Computer and Modernization

Multi-feature Fusion Fundus Image Segmentation Based on Codes Structure

DING Wan-ying, CHEN Wei, LI Zhao-hui

2022, 0(07): 1-7.

Asbtract ( 224 )

PDF (2297KB) ( 108 )

References | Related Articles | Metrics

In order to solve the problem that the existing fundus images segmentation methods have low segmentation precision and low accuracy for micro vessels, an improved U-Net network model based on codec structure is proposed. Firstly, the data is preprocessed and expanded, the green channel image is extracted, and the contrast is enhanced by contrasting limited histogram equalization and Gamma transform; Secondly, the training set is input into the neural network for segmentation, the residual module is added in the coding process, the high and low feature information are fused by short jump connection, the receptive field is increased by hole convolution, and the attention mechanism is added in the decoding module to increase the segmentation accuracy of fine blood vessels; Finally, the trained segmentation model is used to predict the retinal vascular segmentation results. Comparative experiments on DRIVE and CHASE-DB1 fundus image data sets show that the average accuracy, specificity and sensitivity of the model algorithm are 96.77% and 97.22%, 98.74% and 98.40%, 80.93% and 81.12% respectively. The results of experiments show that the algorithm can improve the accuracy and efficiency of microvascular segmentation, and can segment retinal vessels more accurately.

Vehicle Target Detection Algorithm Based on YOLO v4

YIN Yuan-qi, XU Yuan, XING Yuan-xin

2022, 0(07): 8-14.

Asbtract ( 331 )

PDF (2959KB) ( 120 )

References | Related Articles | Metrics

Aiming at the problems of low occlusion target detection accuracy and poor small target detection effect in vehicle target detection, an improved target detection algorithm YOLO v4-ASC based on YOLO v4 is proposed. By adding convolution block attention module to the tail of the backbone extraction network, the feature expression ability of the network model is improved; The loss function is improved to improve the convergence speed of the network model, and the Adam+SGDM optimization method is used to replace the original model optimization method SGDM to further improve the model detection performance. In addition, K-Means clustering algorithm is used to optimize the priori frame size, and the car, truck and bus categories in the traffic scene data set are combined as vehicle, which simplifies the problem in this paper into a two classification problem. The experimental results show that on the basis of maintaining the detection speed of the original algorithm, the proposed YOLO v4-ASC target detection algorithm achieves 70.05% AP and 71% F1-score. Compared with the original YOLO v4 algorithm, AP is improved by 9.92 percentage points and F1 score is improved by 9 percentage points.

An Insulator Self-detonation Detection Algorithm on Transmission Line Based on Double Modules

LIN Hang, GENG Duo-fei, YU Hao, HU Dan, ZHANG Ke

2022, 0(07): 15-20.

Asbtract ( 149 )

PDF (5549KB) ( 77 )

References | Related Articles | Metrics

Aiming at the problem that it is difficult to accurately detect the small defects targets of insulator self-detonation in transmission line UAV inspection images, this paper proposes an insulator self-detonation defect detection algorithm based on Faster R-CNN and the improved YOLO v3 cascaded dual model. Firstly, the insulator string defect dataset is constructed using UAV inspection images, and the training image samples are pre-processed by flipping to increase the number of samples and improve the generalization ability of the model and avoid over fitting; then the Faster R-CNN is used to detect the insulator strings in the images, and then the detected insulator string images are fed into the improved YOLO v3 network for locating the self-exploding defects. The improved YOLO v3 network is based on YOLO v3 by borrowing the idea of FPN, adding feature extraction layer and performing feature fusion to make full use of deep and shallow features; meanwhile, the CIoU Loss function is used as the loss function to solve the boundary frame aspect ratio scale information. The experimental results show that the detection accuracy of the proposed algorithm reaches 91.2% on the constructed insulator defect dataset, which is more than 3.31 percentage points higher than that of single-model detection algorithms such as Faster R-CNN or YOLO v3, and can effectively realize the detection of insulator self-detonation defects in UAV inspection, which provides methodological support for intelligent inspection fault diagnosis of transmission lines.

Image Animation Based on Generative Adversarial Networks

ZHAI Hui-cong, ZHANG Ming, DENG Xing, WANG Li-qun

2022, 0(07): 21-26.

Asbtract ( 284 )

PDF (3547KB) ( 102 )

References | Related Articles | Metrics

Anime-style images are highly simplified and abstract. In order to solve the problem of transforming real-world images into anime-style images, this paper proposes an image animation method based on generative adversarial networks. The generation network in this paper is like a U-Net fully convolutional structure. The input image is down-sampled first, and the shallow features are up-sampled by bilinear interpolation. The discriminant network uses Patch GAN and spectrum normalization. Semantic content loss and style loss are calculated separately to improve the stability of the network. Surface representation loss, structure representation loss, and texture representation loss are used to replace style loss to make the effect of generating animation pictures more controllable. We use train2014 for realistic images, and use the CelebA-HQ data set for face images. Experiments are performed on these data sets using this model. The experimental results show that the model in this paper can effectively complete the process of image animation and generate high-quality animation images.

Traffic Sign Detection in Blurred Light Scenes Based on Improved YOLO v4

SHEN Zhi, XU Li, FU Xiang-yuan

2022, 0(07): 27-32.

Asbtract ( 197 )

PDF (1524KB) ( 82 )

References | Related Articles | Metrics

In recent years, autonomous driving has begun to come into people’s sight. For autonomous driving, traffic sign detection in fuzzy light scene is an extremely important part. At present, YOLO v4 algorithm has been widely used in target detection, although its detection accuracy is greatly improved compared with other versions, but it has not reached the expected accuracy. In order to further improve the accuracy of detecting traffic signs, this article makes certain improvements on the basis of the original YOLO v4 and combines it with MSRCR image enhancement processing. Firstly, the original training images are enhanced by MSRCR algorithm, and it is used as the training set image of target detection. This article uses Darknet-53’s YOLO v4 network, labeles BelgiumTS traffic signal data set by labelImg, and uses the improved K-means++ clustering algorithm to determine the priori box and specific parameters, and improves the path aggregation network（PANet） structure and loss function to train the data set. Experimental results show that compared with the original YOLO v4 algorithm, the improved algorithm has an average accuracy increase of 1.86 percentage points.

ERCUnet: An Improved Road Crack Detection Model Based on U-Net

LIU Yu-xiang, SHE Wei, SHEN Zhan-feng, TAN Shuai

2022, 0(07): 33-39.

Asbtract ( 359 )

PDF (3566KB) ( 116 )

References | Related Articles | Metrics

Aiming at the problems of traditional road crack detection methods, such as low flexibility and poor universality, refering to the residual design in ResNet and the U-shaped encoding and decoding structure of U-Net model, an improved road crack detection model based on U-Net, named ERCUnet, is designed. The model takes residual blocks as the main body, and optimizes the number of convolution cores of convolution layers at different depths for crack detection. All residual blocks in the model have the same structure, the overall structure of the model is more neat and simple, with good elasticity and strong structure. The residual structure not only makes the feature fusion more sufficient but also avoids the problem of gradient disappearance of deep convolution neural network. The experiment is conducted on the CrackForest dataset. The 118 labeled pictures of CrackForest are divided into training set and testing set according to the ratio of 5〖DK(〗∶〖DK)〗1. Through a series of data expansion methods, the problem of too little training data is effectively alleviated. The loss function combines cross entropy and F1 score to alleviate the imbalance between positive and negative samples. The final experimental results show that the number of parameters of ERCUnet model is only 13.30% of that of U-Net model, the recall, precision, and F1 are all greater than 70%, and noise rate and accuracy are 29.05% and 99.01% on testing set. ERCUnet-tiny model is obtained by modifying model parameters to confirm the plasticity of ERCUnet, and the number of its parameters is only 2.39% of that of U-Net model, similar effect to U-Net is achieved on testing set.

Mask Detection Algorithm Based on Lightweight Structure and Re-parameterized Network

LI Yan, LU Zheng-song, LI Qing-yun , YANG Shi-hai, ZHANG Xiao-long

2022, 0(07): 40-46.

Asbtract ( 157 )

PDF (6259KB) ( 76 )

References | Related Articles | Metrics

Under the situation of normalized epidemic prevention and control, there are dense crowds in railway stations, subway stations and other public places, which are prone to the spread of virus. Aiming at the problems of small mask targets, large amount of model parameters and difficult to deploy in crowded places, an improved lightweight structure and re-parameterized network is proposed. On the Retinaface algorithm, the dual cascade pyramid network is used to replace the original feature fusion network to enhance the feature information and to improve the detection effect of small-scale targets. At the same time, the structure re-parameterized network RepVGG is used to replace the original MobileNet0.25 backbone network. During model training, the residual structure is used to improve the feature extraction ability of the model. During model reasoning, the model parameters are reducedand the reasoning speed is improved by re-parameterization of the model structure. The experimental results show that the frame rate is 92.59 fps and the average accuracy rate (mAP) of the proposed algorithm on three different levels of verification sets of self-building data sets is 94.17%, 93.30%, 86.88%, which is 1.17 percentage points, 2.89 percentage points and 5.35 percentage points higher than the original Retinaface algorithm respectively. Mask wearing detection can be better carried out in natural scenes.

User Purchase Forecast Method Under Softvoting Strategy Based on Improved EasyEnsemble

YANG Jin, ZHANG Chen

2022, 0(07): 47-53.

Asbtract ( 156 )

PDF (1512KB) ( 87 )

References | Related Articles | Metrics

With the development of Internet, shopping online has become an increasing choice for people. In order to better achieve the purpose of helping customers to recommend products, the feature of original data is extracted and the feature of the data is selected by mutual information method. The improved EasyEnsemble algorithm is used to deal with the problem of category imbalance, and the defect of under-sampling is compensated by integration strategy. The sample data is fully utilized and the influence caused by positive and negative sample difference is reduced. Finally, the softvoting method is used to combine XGBoost and random forest into a final classifier for prediction, which is compared with the single algorithm, so as to get better results. Based on the data provided by Alibaba Tianchi Competition, the precision rate P, recall R and F1 values are taken as evaluation indexes to compare with the current popular machine learning algorithms respectively to verify the effectiveness of this method.

Short-term Traffic Flow Prediction Model Based on Deep Learning

ZHANG Ling-yun, HAN Ying, ZHANG Kai, LU Hai-peng, DING Yu-jie

2022, 0(07): 54-60.

Asbtract ( 329 )

PDF (2733KB) ( 91 )

References | Related Articles | Metrics

Traffic flow prediction has important and practical significance in the field of intelligent transportation. Because traffic flow data is affected by many factors, leads to poor stability, strong randomness, and presents a highly non-linear characteristic, it is extremely difficult to predict traffic flow. Aiming at the requirements of the accuracy of short-term traffic flow prediction, this paper proposes a short-term traffic flow prediction method based on CEEMD(Complete Ensemble Empirical Mode Decomposition, CEEMD), combined with CNN(Convolutional Neural Networks, CNN) and LSTM(Long Short-Term Memory, LSTM). The model uses CEEMD signal decomposition to reduce the impact of noise on traffic flow data prediction, CNN and LSTM are used to fully mine the temporal and spatial characteristics of the data, so that the model can make more accurate judgments and improve the learning efficiency of the neural network. Experimental verification on real traffic flow data shows that the model proposed in this paper can effectively improve the accuracy of traffic flow prediction.

An Improved LDA Algorithm Based on Graph Mining

LI Shan, CHEN Miao-miao, ZHENG Chen

2022, 0(07): 61-66.

Asbtract ( 143 )

PDF (2296KB) ( 73 )

References | Related Articles | Metrics

As one of the most widely used models in the field of text topic recognition, LDA simplifies the assignment of the same weight to words based on the assumption of bag-of-words model, which makes the topic distribution inclined to high-frequency words, as well as affects the semantic coherence of the recognized topics. This paper proposes an improved LDA algorithm based on graph mining, named GoW-LDA, which firstly builds a semantic graph model based on the co-occurrence of feature word pairs in the text, then uses the weighting degree of nodes in network statistical features to integrate the semantic structure characteristics and relevance of the text into the LDA topic modeling in the form of weight correction. Experimental results show that, compared with traditional LDA and TF-IDF-based LDA, GoW-LDA can greatly reduce the complexity of topic models, improve the PMI of topic recognition, and effectively reduce the training time, which provides for a new solution idea text topic recognition.

Real-time Self-correction of Digital Twin Model Based on Consistency Measurement

XU Lin, HE Yue-shun, SONG Wei-ning, XU Ting-ting

2022, 0(07): 67-73.

Asbtract ( 223 )

PDF (1651KB) ( 73 )

References | Related Articles | Metrics

Digital twin technology solves the problem of integration in the cyber-physical world, and has been widely used in the field of industrial Internet. In order to solve the problem of dynamic correction between digital twin and physical entity, this paper proposes a real-time self-correction method for digital twin model based on consistency measurement. Using the speed of data change, the model is divided into two parts: a gradual model and a fast model. A quick parameter search method is constructed. Combined with Latin hypercube global search and greedy local search, an iterative update mechanism is introduced to achieve the consistency measurement of physical entities and digital twins. The results show that the digital twin model improves the randomness of the adjustable parameter selection by optimizing the model adjustable parameter selection process, achieves ahigh degree of consistency between the model and the physical entity, and meets the requirements of model real time self-correction.

Flow Table Merging on Programmable Data Planes Based on MAT Model

LING Zhi-yuan, CHEN Xiao, SONG Lei,

2022, 0(07): 74-78.

Asbtract ( 156 )

PDF (919KB) ( 73 )

References | Related Articles | Metrics

Compared with traditional networking technologies, SDN technologies separate the control plane of the networking from the data plane, making the networking programmable. Most of the SDN switches, like OpenFlow, POF, P4, are implemented based on the match-action table model. Different from protocol-aware SDNs, such as OpenFlow, protocol-oblivious SDNs define protocol fields by {offset, length} structures, so as to realize the parsing and processing of any protocol fields. However, packets to be processed may have packet headers with different lengths. In order to match the fields in packet header protocols, the control plane needs to install more flow tables to parse packets, which results in complex flow tables and pipelines. To address the above problems, this paper proposes a flow table merging scheme on programmable data planes based on the MAT model, which extends the action set in the MAT model. The starting offset of packets can be adjusted dynamically with specific actions when a packet queries the flow table, making the offset of the same matching fields in different packets consistently, so that flow tables with the same matching fields are merged. The scheme reduces the flow table memory consumption by about 69% at the cost of executing one more action when jumping the flow table in the POF Switch experimental scenario compatible with VLAN and QinQ.

Container Cloud Queue Online Task Dynamic Allocation Based on Long Short-term Memory Neural Network

XU Sheng-chao, YE Li-hong

2022, 0(07): 79-84.

Asbtract ( 113 )

PDF (933KB) ( 82 )

References | Related Articles | Metrics

Aiming at the problems of the poor rationality of allocation and resource balance degree and the low task processing efficiency of existing container cloud online task allocation methods, a dynamic online task allocation method of container cloud queue based on long short-term memory neural network is proposed. This paper describes the online task model of container cloud queue, assigns multi-objective functions with node complementarity, resource utilization ratio and energy consumption composition, solves the optimal task allocation scheme with long short-term memory neural network under the constraint condition, and completes the dynamic task allocation of container cloud queue. The experimental results show that the allocation rationality of the allocation scheme proposed in this paper reaches 0.925, the resource balance degree reaches 10.255, the longest queue length is 10, and the maximum energy consumption value is 5000 W. The allocation rationality, resource balance degree and task processing efficiency are all improved, and the allocation scheme is more reasonable.

End-to-end Optical Music Recognition Method Based on Residual Gated Recurrent Convolutional Neural Network and Attention Mechanism

SUN Hong-yang, WANG Shang

2022, 0(07): 85-90.

Asbtract ( 219 )

PDF (1160KB) ( 85 )

References | Related Articles | Metrics

Optical music recognition(OMR) is of great significance to promote the intelligence and digitization of music. The traditional music recognition process is complicated and easy to lead to the accumulation of errors, but current sequence modeling-based optical music recognition methods cannot obtain notes context information from the full scale, there is still room for improvement in the recognition effect. To this end, this paper proposes an end-to-end optical music recognition method based on residual gated recurrent convolution and attention mechanism. The method uses residual gated recurrent convolution as the backbone network to enrich the model’s ability to extract contextual information; Combined with an attention mechanism decoder, the feature information of the music score and its internal correlation can be better mined to enhance the representation ability of the model and identify the notes and notes sequences in the score image. The experimental results show that, compared with the Convolutional Recurrent Neural Network (CRNN) model, the improved network has a significant decrease in both the symbol error rate and the sequence error rate.

Research and Design of Science and Technology Service Resource Pool Oriented to Central Plains Urban Agglomeration

SHAN Ke, ZHANG Yi-ming, LIU Rui-xia,

2022, 0(07): 91-96.

Asbtract ( 118 )

PDF (1355KB) ( 73 )

References | Related Articles | Metrics

With the rapid development of science and technology in China, great progress has been made in the informatization construction of the Central Plains urban agglomeration, and the scientific and technological resources with regional characteristics have been formed. But on the whole, the science and technology service resources of Central Plains urban agglomeration are scattered, the integration is not high, and the service efficiency is low, so it is difficult to achieve the real sense of science and technology resource sharing. In order to solve the above problems, this paper studies data extraction, processing, verification, standardization and fusion technology. Through the technology service data collection integration engine, this paper uses big data integration technology to integrate the scattered and independent various technology service data into one; HBase column storage database and HDFS distributed file system are used to store different types of data to support the parallel processing of structured, semi-structured, and unstructured data, and then providing various data services such as data retrieval and data analysis to the platform, solving various differences problems of source data integration, completing the architectural design of the science and technology service resource pool of the Central Plains urban agglomeration, improving the utilization rate of science and technology resources, and promoting the rational use of science and technology resources.

Discrimination of Converter Steelmaking State Based on Improved Attention Network

HE Yu-xia, CAO Guo

2022, 0(07): 97-102.

Asbtract ( 103 )

PDF (1066KB) ( 79 )

References | Related Articles | Metrics

The status discrimination of converter steelmaking has a direct impact on the quality of finished steel. The manual experience-based state discrimination requires continuous observation of flame changes at the furnace mouth, which is highly subjective and costly. In order to improve the accuracy of the judgment of the converter steelmaking state, a 3D residual convolutional neural network model based on attention mechanism is proposed. The improved channel attention combines average pooling and maximum pooling for feature fusion, which can infer finer channel features, and the spatial attention can extract key information in space. The experiment results show that the improved model is better than the Squeeze and Excitation Module(SE), Convolutional Block Attention Module(CBAM) and Efficient Channel Attention Module(ECA). Compared with the 3D residual model without attention mechanism, the F1 score is improved by 1.03 percentage points and the accuracy is improved by 1.06 percentage points. Finally, the influence of channel attention and spatial attention on the model are analyzed through ablation experiments.

Android Application Performance Testing and Monitoring Technology

XIAN Jin

2022, 0(07): 103-109.

Asbtract ( 218 )

PDF (1574KB) ( 76 )

References | Related Articles | Metrics

According to the actual demand of mobile application, the performance test and monitoring scheme of Android application is designed and implemented. The solution is simple to use and does not require code injection. It has SDK basic modules, which can complete configuration file parsing, network identification, location and data transmission, and monitor application CPU and memory resource consumption. User data is collected and sent to the Web server for analysis to obtain user composition and retention; By monitoring Activity and Fragment, the time-consuming, high-frequency and long-used page information are obtained, as well as the jump relationship between pages; By monitoring data access performance, performance bottlenecks and defects are found. Finally, a mobile shopping application is used to verify the effectiveness of the scheme. Through the combination of application performance testing and monitoring, developers can discover application problems in time and continuously track and improve software services.

An Adaptive MIS Generator Based on Object Information

ZHOU Bin,

2022, 0(07): 110-120.

Asbtract ( 80 )

PDF (2499KB) ( 90 )

References | Related Articles | Metrics

In order to solve the problem of increasing workload of development and maintenance caused by frequent changes of requirements during the development and operation of MIS, a new adaptive generation manager of basic modules of MIS software is proposed. MIS version, user and authority, system structure, system module, object information, general view element, general business logic element and general data access mode are managed through informatization, and the database, interactive interface, business logic, data interaction part and software internal structure of MIS software basic modules are adaptively generated or reconstructed based on general interface layer basic element, business logic layer basic element, data access layer basic element and object information, so as to dynamically build and manage the basic module of MIS software, and to realize the informatization of basic module, and the informatization and automatic management of the system. Through the implementation and practical application of the generator technology in MIS development, it is proved that the adaptive MIS generator can better deal with the rapid construction and change of MIS basic modules, and can effectively reduce the workload of MIS development and maintenance. The generator solves the problem of rapid construction and reconstruction of basic modules of MIS software, and can better deal with the frequent changes in the development and operation of MIS.

GRU Adversarial Network Text Generation Model with Reward

PENG Peng-fei, ZHOU Lin-ru

2022, 0(07): 121-126.

Asbtract ( 179 )

PDF (1633KB) ( 71 )

References | Related Articles | Metrics

Aiming at the problems of accumulated errors caused by the supervised form of the current generative adversarial network text generation model and the single generated text information, a text generation model based on GRU generative adversarial network is proposed. The GRU generator uses rollout-policy to update parameters, and Monte Carlo search is added into the model to generate sample sequences. The GRU neural network with fewer parameters is used as the generator and the discriminator. The output loss function of the discriminator guides the parameter optimization in the generation process, and the Monte Carlo strategy is used to supplement the incomplete sequence in the generation process to reduce the accumulation of errors and increase the text richness of generated information. This paper introduces the gate truncation mechanism, replaces the sigmoid function in the GRU network with a custom function, improves the activation function of the implicit variable at the current time, and improves the slower convergence speed of the original function and the problem of gradient disappearance, making it more suitable for this model. The results of simulation experiments show that this model enriches the diversity of text generation, improves the convergence speed of the model, and proves the effectiveness of this model. The model has good applicability.

Table of Content