Table of Content

19 February 2024, Volume 0 Issue 02

Previous Issue   

A Fast Registration Method for Massive Point Clouds Based on 3D-SIFT and 4PCS

LI Jia-le1, LI Zhe-run1, ZHAO Yong2, ZHANG Yang1

2024, 0(02):  1-6.  doi:10.3969/j.issn.1006-2475.2024.02.001

Asbtract ( 109 )   PDF (1952KB) ( 92 )  

References | Related Articles | Metrics

Abstract： The registration of measurement point cloud and model point cloud is the key of visual positioning. Aiming at the problems of poor visual positioning accuracy and low algorithm efficiency caused by large amount of measurement point cloud data and low overlap rate with CAD model point cloud， a registration method of measurement point cloud and model point cloud based on the fusion of 3D scale invariant feature transform （3D-SIFT） and four point fast robust matching algorithm （4PCS） is proposed. Firstly， the depth camera is used to extract the point cloud of the part， and the extracted measurement point cloud is denoised and filtered； Then 3D-SIFT feature point extraction algorithm is used to extract feature points from measurement point cloud and CAD model point cloud； Finally， the extracted feature points are used as the initial values of the 4PCS algorithm to achieve the registration of the two point cloud data. Compared with the commonly used 4PCS algorithm and Super-4PCS algorithm， the algorithm simulation and experimental results show that the proposed algorithm can improve the registration speed by more than 30% on the premise of ensuring the registration accuracy.

Adaptive Bald Eagle Search Algorithm Embedded with Somersault Foraging and Application

XIA Huang-zhi1, 2, CHEN Li-min3, MAO Xue-di1, 2, QI Fu1, 2

2024, 0(02):  7-14.  doi:10.3969/j.issn.1006-2475.2024.02.002

Asbtract ( 20 )   PDF (4952KB) ( 43 )  

References | Related Articles | Metrics

Abstract： An improved bald eagle search algorithm is proposed to address the problems that the bald eagle search （BES） algorithm is easy to slip into local optimum and low solution accuracy. Firstly， a Circle chaotic map is used in place of the original algorithm’s randomly generated initial population to increase the initial population’s diversity. Secondly， in the search selection space phase， adaptive weight is combined to update the bald eagle individual position and balance the search and development ability of the algorithm. Finally， the elite differential variation is fused with a somersault foraging strategy and is used to update the positions generated by bald eagle leader individuals in the subsequent stages. The ability of the algorithm to jump out of local optimum is improved. The method underwent comparative simulation tests in some standard test functions， and the Random Forest classification parameters were optimized using the suggested strategy in this research. The experimental results demonstrate that the improved algorithm outperforms the conventional algorithm in terms of solution efficiency， solution accuracy， and classification accuracy.


Improved Algorithm for Keypoints Detection of Hip Based on U-Net

CHEN Zhen1, YAO Jing-hui2, SU Cheng-yue1

2024, 0(02):  15-19.  doi:10.3969/j.issn.1006-2475.2024.02.003

Asbtract ( 34 )   PDF (1367KB) ( 91 )  

References | Related Articles | Metrics

Abstract： The diagnosis of developmental dysplasia of the hip （DDH） using pelvic X-ray requires accurate mapping of hip key points， and deep learning methods can be used as reliable auxiliary tools. In order to solve the problem of diversified shooting posture and shooting distance for pelvic radiographs， this paper proposed RKD-UNet based on U-Net to detect keypoints of the hip. The model used residual blocks to improve U-Net’s convolution layers and skip-connection paths， as well as introduced the coordinate attention module into the encoder to enhance feature extraction ability for the keypoints neighborhood. Convolution layers and ASPP module were used on top of the encoder to form a Bridge block to fuse feature information at different scales and enhance the receptive field of the model with an atrous rate of ［3， 6， 9］. The model was trained and tested using radiographic data containing types of pelvic orthostasis， frog， full-length lower extremity， and postoperative pelvis. RKD-UNet achieves an average keypoints detection error of 3.19 ± 2.19 px and an average acetabular angle measurement error of 2.83°± 2.59°. The F1 score for the normal， mild， moderate， and severe dislocation cases were 89.6， 77.1， 57.9， and 94.1， respectively， which were higher than the doctors’ diagnostic results. Experiments have shown that RKD-UNet can accurately detect keypoints of the hip and assist doctors in diagnosing DDH.

Front-view Sonar Imaging Method Based on Sparse Reconstruction

XU Yun-yan1, ZHENG Wei2, LIU Jian-guo3, BI Yang4, GUO Tuo1

2024, 0(02):  20-28.  doi:10.3969/j.issn.1006-2475.2024.02.004

Asbtract ( 23 )   PDF (9653KB) ( 25 )  

References | Related Articles | Metrics

Abstract：The DOA estimation algorithm based on sparse reconstruction can obtain higher-resolution spatial spectrum estimation by strengthening the sparsity of the representation， which is helpful to realize the differentiation of adjacent targets， and a sonar imaging method with sparse reconstruction at each distance is proposed. This method uses the sparsity of the target itself in sonar imaging and the norm constraint in the sparse reconstruction algorithm to obtain higher resolution and ultimately achieve the improvement of imaging effect. In the simulation and pool experiments， the performance of l1-SVD and SpSF sparse reconstruction algorithms is compared with the traditional azimuth estimation methods MUSIC， CBF， SFW-L21 and NN-SpSF， and the experimental results show that the l1-SVD algorithm and SpSF algorithm are better than the traditional methods， with narrower main lobes and lower side lobes， and have a certain suppression effect on background noise. At the same time， two targets that are close to each other can also be well distinguished， indicating a higher resolution.

Design Scheme of User Clustering and Power Distribution for Millimeter-Wave Massive#br# MIMO-NOMA Systems#br#

LI Wang-wang, HUANG Xue-jun

2024, 0(02):  29-35.  doi:10.3969/j.issn.1006-2475.2024.02.005

Asbtract ( 18 )   PDF (2532KB) ( 29 )  

References | Related Articles | Metrics

Abstract： To solve the problem of high computational complexity in millimeter wave massive multiple-input multiple-output non-orthogonal multiple access （MIMO-NOMA） systems， a scheme of user clustering and power distribution is proposed to improve the spectral efficiency. Firstly， user clustering scheme based on cluster-head selection is improved， in which the threshold and the number of clusters are determined dynamically according to the real channel. The clustering result is more suitable for the actual situation， and users can get greater gain from the beams. Then power allocation is designed with the goal of maximizing the weighted sum of spectral efficiency and energy efficiency of the system and solved by an improved meta-heuristic algorithm. By introducing new vector components to Particle Swarm Optimization （PSO） algorithm and adding cosine perturbation， the algorithm can converge to the global optimal value more quickly. Sand Cat Swarm Optimization （SCSO） algorithm is integrated to make the algorithm more accurate. The simulation results show that compared with the existing algorithms， the spectral efficiency and energy efficiency of the proposed scheme are better than the traditional schemes， and it is more suitable for multi-user cases.

A Partition Inverted Index Compression Algorithm Based on CRF

WANG Zi-chen, QU You-li

2024, 0(02):  36-42.  doi:10.3969/j.issn.1006-2475.2024.02.006

Asbtract ( 12 )   PDF (1927KB) ( 36 )  

References | Related Articles | Metrics

Abstract： The inverted index is the core data structure of a large search engine， which is essentially a collection of integer sequences in an inverted table. Inverted index compression can effectively reduce the space occupied by inverted indexes and improve retrieval efficiency of keywords. The partition inverted index compression algorithm based on conditional random field （CRF） mainly focuses on the partition mode of universe partition. The algorithm pre-partitions the sequence and uses conditional random fields to label and reorganize the pre-partitions， which effectively reduces the compression time. According to the partition type， the algorithm uses the corresponding encoding method to further reduce the space occupation after compression. Compared with other inverted index compression algorithms， this algorithm outperforms some current universe partition algorithms in compression ratio， and is equivalent to other universes partition algorithms in decompression time. The algorithm achieves a good balance between time and space.

Few-shot Algorithm for Object Detection in Remote Sensing Images

XUE Yang-yi1, ZHOU Li-fan2, GONG Sheng-rong1, 2

2024, 0(02):  43-49.  doi:10.3969/j.issn.1006-2475.2024.02.007

Asbtract ( 15 )   PDF (2327KB) ( 35 )  

Related Articles | Metrics

Abstract：In view of the lack of remote sensing scene data, the obvious size change of surface objects captured by aerial photography, including a large number of objects of multiple categories and complex background, resulting in low detection accuracy and inaccurate classification, a small sample remote sensing target detection network based on the two-stage detection model (Faster RCNN) is proposed. New involution convolution operators are added to build detector backbone to improve feature extraction capability; Integrate multi-scale object-level positive sample features to enhance the original features, suppress the adverse effects of negative samples, fully mine the feature information of each target scale, and help the semantic information to locate; The idea of comparative supervision is adopted to improve the loss function, refine the target classification and reduce the false detection rate. The experimental results on public remote sensing data sets show that the network can adapt to the multi-scale characteristics of remote sensing images and effectively alleviate the over-fitting phenomenon caused by data scarcity under the condition of only a small number of remote sensing labeled samples. Compared with the previous Meta RCNN and FsDet networks, the average accuracy has been further improved by 3.8 percentage points and 2.5 percentage points, providing a meaningful reference for image target detection in the remote sensing field.


Multi-stage Rain Removal Algorithm Based on Multi-scale Frequency Attention

WU Tian-tian, LI Yan-kai, LIU Yang

2024, 0(02):  50-55.  doi:10.3969/j.issn.1006-2475.2024.02.008

Asbtract ( 21 )   PDF (12157KB) ( 52 )  

References | Related Articles | Metrics

Abstract：Rain streaks interfere with the photos the outside vision system takes in rainy weather， which caused lowering the image quality and affecting the subsequent vision tasks. Therefore， for the following computer vision tasks， it is very crucial to remove the rain streaks from the photos and get high-quality images. The goal of the multi-stage rain removal method we present in this research is to recover a high quality image by removing rain streaks from a single rain image using multi-scale frequency attention. Firstly， a multi-stage rain removal model is designed by integrating the variety of rain streaks， decomposing the rain streaks removal process into multiple sub-processes， and eliminating rain streaks step by step. Second， a long and short-term memory recurrent network is improved to achieve multi-stage rain streaks removal， in which the frequency attention mechanism is introduced to strengthen the attention to rain streaks and a multi-scale feature extraction method is designed to characterize the global information. This addresses the issue of oversmoothing in the current rain streaks removal algorithms. The detail restoration module’s final purpose is to fortify background elements. Experiment results show that the proposed algorithm can effectively remove rain streaks on both the synthetic data set and the real dataset while preserving complete background information， and has a good rain removal effect.


Image Dehazing Algorithm with Improved Generative Adversarial Network

LIU Yan-hong, YANG Qiu-xiang

2024, 0(02):  56-63.  doi:10.3969/j.issn.1006-2475.2024.02.009

Asbtract ( 25 )   PDF (4948KB) ( 50 )  

References | Related Articles | Metrics

Abstract： In hazy weather， visible light scattering and absorption occur when passing through the atmosphere， resulting in poor image quality， information blocking or loss. Based on this， we propose an improved generative adversarial network （GAN） image dehazing algorithm， which learns to generate dehazed images in the generator and discriminator adversarial. In the generator， a three-row multi-column multi-scale fused attention network （Grid-G） is proposed to introduce channel attention and pixel attention to process the thick haze region and high frequency region of the image from different angles， respectively. In the discriminator， the high and low frequency information in the image is introduced to construct the fused discriminator （FD-F）， which is used as a source of additional a priori discriminative images. Experiments on synthetic and real data in the RESIDE dataset show that the algorithm outperforms the rest of the comparison algorithms in terms of peak signal-to-noise ratio and structural similarity， achieves better dehazing effects， and effectively improves problems such as color distortion.


Optical Flow Estimation Based on Inverse Residual Attention

LIANG Jian-ye1, CHEN Jun-hong1, FANG Gui-biao1, WU Xing-cai2, LIU Wen-yin1

2024, 0(02):  64-68.  doi:10.3969/j.issn.1006-2475.2024.02.010

Asbtract ( 28 )   PDF (1600KB) ( 22 )  

References | Related Articles | Metrics

Abstract: Optical flow estimation is a basic task of video understanding and analysis. Many existing methods directly take occlusion as the outer point and eliminate it， so as to improve the ability of the model to calculate the optical flow， but it is also easy to cause the image gray discontinuity， leading to the failure of optical flow estimation. In addition， the problem of large displacement caused by high speed motion of objects has always been a difficulty in optical flow estimation. In order to solve the above problems， this paper proposes a generative adversarial learning framework based on reverse residual attention （FlowTranGAN， FTGAN） for optical flow estimation. The proposed framework enhances the spatial information of features by designing a reverse residual attention module to improve the matching degree between pixels. Besides， we use a discriminator based on U-Net to constrain the generator to reduce the error and discontinuity of optical flow estimation， and improve the generalization ability of the model. Experiment results on the KITTI-2015 dataset and MPI-Sintel dataset demonstrate the effectiveness and superiority of the proposed FTGAN.


Nested Named Entity Recognition Based on Semantic Segmentation

CUI Shao-guo, HU Guang-ping

2024, 0(02):  69-74.  doi:10.3969/j.issn.1006-2475.2024.02.011

Asbtract ( 23 )   PDF (1307KB) ( 66 )  

References | Related Articles | Metrics

Abstract： Named entity recognition aims to extract entities from an unstructured text， and a nested structure often exists between entities. However， most of the previous studies only focused on the recognition of flat named entities while ignoring nested entities. Therefore， a nested named entity recognition method based on semantic segmentation is proposed， which describes the task of nested named entity recognition as a semantic segmentation task. First， we calculate the element similarity， cosine similarity and bilinear similarity between words and words. Then， the 3 similarity features are spliced as an image which will be input into the semantic segmentation model to obtain the relationship matrix between words and words. Finally， we extract nested entity from the relationship matrix. The experimental results show that the proposed method can effectively recognize nested entities， and the F1 value on the public nested named entity recognition dataset GENIA reaches 80.0%， which is superior to most existing nested entity recognition methods.


Anomalous Behavior Detection Network Based on Dilated Convolution and Fused Temporal#br# Features

MA Cai-sha, JIAO Li-nan, LIU You-quan, LI Xin

2024, 0(02):  75-80.  doi:10.3969/j.issn.1006-2475.2024.02.012

Asbtract ( 12 )   PDF (2029KB) ( 61 )  

References | Related Articles | Metrics

Abstract: In this paper， we propose a multi-scale deep autoencoder network based on dilated convolution， incorporating pedestrian prototypes and spatio-temporal features. To better exploit the temporal features of pedestrians in videos， a dual-branch structure is added to the potential space of the encoder and decoder， the ST-RNN branch of the recurrent neural network for predicting spatio-temporal features and the memory storage module for preserving the normal patterns of pedestrians. To enhance pedestrian feature extraction， ignore the effect of background information，and improve the generalization ability of the model， an improved atrous spatial pyramid pooling （ASPP） module is added in the encoder， the hybrid dilated convolution （HDC） principle is used in the convolution block to solve the pedestrian size variation problem， while a multi-level residual channel attention mechanism is introduced in the decoder to obtain more contextual information. The corresponding area under the ROC curve （AUC） of this model reaches 0.982， 0.928 for USCD ped2， CUHK Avenue datasets， respectively.


Image Classification of COVID-19 Based on Contrast Learning MocoV2

XU Yue-wen1, LI Ming1, LI Li2

2024, 0(02):  81-87.  doi:10.3969/j.issn.1006-2475.2024.02.013

Asbtract ( 18 )   PDF (3940KB) ( 57 )  

References | Related Articles | Metrics

Abstract: Pneumonia is a common multi-infectious disease that predisposes the elderly and those with weakened immune systems to infection， and early detection can help with later treatment. Factors such as the location， density and clarity of lung lesions can affect the accuracy of pneumonia image classification. With the development of deep learning， convolutional neural network is widely used in medical image classification tasks， however， the learning ability of the network depends on the number of training samples and labels. Aiming at the classification of pneumonia images in computed tomography （CT）， a network model based on self-supervised comparative learning （MCLSE） is proposed， which can learn features from unmarked data and improve the accuracy of the network model. Firsly， auxiliary tasks were designed to mine representations from unmarked images to complete pre-training， improving the ability of the model to learn data mapping relationships in vector space. Secondly， the convolutional neural network is used to extract features. In order to effectively capture higher level feature information， the compression excitation network is selected to improve the classification model and the correlation between the feature channels is modeled. Finally， the trained weights are loaded into the improved classification model， and the network is trained again with marked data in the downstream task. Experiments were carried out on open data sets， SARS-CoV-2 CT and CT Scan for COVID-19 Classification. The results show that the accuracy of the MCLSE model in this paper for the overall sample classification reached 99.19% and 99.75%， respectively， which was greatly improved compared with the mainstream model.


Lightweight PCB Defect Detection Method Based on Multi-scale Features and Attention Mechanism

ZHOU Yong-qin, WANG Yong, WANG Ying

2024, 0(02):  88-92.  doi:10.3969/j.issn.1006-2475.2024.02.014

Asbtract ( 21 )   PDF (2101KB) ( 37 )  

References | Related Articles | Metrics

Abstract： To address the issues of surface background interference in PCB defect detection and resource consumption caused by large detection models， a lightweight network model called SL-Unet is proposed for effective extraction of multi-scale and edge information in PCB surface defect detection. SL-Unet utilizes the U-Net structure as the backbone network. Firstly， the U-shaped residual structure is used to capture multi-scale information in each dimension of the backbone network， strengthening the communication between shallow and deep information， and introducing the DropBlock module to improve the model’s generalization ability. Secondly， the edge information of the decoder is used to complete deep supervision， and a lightweight channel attention module is incorporated to enhance the feature dependence of edge information， guiding the backbone network to perceive the edge information of defects when extracting features. Then， a multi-level joint loss is constructed through the edge-aware module for the optimization of the overall model. Finally， the Leaky ReLU function is used to replace the ReLU function in the network， improving the model’s feature extraction ability in the negative interval. Experimental results show that the Dice coefficient， intersection over union， image detection frame rate and model size indicators of SL-Unet reach 79.3%， 67.4%， 22 frames/s and 3.46MB， respectively， greatly ensuring the lightweight of the model and significantly improving the detection accuracy of PCB surface defect images.


Lightweight YOLOv4-based Target Detection Method for Remote Sensing Images of#br# Airport Fields

YANG Ke, DONG Bing, WU Yue, HAO Kuan-gong, PENG Zi-chen

2024, 0(02):  93-99.  doi:10.3969/j.issn.1006-2475.2024.02.015

Asbtract ( 21 )   PDF (4815KB) ( 62 )  

References | Related Articles | Metrics

Abstract： Aiming at the problems that existing remote sensing image target detection methods suffer from the loss of local feature information in deep CNNs and low detection accuracy of complex scenes， a target detection method based on lightweight YOLOv4 is proposed. Firstly， the lightweight neural network Ghostnet is used to replace the cspdarknet53 network used as the backbone feature extraction in YOLOv4. Secondly， to improve the complex environment detection capability， CycleGAN is used to simulate night scenes， and again， the transformer module is fused to make the model easy to capture inter-feature relationships and local information of the network. Finally， Adam optimiser and K-means++ screening anchor frame are used to accelerate the convergence speed， and the example is validated with RSOD aerial remote sensing dataset. The experimental results show that the MAP value is improved by 6.65 percentage points and the number of parameters is reduced by 84.7% compared with the original YOLOv4， i.e. the algorithm in this paper can meet the real-time target detection of aircraft on the airport field in complex scenes.


Multi-view Knowledge-aware Recommender System

WANG Xiao-xia, MENG Jia-na, JIANG Feng, DING Zi-qing

2024, 0(02):  100-107.  doi:10.3969/j.issn.1006-2475.2024.02.016

Asbtract ( 16 )   PDF (2064KB) ( 77 )  

References | Related Articles | Metrics

Abstract： At present， most of the recommendation methods based on knowledge graph use single user or item representation， which has the problems of user interest interference， incomplete use of information and sparse data. This paper proposes a multi-view knowledge-aware recommendation model （MVKA）. Firstly， the model captures the user’s interest representation in the user-item graph fusion attention mechanism. Introduce the project-entity diagram， the graph attention network is used for feature extraction to obtain the embedded representation of the item. Then， a comparative learning method of graph perspective is constructed between the two views. Finally， summation and concatenation operations are carried out to get the final representation of the user and the project， and the matching score of the user to the project is predicted by the inner product. In order to verify the accuracy and computational efficiency of the experiment， a large number of experiments were carried out on the three public datasets of MovieLens-1M， Book-crossing and Last FM， and compared with other traditional methods and graph neural network models， the AUC and F1 value evaluation indicators were significantly improved， indicating that the MVKA model can significantly use various information relationship data to improve the knowledge perception recommendation task.


Confidence Evaluation of Reliability of Multivariate Degenerate System Considering#br# Different Strength Dependence

MIAO Si-qiao, FAN Hong-mei, YUAN Fei-meng

2024, 0(02):  108-113.  doi:10.3969/j.issn.1006-2475.2024.02.017

Asbtract ( 13 )   PDF (982KB) ( 36 )  

References | Related Articles | Metrics

Abstract： For aero-engine operating systems with multiple performance degradation modes， a reliability confidence evaluation method based on random correlation is proposed considering individual differences and different strength correlations. The Wiener random process and Gamma random process with random effects were used to describe the performance degradation failure processes respectively. The Copula function of the randomization of relevant parameters was used to model the degree of dependence of multiple performance parameters. An analytical expression of the boundary point of the confidence interval based on the Clayton Copula function model is derived， and an overall reliability evaluation model is established by using the edge function inference method to estimate the unknown parameters in the model through two-step optimization. Combined with the performance degradation data of aero-engine EGTM and ZVB2R， the overall reliability model was established， and the comprehensive evaluation was completed. The life interval was obtained as（1.033×104，1.278×104） cycles. The feasibility and accuracy of the model were verified by an example.



A Data-Driven Intelligent Analysis Platform for Ion Source Data

XIONG Qing-zhi1, LI Xiang1, 2, PENG Fang-wei1, JIN An-an1

2024, 0(02):  121-126.  doi:10.3969/j.issn.1006-2475.2024.02.019

Asbtract ( 21 )   PDF (2431KB) ( 36 )  

References | Related Articles | Metrics

Abstract： Food safety is a matter of great significance concerning people’s health and safety. Food and drug residue detection technology is a crucial means to ensure food safety. Ion source data encapsulation and processing technology constitute a bottleneck affecting the quality of food and drug residue detection. Currently， due to restrictions in foreign software functions， ion source data can only be accessed on specific software platforms with limited operations. Addressing the issues related to ion source offline mode data acquisition and processing methods， this paper presents a data-driven intelligent ion source analysis platform. A simple， efficient and accurate intelligent data processing and analysis platform has been built. It enables rapid data acquisition and processing， breaking down technological barriers imposed by foreign software. This proposed solution offers a new solution， particularly in updating and improving domestic ion source equipment.