Computer and Modernization

Multi-scene Fusion Algorithm for Fine-grained Image Caption

LI Xin-ye, ZHANG Cheng-qiang, ZHOU Xiong-tu, GUO Tai-liang, ZHANG Yong-ai

2021, 0(09): 1-6.

Asbtract ( 296 )

PDF (1256KB) ( 247 )

References | Related Articles | Metrics

In terms of the poor performance of image caption task in different scenes, a multi-scene image caption generation algorithm based on convolutional neural network and prior knowledge is proposed. The algorithm generates visual semantic units by convolutional neural network, then uses named entity recognition to identify and predict image scenes, uses the result of classifying to adjust the focusing parameter of self-attention mechanism automatically, and calculate the multi-scene attention score. Finally, the obtained region coding and semantic prior knowledge are inserted into Transformer text generator to guide sentence generation. The results show that the algorithm can effectively solve the problem that the caption lacks the key scene information. Evaluation indicators are used to evaluate the model on the MSCOCO dataset and Flickr30k dataset, and the CIDEr score of MSCOCO dataset reaches 1.210, which is better than similar image description generation models.

Leaf Recognition Method of Invasive Alien Plants Based on Improved VGGNet Model

YUAN Zhong-hu, WANG Wei, SU Bao-ling

2021, 0(09): 7-11.

Asbtract ( 282 )

PDF (856KB) ( 194 )

References | Related Articles | Metrics

In view of the leaves of different species of plants in nature may have small differences, which leads to the problem of leaf recognition errors of some native plants and invasive plants with similar edge profiles, a PF-VGGNet model is proposed. The common VGGNet model performs well in image classification. Using the sequential connection structure, it can extract the high-level semantic information features of the image, but the shallow contour and texture features of some images also play a key role in the classification. The PF-VGGNet model can fuse the shallow contour and texture features with the deep semantic information of the network to realize the automatic recognition of plant leaves. The experimental results show that the PF-VGGNet model has better recognition effect than other algorithms on the self built data set of alien invasive plant leaves, and the accuracy rates in training set and test set are 99.89% and 99.63% respectively. The PF-VGGNet can effectively reduce the problem of recognition error caused by the similar edge contour of leaves, can quickly identify the leaves of alien invasive plants, and provide support for the prevention and control of alien plants.

Face Mask Detection Algorithm Based on DCN-SERes-YOLOv3

LI Guo-jin, RONG Yu

2021, 0(09): 12-20.

Asbtract ( 317 )

PDF (13180KB) ( 302 )

References | Related Articles | Metrics

With the outbreak of the COVID-19 epidemic in 2020, wearing mask is one of the important measures to effectively suppress the rebound of the epidemic. It is of great practical significance to study the use of machine vision technology to detect whether face masks are worn or not. This paper proposes a face mask detection algorithm based on DCN-SERes-YOLOv3 to solve the problems of occlusion, small detection targets, unobvious feature information, and difficult identification of the target group when wearing masks in video image. Firstly, the algorithm uses the combination of ResNet50 and YOLOv3 to replace the backbone network with the ResNet50 residual network. In order to balance the accuracy and speed of the model, the convolutional layer in the residual block is improved and the average pooling layer is added to reduce the model’s loss and complexity, improve the detection speed. Secondly, the conventional convolution of the fourth residual block in the ResNet50 residual network is replaced with DCN deformable convolution to improve the model’s ability to adapt to geometric deformation when wearing masks. Finally, the SENet channel attention mechanism is introduced to enhance the ability to express characteristic information. The experimental results show that the average accuracy of the algorithm proposed in this paper is as high as 95.36%, which is about 4.1 percent point higher than the traditional YOLOv3 algorithm, and the detection speed is increased by 11.7 fps. The proposed algorithm improves the precision and the speed of the task of detecting faces wearing masks and has high application prospect.

Ship Object Detection in Any Direction at Sea Based on Active and Transfer Learning

SU Hao, DING Sheng, ZHANG Chao-hua,

2021, 0(09): 21-30.

Asbtract ( 224 )

PDF (8991KB) ( 133 )

References | Related Articles | Metrics

In remote sensing images object detection tasks based on deep learning, the ship usually exhibit features arranged in any direction. The common algorithms of object detection adopt horizontal detection that generally cannot meet the application requirements of such scenarios. Therefore, this paper adds a rotation angle prediction branch to the single-stage Anchor-Free object detector CenterNet, it can output a rotating bounding box for the detection of marine ship objects. At the same time, in view of the problem that maritime ship remote sensing data sets only have horizontal bounding box labels, which cannot be directly applied to rotating boxes object detection, and manual labeling of rotating boxes labels is expensive, an active and transfer learning method of rotating boxes label generation is proposed. Firstly, a horizontal box-rotating box constraint screening algorithm is proposed. The rotating prediction box is supervised and constrained by the horizontal ground truth bounding box. The image with higher detection accuracy is selected and added to the training set. Then this process is iterated to filter out more images. Finally, the automatic labeling of the rotating box of the data set is completed by matching the label categories. In this paper, about 65.59% of the pictures in the remote sensing image data set BDCI of marine ships are finally marked with a rotating boxes, and some unmarked pictures are manually marked as the test set. The pictures marked by the method in this paper are used as the training set for verification. The evaluation index AP50 reaches 90.41%, which is higher than other rotating boxes detectors, indicating the effectiveness of this method.

Infrared and Visible Images Registration Method for Electric Power Inspection

LIU Xiao-kang, XIA Tian-lei, WU Chen-yuan, JIANG Xiong-biao, ZHOU Ming-yu, WANG Qing-hua

2021, 0(09): 31-36.

Asbtract ( 248 )

PDF (6212KB) ( 122 )

References | Related Articles | Metrics

The failure of power equipment may cause power outages and affect the safe and stable operation of power grid. According to the characteristics of heat generated by power equipment during operation, an infrared and visible images registration method for power equipment is proposed to facilitate detection of abnormal heating faults. Firstly, the Sobel edge detection operator extracts edge information from infrared and visible images of power equipment to obtain edge images. Then the feature points of two edge images are detected by the SuperPoint algorithm and the descriptors are calculated, and the feature points are matched by the SuperGlue algorithm. Finally, the affine transformation model parameters are calculated by the least square method to realize the infrared and visible image registration of power equipment. The experimental results show that the method in this paper can achieve high-precision registration of infrared and visible images of power equipment.

Rumor Source Detection Algorithm in Social Networks Integrating Objective Weighting Method

ZHOU Zhong-yue, ZHANG Hai-jun, PAN Wei-min

2021, 0(09): 37-42.

Asbtract ( 184 )

PDF (847KB) ( 92 )

References | Related Articles | Metrics

At present, many detection methods only judge whether the information is a rumor, and there is less research on the source of the rumor. Aiming at the problem of ignoring the application of node weights as an important parameter in rumor source detection in previous studies, a model of objective weighting algorithm based on rumor centrality, namely BEW algorithm, is proposed. The model firstly calculates the weights of network nodes through the entropy weight algorithm, and then simulates network propagation based on the SIR model, while considering the embedded characteristics of the network node weights, and finally achieves the purpose of source point prediction through the MLE algorithm. Simulation experiments are carried out through 4 real networks. The experimental results show that the algorithm can achieve better results in identifying the source of rumors.

Multi-objective Shark Smell Optimization Algorithm Based on Decomposition and Vector

LI Hong-wei

2021, 0(09): 43-50.

Asbtract ( 248 )

PDF (1397KB) ( 151 )

References | Related Articles | Metrics

In order to improve the convergence rate and the distribution of the solution set of the multi-objective shark algorithm, this paper proposes a multi-objective shark smell optimization algorithm based on decomposition and vector (DVMOSSO). Firstly, aiming at the problem that the convergence and diversity of the basic shark algorithm are difficult to balance, this paper uses the reference vector to calculate the angle penalty distance scalar value to balance the convergence and diversity of the solution in the target space in the process of elite centralized mining. In addition, the basic shark algorithm is easy to converge prematurely and fall into local optimum in the late iteration. In this paper, Gaussian mutation strategy is used to reinitialize the particles, and polynomial mutation is used to increase the diversity of the population in the elite solution set. Finally, in order to verify the effectiveness of the proposed algorithm, the proposed DVMOSSO algorithm is compared with NSGAII-DS, MOEA/D, MMOPSO, MOSSO and dMOSSO algorithm in the standard test function. The experimental results show that the proposed algorithm has good convergence and distribution, higher convergence accuracy and stronger optimization ability.

Kriging Parameter Estimation Algorithm Based on Combinatorial Optimization

WANG Hong

2021, 0(09): 51-56.

Asbtract ( 188 )

PDF (997KB) ( 82 )

References | Related Articles | Metrics

Kriging surrogate can be used to identify the parameters for the mathematical models described by ordinary differential equations. By training a relatively small set of samples, Kriging surrogate can partially replace the time-consuming original objective function optimization process, so it can save a lot of computation time. And the optimization algorithm of searching new samples has major impacts on result of the parameter estimation during the process of refining Kriging surrogate models. For the problem of parameters estimation described by ordinary differential equations with nonlinearity and sloppiness, this paper combines the advantages of Adam with second order momentum and SGD with momentum to search for new sample points that need to be added during model refinement, so as to improve the convergence speed and search quality. Compared with other optimization algorithms, the effectiveness of the proposed algorithm is verified.

High Precision Display Method of BIM Model Based on Triangulation Algorithm

WANG Ni, WANG Shu-ying, SHI Hai-ou, YUAN Quan

2021, 0(09): 57-62.

Asbtract ( 292 )

PDF (3033KB) ( 113 )

References | Related Articles | Metrics

The rendering problem of BIM model in the web front end is an important problem in the practical application of BIM Technology. The solution to this problem is to use triangular patches to accelerate the rendering efficiency of the model front end (model lightweight). According to the problem that the average quality coefficient of triangle mesh of BIM model is low in the secondary development technology of Revit, aiming at the application requirements of BIM model lightweight and web-based sharing, an improved algorithm combining the secondary development of Revit and Delaunay subdivision algorithm is proposed. By adding points to the original points of BIM model obtained from the secondary development of Revit, the original points and the added points conform to the Delaunay criterion according to the B-W algorithm, and more precise triangular patches are generated. At the same time, the generation of extraterritorial triangles is avoided, and the practical application effect of the algorithm is improved. The experimental results show that the average mesh quality coefficient and mesh association quality coefficient of the improved algorithm are improved compared with the original algorithm. Finally, the BIM model is rendered by WebGL according to the triangle patch generated by the optimization algorithm, and the rendering of BIM model on the web side is realized, which verifies the effectiveness of the method.

An NLP Based Review Method for Compliance of Technical System

FAN Zhi-qiang, LING Dong-yi , NIU Chan

2021, 0(09): 63-67.

Asbtract ( 220 )

PDF (1195KB) ( 91 )

References | Related Articles | Metrics

Technical system compliance review is an important work before the information system project is approved. It is an effective assessment of the new project approval system's positioning role in the system and standard compliance, and it is also an important measure for project risk control. At present, there is a lack of effective technical means to review and verify the compliance of technical system and standards. The commonly used method is manual review based on expert's experiments. This paper studies the implementation method of technical system compliance review. Taking Natural Language Processing(NLP) as the main technology, the NLP named entity recognition algorithm for technical regime review is proposed. Based on the technical regime review business and review algorithm, the technical regime review system is implemented.

Speech Recognition in Complex Noise Environment

ZHANG Yun-yao, HUANG He-ming, ZHANG Hui-yun,

2021, 0(09): 68-74.

Asbtract ( 204 )

PDF (1507KB) ( 217 )

References | Related Articles | Metrics

Speech recognition is an important way of human-computer interaction. Aiming at the poor performance of traditional speech recognition systems for noisy speech recognition and inappropriate feature selection, a deep autoencoder recurrent neural network model based on transfer learning is proposed. The model consists of encoder, decoder and acoustic model. Among them, the acoustic model is composed of stack bidirectional recurrent neural network, which is used to improve the recognition performance. The encoder and decoder are composed of full connected layers for feature extraction. The structure and parameters of the encoder are transferred to the acoustic model for joint training, the experimental results on noisy Google commands dataset show that the proposed model can effectively enhance the recognition performance of noisy speech and has good robustness and generalization.

Application of YOLOv4 with Mixed-domain Attention in Ship Detection

ZHAO Yu-rong, GUO Hui-ming, JIAO Han, ZHANG Jun-wei

2021, 0(09): 75-82.

Asbtract ( 300 )

PDF (2297KB) ( 304 )

References | Related Articles | Metrics

Marine ship detection plays an important role in the maritime field. Due to the complicated environment and the diversity of ships, existing methods based on convolutional neural network cannot achieve both high accuracy and real-time performance. To solve the problem of ship’s real-time detection in complicated environment, a YOLO-marine model based on YOLOv4 is proposed in which the domain attention mechanism is introduced into backbone. Firstly, the Mosaic method is used to preprocess the ship data. Then the K-Means++ algorithm is used to get initial anchors. The model is implemented on Darknet for training and evaluation with the real ship dataset. The experimental result show that compared with YOLOv4, YOLO-marine improves the mAP of ship detection task by 2.1 percentage points. The model can effectively improve the accuracy of ship detection while ensuring real-time performance. It also gives outstanding results in small and occluded target.

Compression Method of CNN Model for Parameter Reduction

ZHU Xue-chen, CHEN San-lin, CAI Gang, HUANG Zhi-hong

2021, 0(09): 83-89.

Asbtract ( 314 )

PDF (1418KB) ( 136 )

References | Related Articles | Metrics

In order to solve the problem that it is difficult to deploy convolutional neural network model on embedded devices with limited computing and storage resources due to the increasing scale of parameters, a convolutional neural network model compression method is proposed to reduce the scale of parameters. It is found that the number of convolution layer parameters is related to the number of input and output feature maps and the size of convolution kernel, while the number of full connection layer parameters is large and difficult to be reduced significantly. The number of input and output feature maps is reduced by grouping convolution, and the convolution kernel size is reduced by convolution resolution. At the same time, the global average pooling layers are used to replace the fully connected layers to solve the problem of large number of parameters in the fully connected layers. The above methods are applied to LeNet5 and AlexNet for experiments, the experimental results show that the parameters of LeNet5 model can be reduced by 97% and the recognition accuracy can be reduced by less than 2 percentage points by using the combined compression method, the parameters of AlexNet model can be reduced by 95% and the recognition accuracy can be improved by 6.72 percentage points after compression. On the premise of ensuring the accuracy of convolutional neural network, the parameters of the model can be greatly reduced.

Recommendation Algorithm Based on Knowledge Graph and Bi-LSTM

WANG Yu-ying, WANG Yong

2021, 0(09): 90-98.

Asbtract ( 416 )

PDF (1255KB) ( 203 )

References | Related Articles | Metrics

At present, most of the existing model-based recommendation algorithms input the score data into the deep learning model for training to get the recommendation results. Its defect is that it is unable to analyze the interpretability of the prediction results. In addition, the algorithm can not effectively solve the cold start problem. Therefore, this paper proposes a recommendation algorithm based on knowledge map and Bi-LSTM to effectively solve the problem of interpretability and cold start of the algorithm. Firstly, the data set is preprocessed to generate precoding vector. According to the connectivity of data aggregation points, the domain knowledge map is constructed. Secondly, the meta path extraction technology of knowledge map is used to obtain multiple user item path information, which is input into Bi-LSTM. A layer of attention mechanism is added to each node of the path, so that the model could effectively obtain the information of remote nodes. Finally, the training results of multiple paths are input into the average pooling layer to distinguish the importance of different paths. The cross-entropy loss function is used to train the model and the prediction results are obtained. The experimental results show that, compared with the traditional recommendation algorithm based on the cyclic neural network model, this algorithm can effectively improve the interpretability and prediction accuracy of the algorithm, and alleviate the cold start problem of the algorithm.

Short-term Traffic Speed Prediction Based on Graph Convolutional Network

WANG Zeng-guang, WANG Hai-qi, CHEN Hai-bo

2021, 0(09): 99-105.

Asbtract ( 182 )

PDF (2722KB) ( 164 )

References | Related Articles | Metrics

Traffic forecasting is an important technology for building intelligent transportation systems. Real-time and accurate traffic forecasting is beneficial to route planning and improve travel efficiency. In order to improve the accuracy of traffic speed prediction, the article proposes a short-term traffic speed prediction model based on graph convolutional network. Firstly, the spatial and temporal characteristics of the traffic speed data are analyzed, and then the learnable adjacency matrix is constructed in combination with the data space characteristics to establish the graph convolution network. At the same time, considering the time characteristics of the traffic data, the long-term and short-term memory network and attention mechanism are added on the basis of graph convolution to jointly construct the prediction model. The experimental results show that due to the consideration of the temporal and spatial characteristics of traffic speed data, the root mean square error, average absolute error and average absolute percentage error of this model are all smaller than the traditional model and the single model, which verifies that the proposed model has higher prediction accuracy.

Textbook Classification Method of Index of Moral Education Based on Deep Learning

GUO Shu-wu, CHEN Jun-hua

2021, 0(09): 106-112.

Asbtract ( 204 )

PDF (1731KB) ( 144 )

References | Related Articles | Metrics

Moral education is the cornerstone of personal development and one of the important responsibilities of schools. As an important carrier of moral education, the index of moral education has naturally become one of the important standards for textbook revision. It is more efficient and reliable to use deep learning to realize automatic classification of textbook titles. However, the text data set of the textbook has the characteristics of abundant text information, unobtrusion of features and unbalanced sample distribution. To solve these problems, a new data enhancement method is combined. According to the contribution of text words vector to the classification results, the attention matrix is calculated by the attention mechanism, and then the word vector matrix is input into the model together, a text classification model IoMET_A bonded with attention mechanism is proposed. IoMET_A is used to study the textbooks of primary and secondary schools in Shanghai. The experimental results show that compared with the original IoMET text classifier, IoMET_A effectively improves the evaluation effect.

S2R2: Semi-supervised Feature Selection Based on Analysis of Relevance and Redundancy

ZHANG Dong-fang, CHEN Hai-yan, YUAN Li-gang

2021, 0(09): 113-120.

Asbtract ( 230 )

PDF (1177KB) ( 99 )

References | Related Articles | Metrics

Feature selection is one of the key problems of pattern recognition and data mining, which can be removed dataset redundant and irrelevant features to improve learning performance. Based on the max-relevance and min-redundancy criteria, a novel semi-supervised feature selection method based on relevance and redundancy analysis is proposed. This new method is independent of any classification learning algorithm. Firstly, unsupervised relevance is analyzed and expanded. Then it is combined with information gain to form a semi-supervised feature relevance and redundancy measures, which can effectively identify and remove irrelevant and redundant features. Finally, an incremental forward search is used to construct feature subset in a greedy manner, which avoiding the search for exponential solution spaces and improving algorithm efficiency. This article also proposes the FS2R2 method as a fast version of the S2R2 method to deal with large-scale problems. The experimental results on standard data sets illustrate the effectiveness and superiority of the proposed approaches.

Combination Forecasting Model of Time Series Data in Smart Agriculture

CHEN Xiao-lei, WANG Xing-xing, SHEN Hao-yang

2021, 0(09): 121-126.

Asbtract ( 281 )

PDF (1340KB) ( 111 )

References | Related Articles | Metrics

Smart agriculture is a technical solution to achieve precision in agriculture. The smart agriculture system can monitor various environmental parameters of plant growth in real time, and many predictive models are applied to simulate the changing trend of crop growth environment and provide a basis for scientific decision-making. In recent years, many scholars have proposed prediction model algorithms for time series, which have achieved good results in terms of prediction stability. In order to further improve the prediction accuracy of time series, a combined prediction model based on autoregressive integrated moving average model and wavelet neural network is proposed. The combined model combines the advantages of two single models, the autoregressive integrated moving average model is used for fitting the linear part of the sequence, and the wavelet neural network is used for correcting the residual error to make the fitting curve closer to the actual value. History within the greenhouse temperature data is used to validate the precision of combination model. Finally, the results of the combined model and the traditional prediction model are compared. The results show that the combined model has higher accuracy and better fitting effect for greenhouse temperature prediction, and the calculation efficiency is about 20% higher than that of the traditional model prediction algorithm.

Table of Content