Computer and Modernization

A Hybrid Brain Tumor Classfication Study Based on CBAM and EfficientNet with Improved Channel Attention

HUA Xin-yu, QI Yun-song

2023, 0(05): 1-7.

Asbtract ( 515 )

PDF (1818KB) ( 104 )

References | Related Articles | Metrics

In order to further improve the accuracy and robustness of brain tumor image diagnosis， a novel hybrid brain tumor classification method based on CBAM（Convolutional Block Attention Module） and EfficientNet with improved channel attention mechanism （IC+IEffxNet） is proposed. The method is divided into 2 stages. In the first stage， the features will be extracted by CBAM model based on improved spatial attention mechanism. In the second stage， the sequence and exception （SE） block in EfficientNet architecture is replaced by the efficient channel attention （ECA） block， and the combined feature output of the first stage is used as the input of the second stage. Experiment shows the 4 classifications of glioma， meningioma， pituitary and normal images from the mixed brain tumor MRI dataset. The results show that the average classification accuracy is about 0.5~2 percentage points higher than the existing methods. The experimental results demonstrate the effectiveness of the method and provide a new reference for medical experts to accurately judge brain tumor.

Image Caption Generation Method Based on Channel Attention and Transformer

LIU Jing, CHEN Jin-guang

2023, 0(05): 8-12.

Asbtract ( 276 )

PDF (10305KB) ( 80 )

References | Related Articles | Metrics

Image caption generation refers to translating an image into a sentence by computer. Aiming at the problems of existing image caption generation tasks， such as insufficient use of local and global features of images and high time complexity， this paper proposes a hybrid structure image caption generation model based on Convolution Neural Networks (CNN) and Transformer. Considering the spatial and channel characteristics of the convolutional network， firstly， the light-weight and high-precision attention ECA is fused with the convolutional network CNN to form an attention residual block， which is used to extract visual features from the input image. Then the features are input into the sequence model Transformer. At the encoder， use self-attention learning to obtain the participating visual representations. At the language decoder，capture the fine-grained information in the caption and learn the interaction between the caption sequences. The model was validated on MSCOCO dataset， and the BLEU-1， BLEU-3， BLEU-4， Meteor， and CIDEr metrics were improved by 0.3， 0.5， 0.7， 0.4， and 1.6 percentage points， respectively.

Tibetan Named Entity Recognition Based on Small Sample Learning

YU Tao, ZHANG Ying, Yong Tso,

2023, 0(05): 13-19.

Asbtract ( 249 )

PDF (1601KB) ( 59 )

References | Related Articles | Metrics

The task of Tibetan named entity recognition is to identify the names of people， places and organizations in the text. This paper proposed a Tibetan named entity recognition method based on small sample learning. In the training process， the feature fusion of entity location information， word segmentation information and Tibetan syllable information in the form of dimensional splicing could better represent the boundary information of Tibetan long entities. Ablation experiments were designed to explore the effect of different feature information on model performance. The experimental results show that our method is effective， and the F1 value is improved by 22.22~38 percentage points compared with the baseline experiment.

Trend Prediction of Infectious Diseases Based on Logistic-GF-SEIR Model

WU Le, CHEN Gang, LI Zhu

2023, 0(05): 20-25.

Asbtract ( 226 )

PDF (14136KB) ( 58 )

References | Related Articles | Metrics

In order to improve the prediction accuracy of the epidemic trend of new infectious diseases， this paper improves the traditional SEIR model and proposes the Logistic-GF-SEIR model. Firstly， based on historical data， the Logistic model is used to fit the cumulative rehabilitation， and the daily rehabilitation rate， infection rate and contact rate are inverted. Secondly， Gaussian model and Logistic model are used to fit the optimal time-varying parameters. Finally， the initial value of the model is initialized to predict the trend of epidemic population. Taking the epidemic development trend of Wuhan and Japan in the early stage of COVID-19 outbreak as an example， the simulation test is carried out and compared with Logistic， SEIR， ARIMA， BP neural network and other prediction models. The results show that the fitting and prediction performance of the Logistic-GF-SEIR model is better than other models in the prediction of the epidemic situation in Wuhan， and the root mean square error is better than other models in the prediction of the epidemic situation in Japan， which verifies the feasibility， effectiveness and robustness of the proposed model. It can provide a basis for China to formulate prevention and control policies for similar infectious diseases.

Asymmetric Deep Supervised Hashing with Attention Mechanism

WANG Xin-yi, YIN Si-qing, HONG Jun

2023, 0(05): 26-31.

Asbtract ( 284 )

PDF (5364KB) ( 69 )

References | Related Articles | Metrics

With the advent of the era of big data， the information data on the Internet is growing exponentially. Among these data， image resource accounts for a very large proportion， so how to carry out accurate and efficient image retrieval from massive images has become one of the important research topics today. At present， there are some problems in large-scale image retrieval， such as poor retrieval performance and low accuracy due to the inability to effectively focus on the key areas of the image. Based on the above shortcomings， an asymmetric deep hash algorithm that integrates the attention mechanism is proposed， which is modified based on convolutional neural network. The existing mixed attention mechanism guided by semantic features is improved and embedded into the network， so that the hash network can analyze the global semantic information and local semantic information together. At the same time， a new quantization function is designed to reduce quantization error， so as to enhance the feature expression ability of hash coding. This method is compared with other hashing methods on the CIFAR-10 and NUS-WIDE datasets with evaluation standard mAP. The results show that the proposed network model can combine global and local spatial features well， and improve the image retrieval performance.

Prediction of Short-term Taxi Flow Based on Spatio-temporal Characteristics

SU Jin-ku, GUI Zhi-ming

2023, 0(05): 32-38.

Asbtract ( 235 )

PDF (2206KB) ( 59 )

References | Related Articles | Metrics

Traditional passenger flow forecasting research only focuses on the time-series characteristics of data and ignores the spatial dimension characteristics， weather or other external factors. This paper proposes a convolution gated spatio-temporal forecasting model (KSTCGN) combined with attention mechanism to predict taxi passenger flow. In this model， convolutional neural network (CNN) is used to extract the spatial features of the traffic in each period of the grid， and gated recurrent unit (GRU) is used to extract the temporal features of the passenger traffic. The convolution layer introduces CBAM attention mechanism to pay more attention to important spatial points. GRU layer combines attention mechanism to focus on the time period that has an important impact on traffic， and uses K-means clustering algorithm to cluster different time periods. Through experimental analysis and comparison with other traditional prediction algorithms， it is proved that the proposed combined model can effectively improve the prediction accuracy.

Prediction Method of Foundation Pit Displacement Based on Spatiotemporal Attention Mechanism#br#

WANG Yu-li, YANG Chang-song, QIU Jing, WEI Jun, WU Hong-jie,

2023, 0(05): 39-45.

Asbtract ( 194 )

PDF (1807KB) ( 41 )

References | Related Articles | Metrics

The safety management of the foundation pit is the key content of the construction of large-scale building foundation pits， and the displacement prediction of the foundation pit structure is an important means to prevent the maintenance accident of the foundation pit. However， due to the complex causes of local pit displacement in the pit， the existing support vector regression （SVR） and random forest （RF） methods ignore the characteristics of local weakening of the pit displacement with spatial displacement and accelerating growth with time local displacement， resulting in low prediction accuracy. Therefore， in this paper， a GA-BP neural network method that integrates the spatiotemporal attention mechanism （A-GA-BP） is proposed， which accurately represents the spatiotemporal dimensions and feature correlations of the foundation pit displacement prediction through spatiotemporal features， and improves the effectiveness of the foundation pit displacement prediction. Finally， taking a large-scale project in Suzhou as an example， this paper trains and evaluates the horizontal and vertical displacement monitoring data of the foundation pit， and quantifies the temporal features， spatial features and multi-order temporal and spatial features， and compares them with the existing methods. Experiment results show that the fitting index of this method is 29.19% and 41.25% higher than that of other methods， and the multi-order temporal and spatial features are 3.08% and 1.83% higher than the temporal or spatial features alone.

Dynamic Transfer Method Based on Sensitivity in Industrial Control Network Anomaly Detection

YANG Jun, WANG Jin-lin, NI Hong, SHENG Yi-qiang,

2023, 0(05): 46-51.

Asbtract ( 177 )

PDF (1304KB) ( 26 )

References | Related Articles | Metrics

With the continuous improvement of the informatization of industrial control networks， industrial control networks have gradually become more open， which on the one hand provides convenience for industrial production， but also brings security risks on the other hand. As an important infrastructure， the industrial control network will cause serious damage once it is attacked. In recent years， scholars have used network anomaly detection technology to discover potential security risks in industrial control networks， and have achieved great results. However， the data in the industrial control network often lack labels， which limits the application of traditional supervised learning algorithms in the field of industrial control network security. Algorithms based on unsupervised learning can detect anomalies in scenarios of lacking labels， but there is often a problem of poor algorithm performance， while transfer learning algorithms can get a better result by migrating to the target domain with only a few labels after learning on the source domain. In order to further improve the performance of anomaly detection in industrial control networks with few labels， this paper proposes a dynamic transfer method based on sensitivity in industrial control network anomaly detection. First of all， the algorithm is based on the idea of transfer learning， which is trained in the labeled source domain and then migrated to the target domain with a small number of labels， which can detect anomalies in the industrial control network environment with only a few labels. Secondly， benefits from the memory effect of the GRU structure， the algorithm can effectively utilize the inherent time-series correlation of industrial control network data， which further improves the ability of algorithm anomaly detection. At the same time， the method of dynamic transfer of parameters based on parameter sensitivity factor in the algorithm improves the insufficiency of the traditional transfer learning fine-tuning method for the unbalanced learning of the underlying features of the source domain and target domain data. The comparative experiments on the KDD99 dataset and the Kyoto2016 dataset show that the dynamic transfer learning method based on the sensitivity factor adopted by the algorithm has a better effect than the traditional fine-tuning method. In comparison with the latest series of unsupervised and transfer learning algorithms， the algorithm outperforms the comparison methods in precision， recall， and comprehensive F1 score， achieving excellent performances of 0.97， 0.95， and 0.96.

Improved End-to-end Synthetic Speech Detection Method Based on Auxiliary Learning

YUAN Tian-tian, LI Zhi-hua, QIU Yang

2023, 0(05): 52-57.

Asbtract ( 208 )

PDF (2323KB) ( 40 )

References | Related Articles | Metrics

With the development of deep forgery technology， synthetic speech detection faces more and more challenges， a synthetic speech detection method is proposed， which integrates auxiliary learning into end-to-end model. After data alignment， the audio data is directly input to the improved end-to-end model without extracting any manual features. The main task is to classify real speech and synthetic speech. At the same time， different synthetic speech types are selected as auxiliary tasks to provide a priori hypothesis for the combined speech detection of the main task， and the weight superposition of the main and auxiliary tasks is optimized. The experimental results on the open datasets ASVspoof2019 and ASVspoof2015 show that the improved model in this paper can effectively reduce the equal error rate compared with the model using manual features， and is better than the end-to-end model before the improvement， and has better generalization ability in the face of unknown attack types.

Overview of State Machine Inference Technology for Unknown Protocols

SHENG Jia-jie, NIU Sheng-jie, CHENG Yang, FANG Wei-qing, ZHANG Yu-jie, LI Peng, HU Su-jun

2023, 0(05): 58-67.

Asbtract ( 294 )

PDF (2194KB) ( 88 )

References | Related Articles | Metrics

Protocol reverse engineering (PRE) describes the behavioral logic of the protocol， which is generally divided into 2 steps: protocol format extraction and state machine construction. These two steps are both interrelated and independent. PRE has important significance in the field of network security. In this paper， we have comprehensively sort out the relevant reference of protocol state machine inference. The research status and development trend of protocol state machine reasoning are summarized and analyzed. Firstly， we introduce the formal definition and basic principles of PRE and discuss the specific requirements of the main fields. Secondly， we analyze the state machine inference methods and divide them into three patterns: clustering method， state-related method， and polling state entity. Then we compare the inverse ability and time efficiency of the algorithms from different perspectives. Finally， the development trend of protocol state machine reasoning is prospected.

A Smart Park Management Platform for Multi-source Data

ZHOU Ming-sheng, ZHANG Wen

2023, 0(05): 68-74.

Asbtract ( 199 )

PDF (2755KB) ( 51 )

References | Related Articles | Metrics

Through the application of new generation information technology such as Internet of Things， cloud computing and big data， the smart park realizes the automation and intelligence of park management or services， and promotes the high-quality development of the park. This paper proposes a smart park management platform architecture for multi-source data， constructs a smart park data management gateway. The gateway integrates and extends the park equipment and facilities management system， integrates the existing informatization achievements of park management and services， and builds a new park real-time state awareness system based on the Internet of Things， which solves the problems of “information islands” and data interaction of multi systems， multi platforms and multi criteria in the park management. Through GIS map panoramic presentation and business function linkage， we realize the centralized display， integrated management and unified dispatching of the park’s real estate， equipment and facilities， which provides real-time， comprehensive and accurate technical support for the park’s operation monitoring and management. The architecture has been realized in Shanghai free trade zone and it achieves good economic and social benefits.

Design and Implementation of Effectiveness Evaluation System of Meteorological Disaster Early Warning Signal in Hainan Province

TIAN Guang-hui, SHEN Xiao-yun, WU Yu

2023, 0(05): 75-79.

Asbtract ( 195 )

PDF (2344KB) ( 33 )

References | Related Articles | Metrics

The issuance of early warning signals for meteorological disasters is the most important means for governments at all levels to prevent meteorological disasters. According to the actual needs of meteorological early warning signal evaluation business， the effectiveness evaluation system of meteorological disaster early warning signal in Hainan province is designed to strengthen the quality management of Hainan meteorological disaster early warning signal release in Hainan. The system design is based on object-oriented technical methods， and uses UML language to abstract describe the evaluation of early warning signals， so as to improve the level of system development and application. The system adopts C/S architecture， takes SQL server database as the support for managing massive meteorological data， and is developed and implemented in VS.NET C# language. The system can provide fast and convenient technical support for the early warning business management department and forecast and early warning release personnel， effectively improve the modernization level of meteorological business.

Pedestrian Visual Tracking Algorithm Based on Improved UpdateNet

HU Xiao, JIAO Li-nan, LIU You-quan

2023, 0(05): 80-85.

Asbtract ( 180 )

PDF (3087KB) ( 46 )

References | Related Articles | Metrics

At present， trackers based on siamese networks mainly regard tracking as a cross-correlation calculation between the target template branch and the branch of the region to be detected， which can achieve a good balance between speed and accuracy. However， in the tracking process， the current frame template is a linear combination of the previous cumulative frame template， which leads to the object occlusion difficult to solve. In order to cope with this thorny problem， we adopt the improved SiamRPN tracker and integrate UpdateNet network to track the single pedestrian target. Firstly， the improved SiamRPN network module is used to generate the linear template， then the UpdateNet network is integrated to generate the updated template and perform multi-stage training. Finally， the optimal parameter model is selected to complete the pedestrian tracking task， according to the evaluation index of the dataset. We make the experiment in the benchmark data sets of the OTB2015 and its subset， the results show that the proposed method has obvious improvement than the original method， accuracy and success rate are increased by 2.1 and 1.6 percentage points respectively， while the real-time tracking frame rate is kept. It is also better than many advanced methods to deal with occlusion， such as DaSiamRPN， SiamDW， etc.

Surface Defect Classification of Aluminum Profiles with Weighted Non-local Modules

WANG Jie, PAN Feng, ZHANG Yan-sha, TAN Mian, YAN Xiao-bo, WANG Lin,

2023, 0(05): 86-92.

Asbtract ( 152 )

PDF (1590KB) ( 29 )

References | Related Articles | Metrics

To address the issues of extreme aspect ratio and difficult classification of small area defects in the task of aluminum profile surface defect classification， we propose a surface defect classification method named Fusion of Weighted non-local Modules and Auxiliary classifier network （FWACNet）. This method proposes a weighted non-local module and uses the dot product similarity to calculate the similarity of different positions in the feature map space to improve the model’s ability to capture long-distance dependencies and contextual information. Meanwhile， we designed an auxiliary classifier to strengthen the mining ability of details in shallow features by integrating deep and shallow features， taking into account the effect of texture， edge， and other details in shallow features on surface defect classification. Finally， we implement simulation experiments on an open data set of aluminum profile surface defects to validate the efficacy of the proposed FWACNet method. The results show that FWACNet outperforms mainstream classification methods in the task of extreme aspect ratio and difficult classification of small area defects， with a classification accuracy of 95.7%.

Person Re-identification Based on Dual-branch Feature Concatenation

PAN Feng, WANG Jie, ZHANG Yan-sha, TAN Mian, HE Xing, WANG Lin,

2023, 0(05): 93-99.

Asbtract ( 169 )

PDF (1864KB) ( 34 )

References | Related Articles | Metrics

For the shooting of different monitoring vision， the person re-identification task is greatly misjudged within the intra-class （the same person）， and the ambiguity inter-class （similar persons） causes low differentiation. A method （Dual-branch Feature Concatenation Network， DFCNet） that integrates depth features is proposed in this paper， the deep features of the network are assembled to complement the feature information and obtain discrimination feature， and the BN layer replaces the full connection layer after the global average pooling layer of backbone network， the network is trained with label smoothing cross-entropy loss function， which solves the problem of within the intra-class changes and the ambiguity inter-class with low differentiation. To verify the effectiveness of the proposed method， the validation was performed on the Market1501， DukeMTMC-reID public datasets， which Rank-1 and mAP can reach 95.8% and 94.3% on Market1501. The proposed method has good performance in dealing with intra-class miscalculation and inter-class difficulty discrimination， and the recognition accuracy outperforms the state-of-the-art algorithms of comparison.

Small Object Detection Method Based on Improved YOLOv5

WANG Yi-cheng, ZHANG Guo-liang, ZHANG Zi-jie,

2023, 0(05): 100-105.

Asbtract ( 369 )

PDF (2253KB) ( 78 )

References | Related Articles | Metrics

In order to solve the problems of low detection accuracy and missing detection in traditional YOLOv5 object detection algorithm， a small object detection method based on improved YOLOv5 was proposed. Firstly， to make anchor box better adapt to small targets， IOU （interp over union） is used to replace the Euclidean distance formula originally used in the K-means clustering process to redefine the distance between anchor box and ground truth. Secondly， a maximum pooling of 3×3 kernel size is added to spatial pyarmid pooling （SPP） to improve the detection accuracy of small targets. Finally， a data set containing a variety of small object is designed to verify the algorithm performance. Experimental results show that the mean average precision （mAP） of the improved YOLOv5 algorithm reaches 76.92%， which is 3.56 percentage points higher than that of the classical YOLOV5 algorithm. The detection performence is improved and missed object can be detected.

Digital Identification of Electric Meter Based on Image Threshold Optimization and Improved SVM

YIN Jian-feng, WEI Xin, GU Xiong-wei, HUANG Kai, WEI Min-jie

2023, 0(05): 106-110.

Asbtract ( 195 )

PDF (1509KB) ( 38 )

References | Related Articles | Metrics

Aiming at the calibration of the electric meter and the test work in extreme environments， it is still necessary to manually detect whether the electric meter has internal component fault or errors. A research method of electric meter digital recognition based on image threshold optimization and improved SVM is proposed. First， we use edge search to obtain the display area of the image， use adaptive threshold for binarization， and then perform a series of filtering processing on the image， and then further extract the image of a single number， combined with image threshold optimization， before retaining the digital image. On the premise of eigenvalues， the redundant eigenvalues are removed， and the display area image is divided into several single digital images. Finally， based on the improved SVM multi-class recognition model， each digit from 0 to 9 is trained， and the trained model is used to identify the single digit image in turn. The experimental results show that compared with the classical convolutional neural network model for the recognition of LED liquid crystal digits， the optimization and improvement of the SVM model based on the image threshold have faster recognition speed and higher accuracy.

Segmentation Method of Knee Meniscus Based on Multiscale-net

WANG Juan, LI Chuan-geng, ZHANG Qing-yuan, XIA Cheng-yi

2023, 0(05): 111-116.

Asbtract ( 179 )

PDF (1641KB) ( 37 )

References | Related Articles | Metrics

The accuracy of knee meniscus segmentation was of great significance to the discrimination and diagnosis of meniscus tear grade. In order to improve the segmentation accuracy， this paper proposed a knee meniscal segmentation method based on Multiscale-Net network. This method combined the convolution layer and pooling layer of visual geometry group network16 and the decoder part of U-Net， and it replaced the 3×3 convolution layer connected with the encoder and decoder with an improved atrous spatial pyramid pooling module. Finally， it was verified on the real data set of clinical patients provided by the first affiliated hospital of Anhui medical university and compared with U-Net， U-Net with ASPP module introduced， and other models. The experimental results showed that the intersection over union and dice similarity coefficient of this method reached 91.25% and 94.89% respectively.

Vehicle Detection of Remote Sensing Images Based on Improved YOLOv5 Algorithm

ZHU Li-qing, LI Xiang,

2023, 0(05): 117-121.

Asbtract ( 277 )

PDF (2658KB) ( 56 )

References | Related Articles | Metrics

An improved model based on YOLOv5s is proposed for the problem of target miss detection in remote sensing images with complex targets in the background and blurred imaging due to small vehicles. A new backbone network structure is designed for the improved model: RepVGG network is selected for the backbone feature extraction of the improved model， while an attention mechanism， CoordAttention， is added to the backbone network to improve the perception capability of the model for small targets. Multi-scale feature fusion is added to improve the detection accuracy of the improved model for small targets， and the loss function of border regression is chosen to use DIoU to help the improved model achieve more accurate localization. After experiments， it is demonstrated that the improved YOLOv5 model improves the detection accuracy by 5.3 percentage points for target detection in remote sensing images compared to the original model in small target vehicles， and improves the mAP by 16.88 percentage points compared to Faster R-CNN. The improved model can have a larger detection accuracy improvement compared with the mainstream detection algorithms， and has a better detection accuracy than the original YOLOv5s model for small vehicle detection in remote sensing images.

Improved YOLOv5s for Small Vehicle Object Detection on Remote Sensing Image

QIU Di-fa, YU Shu-fang, LIU Jin-hui, BI Meng-zhao

2023, 0(05): 122-126.

Asbtract ( 281 )

PDF (1941KB) ( 61 )

References | Related Articles | Metrics

Because of its good detection effect and low computational complexity， YOLOv5s is widely used in various target detection tasks. However， its large downsampling stride makes it difficult to obtain satisfactory results for small-sized vehicle detection in satellite remote sensing images. In order to improve the performance of small target detection， on the basis of YOLOv5s， the strategy of reducing the downsampling stride is adopted to protect the texture and geometric features of small targets in the vehicle， and the attention mechanism module is inserted in front of the detection head to suppress the interference of complex background. The tested results on the autonomous data set with a resolution of 0.5 m/pixel show that the AP， recall and precision of SA-YOLOv5s for vehicle target detection reached 94.1%， 99% and 87.3% respectively， which were 16.4， 6 and 5 percentage points higher than YOLOv5s， and showed good detection performance.

Table of Content