention, recurrent neural networks and generative adversarial networks according to the core algorithms, after analyzing the basic principles of SOT. In addition, this paper summarized the current state of research and proposed the development trend and optimization ideas of the algorithms. [ABSTRACT FROM AUTHOR]

In the large data environment, there are many problems in the parallel deep convolution neural network（DCNN） algorithm, such as excessive data redundancy, slow convolution layer operation and poor convergence of loss function. This paper proposed a parallel deep convolution neural network optimization algorithm based on the Im2 col method. First, the algorithm proposed a parallel feature extraction strategy based on Marr-Hildreth operator to extract target features from data as input of convolution neural network, which can effectively avoid the problem of excessive data redundancy. Secondly, the algorithm designed a parallel model training strategy based on the Im2 col method. The redundant convolution kernel is removed by designing the Mahalanobis distance center value, and the convolution layer operation speed is improved by combining the MapReduce and Im2 col methods. Finally, the algorithm proposed an improved small-batch gradient descent strategy, which eliminates the effect of abnormal data on the batch gradient and solves the problem of poor convergence of the loss function. The experimental results show that IA-PDCNNOA algorithm performs well in deep convolution neural network calculation under large data environment and is suitable for parallel DCNN model training of large datasets. [ABSTRACT FROM AUTHOR]

Accurate identification of rice pests is of great significance for timely protection and management of rice. However, the rice pests are similar with the background in the color and texture, and the morphology of the pests varies greatly during different growth stages. Therefore, it is difficult to accurately identify the rice pests in natural scenes. In this paper, the Light Weight Residual Network (LW-ResNet) composed of feature extraction, global optimization and local optimization modules was designed to improve the ability to identify rice pests in natural scene images. Firstly, in order to reduce the influence of complex background and enhance the feature extraction and expression capabilities of the residual network, the residual block is improved to constitute the feature extraction module. The improved residual block increases the number of convolutional layers and branches of the original residual block, which can effectively extract the deep global features of rice pest images. Secondly, the deep global features are further optimized through the convolutional layers in the global optimization module. Finally, in order to obtain the local discriminative characteristics of rice pest images to distinguish the morphological differences between similar pests, the lightweight attention sub-module constitutes the local optimization module. The light weight attention sub-module uses depth separable convolution to reduce the redundancy of channel features and realize the aggregation of different channel characteristics, so it can highlight the local key features of rice pests. Because the improvement of the residual block in the feature extraction module reduces the number of residual blocks, and the use of deep separable convolution in the attention sub-module and the channel-based global average pooling and global maximum pooling encoding operations reduce floating point operations, the LW-ResNet network has achieved lighter weight. In the HSV space, Gamma transform is used to preprocess the v component of rice pest images and then proceed to the data expansion. After the expansion, there are 4 380 images in the training set and 492 images in the test set. In order to verify the rationality and effectiveness of the method in this paper, in the training phase, the cosine learning rate decay strategy was used to train the network model. By analyzing the number of the improved residual blocks in the feature extraction module, the lightweight attention sub-module in the local optimization module, and the global optimization module, the rationality of the method in this paper was verified. In the testing phase, the LW-ResNet network model achieves a identification accuracy of 92.5% on the test data set of 13 types of rice pest images. The identification accuracy of the LW-ResNet network model is higher than that of classic convolutional neural network models such as VGG16, ResNet, and AlexNet. The parameter amount of the LW-ResNet model is 1.62×106, and the amount of floating-point operations is 0.34×109. The number of parameters and floating-point operations of the LW-ResNet model are both lower than those of MobileNetV3, which verified the effectiveness of the method in this paper. The LW-ResNet network model has achieved light weight and a good identification effect, so it can be used for rice pest identification on the mobile terminal. [ABSTRACT FROM AUTHOR]

To solve the problem that the influencing factors of the airport bus operation process are complicated and the operation time is difficult to predict, this paper established an airport bus operation time prediction model based on the subspace identification algorithm. Firstly, based on the multi-source big data generated during the operation process, considering the number of passengers, departure interval, road congestion and other factors at different times, it established the state space model of the airport bus operation. Then it extracted the characteristic variables suitable for describing the operation process of the airport bus as the input and output of the model,and solved the model by subspace identification method. Finally,this paper took an actual operating route of the capital airport bus as a case for simulation analysis. The calculation results show that the mean absolute percentage error and mean square error of the model are 2. 25% and 4. 77, respectively, which are better than the traditional BP neural network prediction model and the least square identification model. The model has good prediction accuracy and has certain practical application value. [ABSTRACT FROM AUTHOR]

This paper proposed the CoBERT-BiGRU model to overcome the shortage that the existing models could not effectively handle multiple turns of dialogue history. The paper marked the conversation history in the passage, and input differences and questions into the CoBERT model to obtain a vectorized representation of multiple sequences. Through history attention network, it combined multiple results into one vectorized representation of a sequence. Then it put fusion results into BiGRU and predicted the answer and dialogue behavior. The experimental results on the real dataset show that the CoBERT-BiGRU model can effectively handle multiple turns of dialogue history. Compared with the baseline model and others that have been published on this dataset, the HEQ-Q, HEQ-D and F1 values have improved. [ABSTRACT FROM AUTHOR]

It is a very meaningful direction to study data stream ensemble classification based on the condition of supervised or semi-supervised learning. This paper introduced three aspects including base classifiers, key technologies and ensemble strategies. The base classifiers mainly introduced decision trees, neural networks, support vector machines, etc. The key technologies were introduced from incremental and online aspects, and the ensemble strategies mainly introduced boosting, stacking, etc. This paper summarized and analyzed the advantages and disadvantages of different ensemble methods, comparison algorithms and experimental data sets. Finally, it gave the further research directions, including the handling of concept drift based on supervised and semi-supervised learning, the study of homogeneous integration and heterogeneous integration, and the classification of data stream ensemble based on unsupervised learning. [ABSTRACT FROM AUTHOR]

This paper constructed a new TerraSAR-X dataset named GDUT-Nansha and then proposed a new lightweight semantic segmentation algorithm using synthetic aperture radar images（ SAR）. It is difficult to apply traditional deep learning models to SAR datasets with small volume due to the tremendous amount of parameters. A revision scheme is, therefore, put forward to deal with such problem. This paper accomplished an improved lightweight convolutional neural network model named revised weighted loss ENet（ RWL-ENet） for SAR images. The current study introduced a weighted loss function to solve the problem of imbalance of training datasets. Compared with other classical convolution neural network models, the efficiency and robustness of the new dataset were validated. Meanwhile, RWL-ENet model attained 0. 884,0. 804, and 0. 645 in terms of three quantitative metrics, including pixel accuracy（ PA）, mean pixel accuracy（ m PA）, and mean intersection over union（ m IoU）. In addition, the parameters of this new proposed model are much less than other classic network models. [ABSTRACT FROM AUTHOR]

In order to improve the accuracy of text classification and solve the problem of insufficient utilization of node features by text graph convolution neural network, this paper proposes a new text classification model, which integrates the advantages of text graph convolution and Stacking integrated learning method. The model first learns the global expression of documents and words and the grammatical structure information of documents through text graph convolution neural network, and then secondary learns the features extracted by text graph convolution through integrated learning, so as to make up for the insufficient utilization of text graph convolution node features, and improve the accuracy of single label text classification and the generalization ability of the whole model. In order to reduce the time consumption of ensemble learning, the fusion algorithm removes the k-fold cross verification mechanism in ensemble learning. The fusion algorithm realizes the correlation between text graph convolution and stacking integrated learning method. The classification effect on R8, R52, Mr, Ohsumed, 20 ng and other data sets is improved by more than 1.5%, 2.5%, 11%, 12% and 7% respectively compared with the traditional classification model. This method performs well in the comparison of classification algorithms in the same field. [ABSTRACT FROM AUTHOR]

Graph convolutional neural networks can distill the effective information of graph data through graph convolution. However, the graph convolutional neural network show vulnerability to adversarial attack, which leads to the degradation of model performance. Adversarial training can be used to improve the robustness of neural networks. However, since the structure and node features of graphs are usually discrete, it is impossible to directly construct adversarial examples based on gradients. Therefore, distilling feature of graph data in the embedding space of models as adversarial examples can reduce the complexity of adversarial training. By using the idea of the idea of ensemble learning, this paper innovatively proposes an adversarial training method based on non-robust features distillation for graph convolution network, VDERG. The method constructed two graph convolution neural networks as sub models from the two types of features of topology and node attributes respectively. Sub models distilled non-robust features through embedding space and used these features to implement adversarial training. Finally, the method combined the embedding given by the two sub models as the nodes’ vectors. Experimental results show that the adversarial training strategy improves the accuracy of graph convolution neural networks in clean data by 0.8% on average, and improves the accuracy by 6.91% at most under adversarial attack. [ABSTRACT FROM AUTHOR]

Objectives: Mars is the main target object for deep space exploration. Mars rovers, or roving probes, are important tools for surface exploration and scientific research on Mars. For the growing amount of remote sensing data collected by Mars rovers, there is an urgent need for a method that can intelligently detect novel targets of scientific value from the massive amount of images, reduce the time cost of detection planning, and provide information for subsequent scientific analysis. The traditional novel detection methods mostly include distance-based metrics and image-based reconstruction methods, distance-based metrics calculate novel scores pixel by pixel without considering spatial contextual information, and image-based re‐ construction methods focus on reconstruction of typical landscape backgrounds, and novelty is manifested by image reconstruction errors, which is not effective in extracting small novel targets such as boreholes and dust removal points. Methods: To address the above problems of traditional novel detection methods in Mars rover novel target detection, this paper proposes an improved Mars rover multispectral image depth novel target detection method, uses full convolutional self-coding neural network to extract deep features for typical landscape reconstruction, and joints Mahalanobis distance for novel target and typical landscape background separation, fully exploits the spatial and spectral dimensional features to improve the accuracy of Mars rover novel target detection results. Results: The experiments use the multispectral image dataset of Curiosity rover released by NASA ( National Aeronautics and Space Administration ), and the proposed convolution auto-encoder combined Mahalanobis distance method(CAE-M) is compared with Reed-Xiaoli detector, principal component analysis, convolution auto-encoder convolution, and generative adversarial networks on the surface of Gale crater. The results show that CAE-M outperforms previous detection methods in terms of detection accuracy and visual interpretation, and has a balanced and stable performance in different classes of novel target detection. Conclusion: The proposed CAE-M method comprehensively utilizes spatial and spectral information of multispectral images, which can help Mars rover exploration missions to quickly and intelligently screen and sort novel data with scientific value in massive data, save the time and cost of route planning, improve scientific returns. [ABSTRACT FROM AUTHOR]

Objectives: The scale of buildings and their distribution is key indicators to measure the economic and social development of a region. Therefore, it is significant to study the extraction of buildings based on remote sensing images. Existing neural network methods still have shortcomings in the completeness of building extraction and the accuracy of building edges. To solve the above problems, this paper proposes a multi-level feature fusion network (MFFNet) based on high-resolution images. Methods: Firstly, we use edge detection operators to improve the ability of the network to recognize the boundaries of buildings. Secondly, we use a multi-path convolution fusion module to extract building features from multiple dimensions, and introduce a large receptive field convolution module to break through feature extraction. The process is limited by the size of the receptive field. After fusing the extracted features, the convolutional attention module is used to compress them, and the global features are further mined by pyramid pooling, so as to achieve high-precision extraction of buildings. Results: The current mainstream UNet, pyramid scene parsing network (PSPNet),multi attending path neural network (MAPNet) and multiscale-feature fusion deep neural networks with dilated convolution (MDNNet) are used as the comparison methods, and we use Wuhan University Aerial Image Dataset, Satellite Dataset II (East Asia) and Inria Aerial Image Dataset as experimental data for testing. Compared with the other four methods, MFFNet improves intersection over union, precision, recall, F1-score and mean average precision by 1.53%, 2.65%, 2.41%, 3.32% and 1.19% on average, achieves a better effect. Conclusions: MFFNet not only accurately captures the detail features of buildings, but also strengthens the extraction and utilization of global features. It has better ex‐ traction effect on large buildings and buildings in complex environment. [ABSTRACT FROM AUTHOR]

In view of the data dependence that is the major drawback of the supervised deep neural network ( DNN )-based denoising models, this paper proposed an unsupervised deep image generation ( UDIG) denoising model. Firstly, it utilized a noise level estimation ( NLE) algorithm to measure the noise level of a given noisy image, the estimated noise level and the noisy image fed into the state-of-the-art denoising model ( i. e., fast and flexible denoising convolutional neural network, FFDNet) to obtain a preliminary denoised image as the input to the UDIG model. Secondly, it chose the encoder-decoder architecture as backbone network, while using the sum of the mean square error among the output image of UDIG model ( i. e . generated image), and the preliminary denoised image, and the given noisy image to define mixed loss function. Then, the regular stochastic gradient descent ( SGD) algorithm optimized the loss function, adjusted the hyper-parameters of the UDIG model to generating a series of generated images. Finally, when the standard deviation of the residual image between the noisy image and generated image approximates the noise level measured by the NLE algorithm, the network iterative process was terminated adaptively, ensuring the image quality of the gene-rated image ( treated as the denoised image) . Extensive experiments show that, the proposed UDIG denoising model has a better performance than other state-of-the-art counterparts with regard to denoising effect. [ABSTRACT FROM AUTHOR]

The application of deep learning to the task of pedestrian re-identification has made great progress. However, deep learning has exposed some security issues in pedestrian re-identification model applications due to the robustness of deep neural networks that are vulnerable to attacks by adversarial samples. To address this problem, this paper proposed a defense method DSN for senseless noise attacks. First, it used grayscale patch images of RGB images to enhance the data during the training process, so as to improve the recognition ability of the pedestrian re-identification model. Secondly, it adopted a defense structure combining the inside and outside of the model and used a new noise reduction network to de -noise the input noisy images, so that the pedestrian re-identification model had higher recognition accuracy and defense against senseless noise attacks. Simulating the senseless noise attack and defense on the marketl501 dataset, the experimental results show that the method improves the mAP identification accuracy from 2.6% to 82.6% and rank-1 accuracy from 0. 8% to 83.5% . In addition, the effectiveness of each module in the method to defend against the senseless noise attack is demonstrated by ablation experiments. [ABSTRACT FROM AUTHOR]

The distributed denial of service ( DDoS) attack detection method in traditional software defined network ( SDN ) requires frequent communication between the control plane and the data plane, which will lead to significant overhead and delay, and the current programmable data plane can not implement complex detection algorithms, so it is difficult to ensure high detection efficiency. To solve the above problems, this paper proposed a DDoS attack detection method based on programming protocol-independent packet processors ( P4 ) programmable data plane. First of all, the method used the improved information entropy based on P4 as initial detection to determine whether suspicious traffic occurred . Then, it took advantage of the microsecond time required for feature extraction by P4 to extract the six-tuple features of suspicious traffic, and imported them into the data standardization-deep neural network ( DS-DNN ) reinspection module to determine whether they were DDoS attack traffic . Finally, it tested the evaluation indicators of the method in the real environment. The experimental results show that this method can better detect DDoS attacks in SDN environments, ensure high detection rate and accuracy, effectively reduce the false alarm rate, and shorten the detection time to millisecond level. [ABSTRACT FROM AUTHOR]

Mobile edge computing is a method to solve the computing-intensive task requirements of robots. Traditional algorithms are based on intelligent algorithms or convex optimization methods, and the iteration time is long. Deep reinforcement learning can be solved in a single forward pass, but only for a fixed number of robots. Through the analysis and research of deep reinforcement learning, this paper performed input regularization before the input layer in the deep reinforcement learning neural network, and added a convolution layer after the output layer so that the network could adaptively meet the unloading requirements of the number of dynamic mobile robots. Finally, it carried out the simulation experiments to verify the effectiveness and feasibility of proposed algorithm, compared with the adaptive gen etic algorithm and reinforcement learning . [ABSTRACT FROM AUTHOR]

In order to understand the latest research progress of TransE based representation learning methods in real time,this paper classified TransE based representation learning methods into four types:the method based on complex relationship,the method based on relationship path,the method based on image information,and the method based on other aspects.Then,this paper analyzed the design ideas,advantages and disadvantages of each method.At the same time,it compared and summarized the common data sets and evaluation indexes of the TransE based representation learning method,as well as the performance of various TransE based representation learning algorithms in the experiment.From the research results,PaSKoGE,NTransGH,TCE and TransD method perform the best in link prediction and triple classification tasks,which are worth promoting and further expanding,and can be further improved in path specific embedding,two-layer neural network,triple context and dynamic mapping matrix construction. [ABSTRACT FROM AUTHOR]

This paper proposed a data-driven Job-Shop scheduling algorithm. It derived the training samples from some benchmark instances and actual production data. It constructed the feature data of the samples using the feature function and then normalized. It constituted the tag data by the mapping relations between the scheduling tasks and the corresponding scheduling rules. This paper embedded a pointer network into the main framework of the LSTM recurrent neural network model so that the workpiece with the highest probability in the current sequence would be passed to the buffer at first, which improved the training speed and training quality of the neural network. The result of an experiment shows that the proposed model is effective in solving Job-Shop scheduling problem after training. This study provides a new idea for solving the Job-Shop scheduling problem. [ABSTRACT FROM AUTHOR]

Most existing positive-unlabeled (PU) graph learning methods exact only node representations to infer node labels in dependently. This paper proposed a collective inference based method that exploited the correlations among nodes to assist in classification of unlabeled nodes. Firstly, it used the personalized PageRank algorithm to approximate correlation degrees between each node and observed positive nodes as a whole (positive correlation degree, PCD for short). Then, it built a local classifier to predict labels of unknown nodes by combining PCD with node representations that were captured via a graph neural network (CNN), and constructed a relational classifier to collectively up date labels of unknown nodes by combining PCD with local label dependency of nodes that were exacted via another CNN. Furtherly, it exploited the Markov CNN (CMNN ) frame-work to train these two classifiers, alternately and iteratively, to form a multi-hop collective inference procedure. Besides, it proposed a mixed non-negative unbiased risk estimator for the two classifiers to estimate empirical loss with only positive and unlabeled nodes. Finally, either of them could predict labels of unknown nodes. Experimental results on real-life datasets show that the proposed method remarkably out performs the state-of-the-art approaches in identifying both single-class target concept and multiple-classes-merged target concept, and performs quite robust against to error of the prior of positive nodes. [ABSTRACT FROM AUTHOR]

In the field of tumor diagnosis, the AI-assisted diagnosis system can accurately distinguish tumor attributes and malignant tumor stages, thereby prolonging the survival time of patients. Taking breast tumor as an example, this paper proposes an artificial intelligence-assisted diagnosis model based on supervised learning to solve the problem of overfitting caused by the large amount of data in the feature extraction process. When extracting features, effective feature dimensionality reduction is achieved by introducing hierarchical clustering analysis, and the classified feature data is used as the feature input of the artificial neural network model, so as to realize the effective training of the classifier. The experimental results show that the accuracy and AUC value of the proposed algorithm are improved compared with the control algorithm, indicating that the model can not only solve the problem of overfitting caused by the description of massive feature regions, but also enhance the generalization of the artificial intelligence-aided diagnosis system. The ability to achieve high-precision differentiation of mammography-targeted breast tumors. [ABSTRACT FROM AUTHOR]

Under the background of the rapid development of feed industry, the financial risk early warning of feed listed companies is of great significance to improve the anti risk ability and prevent financial crisis. The paper takes 23 listed companies from 2019 to 2020 as research sample, selects 13 financial indicators and two non-financial indicators, extracts six common factor indicators with the help of factor analysis, and carries out early warning on financial risk of feed listed companies through the training and test of BP neural network model. The results show that return on net assets, current ratio, asset turnover rate, fixed asset turnover rate, operating income growth rate, and net profit cash content ratio are important factors for evaluating the financial risk warning level of listed feed companies. The research indicates that after model training and testing, the overall accuracy of BP neural network model reaches 81.8%, which is more accurate for financial risk warning results, and it can provide a reference for the decision-making level of feed listed companies to prevent financial risks. [ABSTRACT FROM AUTHOR]

Object tracking is an important topic in computer visual directions, where scale changes, deformation and rotation are difficult to resolve in the field. For the challenging problems faced in the above track, based on existing Siamese network algorithms, this paper proposed multi-layer feature fusion and parallel self-attention Siamese networks (MPSiamRPN) for visual tracking. Firstly, MPSiamRPN used the modified ResNet50 to extract features from the template image and the search image. In order to deal with the loss of some features caused by the deep network, it proposed a multi-layer feature fusion module to fuse the features of the last three layers of ResNet. Secondly, it introduced the parallel self-attention module, which was composed of channel self-attention and spatial self-attention. The channel self-attention could selectively emphasize the beneficial channel features for tracking, and the spatial self-attention could learn the rich spatial information of the target. Finally, it proposed the region proposal network ( RPN) to perform classification and regression operations to determine the location and shape of the target. Experiments show that the MPSiamRPN can achieve competitive results on OTBlOO and VOT2018 test datasets. [ABSTRACT FROM AUTHOR]

In order to improve the performance of frame disambiguation model, this paper used a neural network model to learn frame representation based on sentences in corpus different from the traditional classification algorithm extracting features manually, and employed the learned frame representation on the frame disambiguation task, which significantly improved the performance of the task. Making full use of the CFN example sentence database and being based on the hinge-loss neural network, the algorithm learnt the frame representation that could distinguish the correct frame from the error frame in the largest degree. This paper also used the WSABIE algorithm to learn the representation vector of the context of the target word, and finally used the cosine distance between the context representation vector and the frame representation vector to make a decision for the task. Experiment performed three sets of two-fold cross-validation (3 x 2 BCV) on 88 ambiguous words in CFN, and the best accuracy of frame disambiguation reach to 72. 52%. The t-test results show that the performance of the proposed method is significantly higher than other frame disambiguation methods. [ABSTRACT FROM AUTHOR]

Aiming at the knowledge overload in the process of online learning and the problems of data sparsity and cold start in the traditional recommendation algorithm, this paper proposed an improved learning resource recommendation algorithm based on MLP. The algorithm used the advantage of MLP in nonlinear data processing, converted the prediction mode of vector multiplication of learner characteristics and learning resource characteristics into the input mode of MLP, and improved the DN-CBR neural network recommendation model. In order to verify the effectiveness of the model, this paper built a dataset with the online learning platform data of iCourse as the sample. Compared with DN-CBR model, the improved model improves the normalized cumulative loss gain and hit rate by 1. 2 % and 3 % respectively on the dataset, which effectively improves the recommendation performance of the model. [ABSTRACT FROM AUTHOR]

How to learn and identify new categories from a small number of training samples is a challenging problem for deep neural networks. For how to solve the problem of few-shot learning, this paper comprehensively summarized the existing few-shot learning methods based on deep neural networks, which covered various aspects such as models used in the methods, datasets and evaluation results. Specifically, for the few-shot learning method based on deep neural network, this paper divided it into four categories, named data enhancement method, migration learning method, metric learning method and meta-learning method. For each category, this paper further divided it into sub-categories and conducted a series of comparisons between each category and method to show the pros and cons of the various methods and their respective characteristics. Finally, the paper highlighted the limitations of existing methods and pointed to future research directions in the field of few-shot learning. [ABSTRACT FROM AUTHOR]

Aiming at the problems of time-consuming and high error rate in handwriting test scores statistics a handwriting test paper recognition method based on improved convolution neural network algorithm was proposed. In order to simplify the classification of the score» the score column of the handwritten test paper was divided to obtain ten types of numbers from 0 to 9. In order to improve the efficiency of handwritten score recognition, a classification and recognition algorithm of convolution neural network and Bayesian was proposed. The constructed convolution neural network model was used to extract the characteristics of handwritten digits. The PCA algorithm was used to reduce the dimensionality of the features.The hayes classifier was used to distinguish ten kinds of numbers from 0 to 9. the auuracy and efficiency of the algorithm were verified in the MNIST database.The paper score summation model was established and automatic summation is performed after recognition. The experimental results show that for the recognition uf the handwritten scores of 1 188 lest papers in 3 courses♦ the algorithm in this paper has a recognition rate of 98. 23% compared with other algorithms, the average recognition time per test paper is 7. 5 s, and verified its practicality. [ABSTRACT FROM AUTHOR]

Multi-level in-depth analysis of collective activity is an important issue to be solved in the field of activity recognition. Based on the research of deep neural network, this paper proposed a progressive hierarchical analysis model for activity recognition. Using the modulating network based on transfer learning, it detected multi-person with temporal consistency detection in the crowd. Through integrating spatio-temporal feature learning, it recognized the individual actions in the crowd with unconstrained action duration. Through integrating the individual action category, interaction context and scene context, it re-cognized the crowd activity steady and effectively. A large amount experiments on the benchmark data sets demonstrate that, compared with the current approaches, The proposed model achieves better performance on collective activity analysis and recognition. [ABSTRACT FROM AUTHOR]

In the industrial process, the existing fault diagnosis methods based on deep neural network are a challenging pro-blem due to complicated network structure design and time-consuming parameter optimization. To achieve the problem, this paper proposed an automatic fault diagnosis method（AutoFD） based on neural architecture search. The method used the AutoFD neural architecture search（NAS） algorithm to automatically complete the network structure design of the convolutional neural network and the optimization of the network parameters. On this basis, some new channels were firstly generated by operating the original data. Performance prediction was then used to speed up the process of acquiring channel adaptive sorting, and the optimal number of convolutional channels was quickly selected according to the adaptive sorting of the channels. Finally, through the optimal convolution channels, the best-performing multi-channel convolutional neural network model for automatic fault diagnosis in chemical processes could be searched. It was applied to the Tennessee Eastman（TE） industrial process and numerical system for fault diagnosis to verify the proposed method. The results show that the AutoFD method can automatically design the network structure and optimize the parameters of the multi-channel convolutional neural network, which has excellent performance in fault diagnosis. [ABSTRACT FROM AUTHOR]

Traffic condition prediction is an important part of intelligent transportation system, and traffic flow is the most direct embodiment of traffic condition. Therefore, traffic flow prediction has important application value. On the one hand, the roads in the city have spatial topological properties, on the other hand, the traffic flow changes dynamically with time. Therefore, the key to the prediction of traffic flow is to model the time and space dependence in the data. In view of this characteristic, this paper used neural network model and attention mechanism to explore the temporal and spatial dependence relationship in traffic flow data, and proposed a traffic flow prediction model based on time map attention. In terms of spatial dependence, it used a learning algorithm combining graph convolution network and attention to assign different weights to nodes with different influence degrees, and added node adaptive learning to effectively extract spatial features. In terms of time dependence, it used the temporal convolution network to extract the temporal features, and expanded the sensing domain by expanding convolution, so as to capture the features of longer time series data. A spatial-temporal network layer was composed of graph attention network and time convolution network, which was finally connected to the output layer to output the prediction results. The model used the combination of graph convolution neural network and attention mechanism to extract spatial features, fully considered the spatial relationship between roads, and used temporal convolution network to capture temporal features. After experiments on two real datasets, it is found that it has good performance in the next 15, 30 and 60 minutes, and the results are better than the existing baselines. [ABSTRACT FROM AUTHOR]

This paper proposed an optimization model of deep-knowledge tracking（KTSA-BiGRU） based on self-attention mechanism and bidirectional GRU neural networks owing to the existing deep-knowledge tracking models with weak capture of complex relationships between input exercises and inability to effectively handle long-sequence input data. Firstly, it mapped the learner’s historical learning interaction sequence data to the real value vector sequence. Then, it trained the bidirectional GRU neural network as input to model the learner’s learning process, and finally used, the self-attention mechanism to calculate the probability of the learner correctly answering the next question based on the hidden vectors of the bidirectional GRU neural network output and the attention weight. The performance analysis on the three public datasets can improve the prediction accuracy of deep knowledge tracking. [ABSTRACT FROM AUTHOR]

Recently, the end-to-end automatic speech recognition system based on Transformer has been widely popularized, but the multi-head self-attention mechanism in Transformer is not sensitive to the position information of the input sequence, and its flexible alignment has poor generalization performance in the face of noisy speech. To solve the above problems, firstly, this paper proposed to use TCN to enhance the capture of location information by neural network model. Secondly, on the basis of the above, it proposed the TCN-Transformer- CTC model by fusing CTC. Without using any language model, the experimental results on AISHELL-1,an open source speech database of Mandarin Chinese, show that the word error rate of TCN-Transformer-CTC is relatively reduced by 10.91% compared with Transformer, and the final word error rate of the model is reduced to 5.31% which verifies that the proposed model has a certain degree of advancement. [ABSTRACT FROM AUTHOR]

Using spikes to represent and convey information, SNN is more biologically plausible than traditional artificial neural networks. However, a classical SNN has limited feature extraction ability due to the shallow network structure, leading to inferior classification performance to CNN especially on multi-class classification tasks such as object categorization. Inspired by the powerful convolutional structure of CNN, this paper proposed a SCSNN. By exploiting convolutional structures and dynamic impulse triggered property of biological neurons, the proposed SCSNN organized integrate-and-fire models in a convolutional fashion, and trained by a new surrogate gradient back-propagation algorithm directly. It validated the proposed SCSNN on the MNIST and Fashion-MNIST dataset respectively, obtaining superior performance to state-or-the-art SNN on both datasets. The classification accuracy reaches 99. 62% on the MNIST dataset, and 93. 52% on the Fashion-MNIST dataset, which verifies the effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

Aiming at the problem of increasingly rich multilingual texts and lack of large-scale labeled parallel corpora on the Internet, in order to mine the relevance between languages from multilingual information sources and expand the knowledge map, this paper proposed a cross language relations hip extraction method based on attention transfer. Firstly, according to the actual parallel corpus between languages, it classified the cross language parallel corpus mapping, and for the low resource language pairs lacking seed dictionaries, it proposed a neural network translation model to obtain the target language data set and save the corresponding attention weight relationship between multiple languages, and then it extracted the entity relationship feature of training data by using BERT end-to-end joint extraction model. Finally, it used the reverse transferred attention to extract the enhanced relationship. Experiments show that the relationship extraction effect of this model is better than other models in accuracy and regression, and also shows better performance in the absence of bilingual dictionary. [ABSTRACT FROM AUTHOR]

Implicit aspect extraction is important for improving the accuracy of fine-grained sentiment analysis. However, existing implicit aspect extraction techniques do not have strong generalization ability when dealing with large-scale data. To address the problem, this paper proposed an implicit aspect extraction model combining dependency syntactic parsing and interactive attention mechanism. First, the model generated the initial representation of the text by the pre-trained language model BERT. Then, it passed the initial representation to the self-attention layer guided by the dependency syntactic parsing. Due to the interactive attention mechanism, the model further extracted the results of the above two processes. Finally it used a classifier to determine the implicit aspect of the sentence. Compared with baseline BERT and other deep neural network models, the proposed model has achieved higher F1 and AUC on the enhanced SemEval implicit aspect dataset, which proves the effectiveness of the model. [ABSTRACT FROM AUTHOR]

This paper aimed to summarize the current multimodal deep learning, found common problems in the implementation of multimodal deep learning under different multimodal and learning objectives, as well as made common problems classify and described methods for solving various problems at the early development of multimodal deep learning . Specifically, this paper summarized the current multimodal deep learning that studied on natural language, visual, auditory, and considered the re search direction such as language translation, event detection, information description, emotion recognition, voice recognition and synthesis,and multimedia retrieval and so on, which further concluded that there were four types of common problems :multimodal r epresentation, multimodal interpretation, multimodal fusion and multimodal alignment. Meanwhile, this paper sub-categorized and discussed each common multimodal learning problem, and listed the neural network models generated for solving the problems. Finally, it introduced some actual multimodal system, listed baseline datasets and evaluation criteria used in multimodal deep learning, and prospected the development directions for future research [ABSTRACT FROM AUTHOR]

Aiming at the shortcomings of the traditional image recognition, including low accuracy and relying on experienced staff, a peanut kernel quality detection method based on machine vision and adaptive convolutional neural network (DAL-ACNN) is proposed in this paper. An image acquisition system was built to obtain the color images of peanuts which included normal peanuts, broken peanuts, shriveled peanuts and mildew peanuts, and a black background was selected in the system. Then, a convolutional neural network (CNN) based on Alexnet was selected to automatically extract and detect the peanut defects. To improve the accuracy and generalization of detection, it is necessary to optimize the existed network. Firstly, the key of the stochastic gradient descent is the choice of the learning rate, which directly influences the convergence speed and testing accuracy of the network. In this paper, an adaptive learning rate based on the changes in the loss and the weight was proposed, respectively. A larger learning rate was obtained to accelerate convergence in the initial training stage, and a smaller learning rate produced due to a higher accuracy in the later of network, which can avoid the oscillation. In the process of parameters update, a quadratic function was introduced to the loss to combine with the adaptive learning rate based on loss change for parameters update, and a normal distribution model was introduced to combine with the adaptive learning rate based on weight change to further update parameters. Secondly, to improve the generalization performance of the model, the domain adaption learning (DAL) for transfer learning was joined to the network. A domain classifier was introduced behind the feature layer, which can measure the difference of feature distribution between domains by the domain classification loss, and minimize difference of feature distribution by the marginal distribution adaption loss. A conditional distribution adaptation was introduced behind the Softmax layer, which can measure and adjust the category distribution difference by the conditional distribution adaption loss. These losses for domain adaption learning were optimized by backpropagation, which can locally and globally adjust the distribution to minimize the difference between domains, then the source domain information was migrated to the target domain, so the network trained on source data can also perform well on target data. The loss of marginal distribution adaption and conditional distribution adaption were combined with the classification loss to form the loss of the proposed model. The loss of conditional distribution adaption and classification were used to update the parameters by the adaptive learning rate based on loss change, and the marginal distribution adaption loss was used to further update the feature parameters by the adaptive learning rate based on weight change. The additional domain classifier parameter was updated by the adaptive learning rate based on domain classification loss variation. Finally, to verify the proposed method, the background above was seen as source domain, and a white background was selected as target domain. The parameters trained on source domain were used to initialize the proposed model, the source data and target data were taken as input to train the model. The result showed that the average accuracy for peanut detection in the target domain was 99.7%. And a higher convergence and accuracy generated by comparing with the traditional CNN at different learning rates and training sample sizes. To further prove the performance of the proposed approach, it was compared with the classic VGGNet、GoogLeNet、ResNet and DenseNet, which produced a satisfactory result. [ABSTRACT FROM AUTHOR]

To improve the correct rate of image classification by convolutional neural network, this paper proposed a multimodel fusion convolutional neural network after research on the network structure. By extracting the output feature vectors of a single model and then fusing them, it obtained the new output feature vectors, and then set up a single classifier to classify the images, and improved the accuracy of the classification. By comparing the classification accuracy of single model and multimodel fusion, the class ification accuracy of multi-model fu sion convolutional neural network was improved. This paper analyzed the weight distribution of the last layer of the convolutional neural network, and found that the weight distribution curve of the same model on different data sets was similar and the weight distribution curve of the network model with better classification effect was more gentle. [ABSTRACT FROM AUTHOR]

Firstly, this paper reviewed the development of computer vision, presented the elementary knowledge of neurons, multi-layer perceptron, back-propagation algorithm, and the development of convolutional neural network (CNN) and its basic operation of convolution and pooled etc. Secondly, it discussed the structures of typical convolutional neural networks such as AlexNet, VGGNet, GoogLeNet and ResNet etc., and mainly introduced CapsNet. And then it summaried the research progress of the application of CNN in image classification, image semantic segmentation, image target detection and image generation. At last, this paper presented the research challenge of CNN and the future next work for CapsNet. [ABSTRACT FROM AUTHOR]