• 北大核心期刊(《中文核心期刊要目总览》来源期刊)
  • 中国科技核心期刊(中国科技论文统计源期刊)
  • JST 日本科学技术振兴机构数据库(日)收录期刊

NoticeMore>

Special TopicMore >

Download CenterMore >

Wechat

Articles just accepted have been peer-reviewed and accepted, which are not yet assigned to volumes /issues, but are citable by Digital Object Identifier (DOI).
Display Method:
Articles in press have been peer-reviewed and accepted, which are not yet assigned to volumes /issues, but are citable by Digital Object Identifier (DOI).
Display Method:
Display Method:
A review of behavior recognition methods based on deep learning
YUAN Shou, QIAO Yongjun, SU Hang, CHEN Qinghua, LIU Xing
2022, 39(8): 1-10.   doi: 10.19304/J.ISSN1000-7180.2021.1327
Abstract(0) HTML(0) PDF (0)
Abstract:

As a research hotspot in the field of computer vision, behavior recognition has been widely applied in intelligent security, intelligent monitoring, intelligent medical treatment and other fields in today's society, and the application of deep learning, which has outstanding performance in computer vision, has a more significant effect on behavior recognition research. Compared with the traditional method based on manual feature extraction, the behavior recognition method based on deep learning has the advantages of fast speed, strong robustness and high accuracy. Therefore, this paper summarizes the video behavior recognition method based on deep learning. At home and abroad based on the latest published paper summarizes relevant literature, firstly analyzes the traditional behavior recognition method and the corresponding improvement points, according to different network architecture detailed carding behavior recognition method based on the deep study, then study the comparison of common recognition data sets and the performance of the algorithm in the data set is comparing the advantages and disadvantages, finally summarizes the research of this field, This paper focuses on the existing problems and looks into the future, hoping to enlighten and help researchers.

Knowledge graph double perception network for recommendation algorithm
HAN Chen, YANG Xingyao, YU Jiong, GUO Liang, HU Haoyu
2022, 39(8): 11-20.   doi: 10.19304/J.ISSN1000-7180.2022.0096
Abstract(0) HTML(0) PDF (0)
Abstract:

In recent years, excellent results have been achieved by aggregating the additional item information in the knowledge graph, but there are relatively few sources of user information. At the same time, multiple aggregation will make the expression of the characteristics of the item incomplete and even produce noise. Aiming at the above two points, a KGDP recommendation algorithm based on knowledge graph is proposed. Firstly, some items are randomly selected from user interaction records as user related items, and the neighbor entities of the items are selected as item related entities; Then, the selected user related items are fused into user features through deep neural network, which enriches user features and aggregates the related entities of the items separately; Secondly, through two deep neural networks, users can perceive item characteristics and neighbor characteristics respectively, that is, non-linear interaction. Finally, a single-layer perceptron is used to adjust the output weight of interactive features for score prediction. Experiments on two real datasets commonly used in recommendation algorithm, compared with the baseline model, the AUC index improved by 9.2% and 2.4% respectively; ACC index improved by 6.6% and 1.9%; F1 index improved by 7.0% and 1.1% respectively; Precision@N index improved by 28.8% and 6.5% respectively; Recall@N index improved by 4.0% and 23.7% respectively; F1@N index improved by 43.3% and 8.4% respectively.

Restore local descriptors network for few-shot learning
WANG Ronggui, WANG Wei, YANG Juan, XUE Lixia
2022, 39(8): 21-30.   doi: 10.19304/J.ISSN1000-7180.2022.0107
Abstract(0) HTML(0) PDF (0)
Abstract:

Aiming at the problem that the existing few-shot metric learning methods based on local descriptors fail to consider the correlation between local descriptors and fail to make full use of the global feature information of categories, this paper proposes the Restore Local Descriptors Network (RLDN in short). The adjacent GCN module increases the relevance between local descriptors by using the spatial position relation in the same image. The global feature extraction module outputs the global descriptors of the category by learning and fusing the global features of the image, and then concatenates the local descriptors for further restoration. In addition, a new hybrid loss function is proposed by introducing the triple loss which is integrated into the traditional cross entropy loss. It increases the distance between different categories and helps the classifier to reduce the misclassification. The experimental results show that compared with the traditional local descriptor methods, the Restore Local Descriptors Network can reduce the interference of noise features on the classifier and effectively improve the classification accuracy of the model.

Abnormal behavior detection based on attention-generative adversarial network
WU Lijun, CHEN Shidong, CHEN Zhichong
2022, 39(8): 31-38.   doi: 10.19304/J.ISSN1000-7180.2022.0065
Abstract(0) HTML(0) PDF (0)
Abstract:

To meet the needs of abnormal behavior detection for large-scale video data, methods based on video frame reconstruction and frame prediction have been widely studied. However, because the background environment is almost constant under the monitoring perspective, a lot of resources will be wasted on the constant background, and it is also not conducive to extracting the detection target information. In order to solve this problem, this paper uses an unsupervised learning video frame prediction strategy, and uses generative adversarial networks to learn features of normal behavior to generate better predicted frames. And the attention-driven loss is used to alleviate the problem of the imbalance between the foreground target and the background environment in abnormal behavior detection, and the spatial-channel attention mechanism (CBAM) is used to enhance the prediction effect of the model generator.After the test and verification of public data sets UCSD Ped1 and UCSD Ped2, the detection accuracy on the Ped1 dataset has reached 83.5%, and the detection accuracy on the Ped2 dataset has reached 95.8%.And compared with the classic abnormal behavior detection algorithm and the original generative adversarial network based anomaly detection algorithm, the method adopted in this paper further improves the accuracy of abnormal behavior detection.

Engineering vehicle detection in aerial images with recursive feature fusion and parallel scaling
MA Xuesen, CHU Zhaokun, MA Ji
2022, 39(8): 39-46.   doi: 10.19304/J.ISSN1000-7180.2022.0109
Abstract(0) HTML(0) PDF (0)
Abstract:

Aiming at the problems of poor detection accuracy caused by complex and changeable backgrounds, small objects and large scale changes in UAV aerial photography transmission corridor images, the detection method of recursive fusion and parallel scaling for engineering vehicles based on RetinaNet is proposed in this paper. This method is more suitable for detecting engineering vehicles in complex backgrounds. Firstly, the C2 layer is added as the base layer, which is used to generate feature pyramids together with the original backbone output layers to avoid small object features being highly compressed. Secondly, the original feature pyramid structure is adjusted, and recursive structure with feedback connections is used for feature extraction to enhance the characterization ability. Moreover, a novel and lightweight feature fusion strategy is designed to reconstruct the feature pyramid and makes full use of contextual information to improve the object detection capability in complex backgrounds. Finally, the parallel feature scaling branch is constructed with multiple deconvolution blocks and average pooling layers based on the C5 layer of the backbone to further increase the resolution of the feature maps and improve the detection accuracy of small objects. Experiments are carried out on the engineering vehicle APEV dataset constructed in this paper and the public Pascal VOC dataset, respectively. The experimental results show that the detection accuracy of the proposed method on the APEV dataset and the VOC dataset is 4.9% and 2.7% higher than that of the original RetinaNet network on the premise of meeting the requirements of engineering applications, respectively, Further, the proposed method also has higher detection accuracy compared with Faster R-CNN, SSD, YOLOv3, YOLOv5, LSN, S-RetinaNet and other methods.

Road extraction based on SAR polarization characteristics and SVM
SU Xiaojie, LIU Xiuqing
2022, 39(8): 47-54.   doi: 10.19304/J.ISSN1000-7180.2021.1368
Abstract(0) HTML(0) PDF (0)
Abstract:

Road extraction from SAR images has important application value in road network planning and construction, disaster monitoring and other fields. The traditional SAR image road extraction method is mainly based on the amplitude characteristics of SAR image, which lacks the interpretation of polarization characteristics. In addition, the polarization decomposition methods are mostly used in water extraction, land classification, building extraction, etc., and are less used in road extraction. Aiming at the problems of high data quality requirements of existing road extraction methods, less research on full polarization road extraction and large influence of speckle noise of full polarization data source, this paper first carries out multi view processing and filtering denoising preprocessing on the full polarization data, and obtains 20-dimensional polarization characteristic scattering components through polarization decomposition method. secondly, from the perspective of scattering mechanism, the optimal polarization feature vector for identifying road information is constructed. Finally, the preliminary road extraction results are obtained by SVM classifier, and road data is extracted by mathematical morphology method. Experimental results show that the proposed method achieves 98.4% Acc and 65.3% Iou, which can be effectively applied to the research of road extraction in SAR images.proposed which making full use of the polarizatio information of high-resolution SAR data, has the advantages of high extraction accuracy and wide application range, it can be effectively applied to the research of road extraction method of SAR image. In addition, different from applying the method of optical road extraction directly to the research of SAR image road extraction, this paper explores the application of polarization features in SAR image road extraction, and puts forward new modes and ideas for the research of SAR image road extraction.

Improved monocular visual inertial odometer based on VI-DSO
CHEN Mingda, YING Jun
2022, 39(8): 55-62.   doi: 10.19304/J.ISSN1000-7180.2021.1192
Abstract(0) HTML(0) PDF (0)
Abstract:

Aiming at the problems that the visual odometry of the pure direct method lacks scale information, it is easy to fall into the local optimum when optimizing the pose, and the convergence speed is slow when monocular vision is initialized, an improved direct method monocular visual inertial odometry scheme is proposed. Based on the direct method visual inertial odometer VI-DSO with better effect at present, a modification scheme is proposed. In the initialization, the VI-DSO scheme ignores the initialization of the IMU, and uses the back-end unified optimization method to estimate the IMU bias, which leads to the slow convergence of the scale and the large cumulative error. The fast estimation of the IMU bias and the MAP of the scale accelerates the speed of the scale convergence during the initialization, and also provides a more accurate initial data for the back-end optimization to reduce the cumulative error. In the depth estimation, the depth filtering scheme is improved. Referring to the SVO filtering method, the Gaussian-uniform filter is used to estimate the probability of mismatching, eliminate the wrong depth estimation, and fuse the correct depth data to improve the positioning accuracy. In the process of marginalization, the marginalization strategy of the VI-DSO scheme is improved, and the judgment of the current motion state is increased. The frames that need to be marginalized are selected according to the motion state to ensure sufficient parallax in the sliding window. The test results in EuRoc dataset show that the improved scheme improves the initialization speed by 33 % and the average positioning accuracy by 34.5 %.

Anti-occlusion correlation filter tracking method from a convolution perspective
LI Lihui, HUANG Yuming, DING Can, YU Fei, CHEN Yingpin
2022, 39(8): 63-70.   doi: 10.19304/J.ISSN1000-7180.2022.0081
Abstract(0) HTML(0) PDF (0)
Abstract:

Because the current correlation filter tracking algorithm is prone to tracking drift or even loss when the target rotates, moves fast, and is occluded, an anti-occlusion correlation filter tracking method is proposed under a convoluted perspective. Based on the framework of the correlation filtering algorithm, the method uses context-aware methods to increase background information. It introduces a multimodal history pool update strategy to enhance the anti-occlusion tracking performance. First, a formula derivation system based on the convolution perspective is designed, and the convoluted theorem is subtly introduced to solve the filter in the frequency domain. Compared with the filter solution method of circulant matrix diagonalization in the existing literature, the method is easy to understand. Then, by submitting context-related information, reasonable energy functional is designed to suppress the response value of the background region to achieve the purpose of tracking the target more robustly. Finally, a historical multimodal target pool is established. When the similarity between the sample with a significant relevant response and each multimodal template in the historical template pool is lower than the artificially set threshold, it is determined that the frame is occluded. In this situation, the template pool, appearance model and filter should not be updated, effectively solving the challenge of occlusion and track drift issues. The proposed method was tested on OTB2015. Experiments show that under the conditions of target rotation, fast movement, occlusion, etc., the proposed method can ensure accurate tracking while maintaining a high speed, which is better than other methods proposed in the experiment.

Iterative MPD algorithm based on transform domain
ZHOU Rui, LI Yingshan
2022, 39(8): 71-77.   doi: 10.19304/J.ISSN1000-7180.2021.1042
Abstract(0) HTML(0) PDF (0)
Abstract:

Compared with the traditional orthogonal frequency division multiplexing (OFDM) modulation scheme, filter OFDM (F-OFDM) enjoys the superiorities of flexible subcarrier bandwidth and reduced out-of-band emission, which makes it one of the key candidate waveforms for future wireless communication systems. However, under the scenarios with high mobility, the Doppler frequency shifts (DFOs) can seriously degrade the performance of F-OFDM systems. To cope with this problem, an iterative message passing detection (MPD) algorithm based on transform domain is proposed in this paper. The MPD algorithm is based on a sparse factor graph and utilizes iterative massage transmission and update to achieve the suppression of the DFOs. Besides, the interference in the detection can be modeled as a Gaussian variable, which has the benefit to reduce the complexity of the MPD. Furthermore, based on the transform domain, the proposed iterative MPD algorithm can make full use of the enhanced channel sparsity in the transform domain, which effectively decreases the branches between the transceiver nodes in the MPD, and further reduces the complexity of the algorithm. Finally, based on the F-OFDM system, the simulation results show than the proposed scheme enjoys an improvement over the computational complexity and also the system performance compared with the traditional counterpart.

Design of instruction control system for neural network accelerator
JIAO Feng, MA Yao, BI Siying, MA Zhong
2022, 39(8): 78-85.   doi: 10.19304/J.ISSN1000-7180.2021.1344
Abstract(0) HTML(0) PDF (0)
Abstract:

Deep neural networks are increasingly used in the field of intelligent processing of image and speech, however their multiple operator and parameter types, large computation and storage intensive characteristics restrict the application in embedded scenarios such as aerospace and mobile intelligent terminals. To address this problem, the concept of decoupling input data streams for efficient flowing parallel processing is proposed, and an instruction control system for a neural network accelerator is designed. After the input data of different operators are cyclically chunked, and corresponding to the instruction group configuration, multiple state machines collaborate to complete the three-stage distribution control of instruction information, realising four stages of parallel flow of instruction parsing, data input, computation and data output, fully utilising the data reuse possibilities within the chunks, so as to reduce the access bandwidth and flow cycle idle rate. Deployed on the ZCU102 development board, the test shows support for a variety of common neural network layer types and a wide range of parameter configurations. At a frequency of 200M, with a peak arithmetic power of 800 Gops and running the VGG16 network model, an actual test run of 489.4Gops and power consumption of 4.42W resulted in an energy efficiency ratio of 113.3GOPs/W, superior to similar neural network accelerators and CPUs and GPUs. Experimental results show that the method of decomposing data streams and using instruction scheduling to achieve efficient parallelism solves the two major challenges of generality and energy efficiency, the instruction control system based on this method, can provide a solution for the embedded platform application of neural network accelerators.

A DS-SoC for high speed digital inkjet printing
GAO Haohui, FAN Rong, MIAO Yongjie, CHAI Zhilei
2022, 39(8): 86-96.   doi: 10.19304/J.ISSN1000-7180.2022.0054
Abstract(0) HTML(0) PDF (0)
Abstract:

According to the requirements of high-speed digital inkjet printing for high bandwidth, low delay, low jitter print data transmission, and high concurrent bit operation orifice control, a domain specific SoC architecture is designed and the software and hardware system is implemented. The lightweight network transmission under StandaloneOS reduces the transmission jitter caused by the operating system and maintains stable and high-speed data transmission; Based on the on-chip high-speed bus, the transmission bandwidth between the main control system and the nozzle control module is improved and the delay of signal transmission is reduced; By designing the bit operation coprocessing module, the high concurrency control of orifice array is realized. The above SoC architecture and system are realized based on ZYNQ7020 FPGA SoC platform. The experimental results show that when the system receives print data from the host computer, the transmission rate of Gigabit Ethernet can be stably maintained at 947 Mbps; The on-chip bus transmission bandwidth can reach 800 MB/s, and the instruction transmission delay is within 10ns; The bit operation data processing frequency of the system can reach 64 MHz; The data throughput of the whole system can reach 1500 Mbit/s, and the data transmission jitter is within 20 ns. The system can drive the printer nozzle with 30720 spray holes to complete the printing work of 200 cm/s with the printing accuracy of 600 dpi. It still has good performance under the printing accuracy of 1200 dpi, which provides a new technical idea for breaking through the speed bottleneck of high-speed digital inkjet printing.

A clock synchronization system based on carrier bidirectional frequency transfer
WANG Yingzi, YAN Bing, LI Zhitian, ZOU Xudong
2022, 39(8): 97-106.   doi: 10.19304/J.ISSN1000-7180.2022.0026
Abstract(0) HTML(0) PDF (0)
Abstract:

With the rapid development of wireless sensor networks and technologies, the need for high-precision wireless time-frequency synchronization technology in the collaborative work of distributed systems is increasingly urgent. In response to the demand for high-precision wireless time-frequency synchronization in the line-of-sight distributed network under the condition of satellite rejection, this paper proposes a clock synchronization system scheme based on carrier bidirectional frequency transfer. It innovatively implements the mutual transmission of time-frequency information through millimeter-wave channels by deploying a full-duplex bidirectional time-frequency synchronization protocol, and introduces Xilinx MMCM IP dynamic phase shift function core to control the phase of time-frequency signals and realize frequency difference measurement and dynamic phase shift. This effectively improves the phase shift performance of the time-frequency synchronization architecture, and finally realizes a complete set of sub-nanosecond-level ultra-high-precision wireless time-frequency synchronization scheme. The technical architecture, RF front-end, wireless channel transmission and anti-jamming capability of the entire system are modeled and simulated in this paper, which verifies the effectiveness of the entire technical solution and the optimal phase shift accuracy. Besides, the 60 GHz RF front-end and Xilinx 7 series FPGA are also used to complete the principle prototype design. The experimental results have proved that the time-frequency synchronization system can provide wireless time-frequency mutual calibration service with a synchronization accuracy of up to 322.2 ps between nodes to achieve frequency synchronization and phase alignment, which can also support the development of various distributed collaborative work. Compared with traditional wireless synchronization methods, this scheme has higher precision, smaller influence from wireless channels and stronger anti-interference abilities, which is easy to extend to complex environments such as high dynamics and more suitable for wireless distributed networks.

Research on reading acceleration technology of embedded flash
YANG Yiwei, DU Junhui, HUANG Kaitian, KUANG Xiaoyun, WANG Ke
2022, 39(8): 107-118.   doi: 10.19304/J.ISSN1000-7180.2021.0948
Abstract(0) HTML(0) PDF (0)
Abstract:

Embedded Flash has increasingly become an important program and data memory in microcontroller due to its advantages in cost and storage density. However, the relatively slow read speed of embedded Flash restricts the overall performance of the microcontroller, so it is very important to improve the read performance of instruction and data in Flash. In order to improve the read performance of the embedded Flash in the microcontroller, a Flash controller based on cache and prefetch is proposed, and the current cache and prefetch are optimized. in view of the problem of poor adaptability of the existing cache, a cache line size adaptive technology is proposed for optimization. Aiming at the problem of high missing cost and high power consumption when accessing the set-associative cache in the traditional way, a way hit prediction technology is proposed for optimization. Aiming at the problem of low accuracy of the existing prefetching technology, a stride prefetching technology is proposed for optimization. Finally, an embedded Flash controller was designed and implemented, and integrated into the SoC system, and a verification platform was built for functional simulation and FPGA verification. Experimental results show that after adopting the cache line size adaptive technology, the performance of the processor to read embedded Flash is significantly improved (103%); after adopting the way hit prediction technology, the performance of the processor to read embedded Flash is further improved (2%). After adopting stride prefetching technology, the performance of DMA reading embedded Flash has been significantly improved (50%).

A high reliability, multi-bit parallel readout in-memory computing scheme based on STT-MRAM
YUAN Lei, CHEN Junjie, ZHUO Pengfu, WANG Shaohao
2022, 39(8): 119-126.   doi: 10.19304/J.ISSN1000-7180.2022.0079
Abstract(0) HTML(0) PDF (0)
Abstract:

In-memory computing technology is one of the most effective ways to solve the bottleneck of traditional von Neumann computing architecture. Although STT-MRAM has the advantages of non-volatile, low power consumption and high durability, its small sensing margin poses a challenge to the read reliability of sensitive amplifier (SA) design. Although the in-memory computing scheme based on two transistors and two magnetic tunnel junction (2T2MTJ) cells can effectively improve the read sensing margin and bit operation accuracy, the area of the memory array increases exponentially. For 1T1MTJ unit, we proposed a high reliability, multi-bit parallel readout in-memory computing scheme using an enhanced multi-bit current-sensitive amplifier (MBCSA) for branch current calculation and three independent reference units. The simulation results indicates that, when compared with the 1T1MTJ and 2T2MTJ schemes with precharge current-sensitive amplifier, the proposed scheme can improve the accuracy rates of the bitwise in-memory computing by 4.07% and 1.65%, respectively, in the typical case, and further enhance the read accuracy and robustness when with a small TMR and a low VDD. Moreover, the proposed scheme can simultaneously perform "AND" and "OR" bitwise logical operations for two pairs of data units in a 6 ns read cycle as well as the results of "NAND" and "NOR" and output four kinds of bit logic operation results.

A design of MHz DC/DC converter based on GaN
NAN Lan, WANG Yingwu, WANG Yongjie, WANG Kai, WANG Junfeng
2022, 39(8): 127-134.   doi: 10.19304/J.ISSN1000-7180.2021.1298
Abstract(0) HTML(0) PDF (0)
Abstract:

At present, high-reliability DC/DC power supply modules with medium and low voltage input generally use Si-based power switches, with a typical input voltage of 28V and a switching frequency of 500kHz. With the development of miniaturization of module power supply and the continuous increase of switching frequency, the switching loss increases greatly, which seriously affects the power conversion. efficient. The application of soft-switching technology can greatly reduce the switching loss caused by high frequency. However, the structure of the soft-switching control circuit is complex, and there is currently no high-end integrated controller available in the military high-reliability field. GaN devices have extremely low gate charge, output capacitance and zero reverse recovery charge characteristics, which can effectively reduce switching losses caused by high-frequency applications without increasing circuit complexity. successful application in the device. However, in the medium and low voltage input military high-reliability power modules, with the substantial reduction of bus voltage and the increase of switching frequency, whether the high-speed characteristics of GaN devices can effectively reduce switching losses and improve conversion efficiency remains to be verified. In this paper, a single-ended flyback power topology and synchronous rectification technology are used to design a principle prototype with a typical input of 28V, output of 5V/30W, and a switching frequency of 1MHz. The loss and efficiency curve of the base device at a switching frequency of 1MHz shows that under the condition of medium and low voltage and high frequency, the conversion efficiency of the GaN device is improved by 4% compared with the Si-based device, and the voltage stress of the power switch is controlled within a reasonable range. It is of great significance for the miniaturization and high frequency development of medium and low voltage input DC/DC converters in the field of military high reliability.

Found in 1972
Monthly

Supervisor:
Xi'an Institute of Microelectronics Technology

Sponsor:
China Aerospace Science and Technology Corporation

ISSN 1000-7180

CN 61-1123/TN