ISSN 0253-2778

CN 34-1054/N

2017 Vol. 47, No. 10

Display Method:
Original Paper
Ensemble max-pooling: Is only the maximum activation useful when pooling
ZHANG Hao, WU Jianxin
2017, 47(10): 799-807. doi: 10.3969/j.issn.0253-2778.2017.10.001
The pooling layer in convolutional neural networks performs subsampling on the basis of the local correlation principle, reducing the data size while keeping useful information in order to improve generalization, and effectively increase receptive fields simultaneously. The winner-take-all strategy is used in classical max-pooling, which will affect the generalization of the network sometimes. A simple and effective pooling method named ensemble max-pooling was introduced, which can replace the pooling layer in conventional convolutional neural networks. In each pooling region, ensemble max-pooling drops the neuron with maximum activation with probability p, and outputs the neuron with second largest activation. Ensemble max-pooling can be viewed as an ensemble of many basic underlying networks, and it can also be viewed as the classical max-pooling with some local distortion of the input. The results achieved are better than classical pooling methods and other related pooling approaches. DFN-MR is derived from ResNet, compared with which it has more basic underlying networks and avoids very deep networks. By keeping other hyperparameters unchanged, and replacing each convolutional layer in DFN-MR with a tandem form, i.e., a combination of an ensemble max-pooling layer and a convolutional layer with stride 1, it is shown to deliver significant gains in performance.
Parallel implementation of surf algorithm based on OpenCL
GUO Jing, CHEN Xianfu
2017, 47(10): 808-816. doi: 10.3969/j.issn.0253-2778.2017.10.002
SURF algorithm has high computational complexity and can not meet the real-time requirement. To solve these problems, a parallel SURF algorithm based on OpenCL was presented. Firstly, data parallelism and task parallelism model were applied to the calculations of the integral images, Hessian detector, orientation and descriptor, and the detailed algorithm process was given. Secondly, the performance of the parallel algorithm was optimized from data transmission, memory access and load balancing. The experimental results show that the algorithm can achieve more than 10 times speedup for images with different resolution, and some high-resolution images can even reach up to 39.5 times. Furthermore, it can be applied to a variety of general purpose computing platforms.
Life span prediction of Huizhou architecture based on improved Elman neural network
ZHANG Guangbin, ZHANG Runmei
2017, 47(10): 817-822. doi: 10.3969/j.issn.0253-2778.2017.10.003
Huizhou architecture comprises one of the four ancient architectural schools in China, with wood components being its core. The accurate prediction of Huizhou architectures wood life is of great significance for the protection of ancient buildings. At present, there are few studies have been conducted on the influence of various factors on the service life of the wood components. Elman neural network is typical multi-layer dynamic recurrent neural network, which has the function of mapping dynamic characteristics by storing internal state. This gives the network the ability to adapt to time-varying characteristics, which can be used to predict the complex nonlinear time-varying system. The basic Elman neural network has the characteristics of slow training speed and the tendency to fall into local minimums. Therefore the particle swarm optimization algorithm with adaptive mutation operator is used to improve the basic Elman neural network. The algorithm optimizes the weights of each layer in the network, improves the learning speed, and finds the optimal solution in the global range. The improved network can fit the training value more accurately and can effectively predict the test value. The simulation results show that the network structure can be well applied to the life span prediction of Huizhou architecture.
Distributed keyword approximate search method for RDF
CHEN Yuan, WANG Jingbin
2017, 47(10): 823-836. doi: 10.3969/j.issn.0253-2778.2017.10.004
Existing RDF keyword search methods mainly search on the large-scale RDF data graph directly and do not make full use of the semantic information in the RDF ontology. Too many iterations lead to unfavorable search efficiency and unsatisfactory results. To solve these problems, a distributed keyword approximate search algorithm (DKASR) for RDF based on Redis memory database cluster was proposed and the parallel search of large-scale data on the distributed platform was realized. The algorithm constructs ontology sub-graphs by using the semantic information of RDF ontology, uses the semantic scoring function to sort ontology sub-graphs, and searches and returns the Top-k results concurrently with the aid of MapReduce computation model. If the results do not meet Top-k, ontology sub-graphs are extended to generate approximate ontology sub-graphs and the semantic similarity function is used to sort approximate ontology sub-graphs. Then, MapReduce computation model was used to realize the parallel search until the results meet Top-k. Finally, the results of experiments show that the DKASR algorithm can realize the RDF keyword approximate search and return the Top-k results efficiently and accurately.
Face tracking algorithm of color and edge features based adaptive fusion
ZHOU Pingping, WAN Honglin, LIU Hui, LI Tianping
2017, 47(10): 837-842. doi: 10.3969/j.issn.0253-2778.2017.10.005
In view of the imprecision of traditional tracking algorithms based on color histogram, a face tracking algorithm combining face multiple features based on adaptive fusion in the basic frame of particle filtering was presented.First, the color and edge features of the human face were extracted in the video sequence, while the weighted color histogram and edge orientation histogram(EOH) described facial features.Then, a self-adaptive features fusion strategy was employed to calculate particle set weight.The reliability of face tracking was enhanced by the self-adaptive features fusion strategy.Experimental results show that in the cases of complex backgrounds such as similar skin color, illumination change and so on, the proposed approach improves the tracking effect and has strong robustness.
A fast solution to monostatic RCS based on SVD-CBFM and RACA
ZHU Jinyu, SUN Yufa, FANG Hongyu, DING Yahui, YANG Ming
2017, 47(10): 843-847. doi: 10.3969/j.issn.0253-2778.2017.10.006
An efficient method was proposed to solve monostatic RCS based on singular value decomposition-characteristic basis function method (SVD-CBFM). To reduce the numbers of incident wave excitations, the method considers the coupling effect among the sub-blocks, and calculates the secondary characteristic basis function (SCBF) of each sub-block. The recompressed adaptive cross approximation (RACA) algorithm was applied to recompress the characteristic basis functions (CBFs), which can accelerate the generation of CBFs. In order to further improve the speed of the matrix vector multiplication in the construction process of the SCBF and reduced matrix, the RACA algorithm was also applied to fill the impedance matrix of the far field. The numerical examples demonstrate the accuracy and efficiency of the proposed method.
Segmented reconstruction for SAR imaging based on 1-bit compressed sensing
LYU Yuanhao, LIU Falin, LI Bo, HU Jingqiu
2017, 47(10): 848-853. doi: 10.3969/j.issn.0253-2778.2017.10.007
In Synthetic Aperture Radar (SAR) systems, imaging algorithms based on 1-bit have a great advantage because the echo’s SNR is generally not high. However, those algorithms often require more time than the match filter strategy. A segmented method based on 1-bit compressed sensing was proposed for reducing reconstruction time. Simulation experiments verify the effectiveness of the proposed algorithm. This method can not only reduce time consumption, but also the required memory size. The stepped-frequency waveform is selected not only because it can make sampling easier than other waveforms, but also because it takes less time to recover the same scene when the stepped-frequency waveform is used instead of linear frequency modulation (LFM) waveform.
Design of multiple pointer elastic buffer for high-speed interface
CHENG GuoLin, CHANG Hong, KE Daoming, ZHANG Ping
2017, 47(10): 854-861. doi: 10.3969/j.issn.0253-2778.2017.10.008
Elastic buffers are widely used in the high speed interface of the physical layer, which usually completes the addition and deletion of skip(SKP) by reading/writing pointer jumping and breakpoint preservation. However, common single pointer elastic buffer must be operated at high frequencies, which would make it easy to create complex timing problems. To solve these problems, based on the FPGA and the USB3.0 protocol, a four read/write pointer addressing elastic buffer to complete the addition and deletion of SKP has been proposed. First, the elastic buffer makes use of the input control unit to change the sequence of the SKP pairs in the input data and the output control unit to change the output data. Then, the threshold detection unit sends the valid instructions to the read/write pointer control unit by checking whether the amount of valid data in the elastic buffer achieves the threshold which is added or deleted. Last, to maintain the elastic buffer in half full state, the SKP in data is added or deleted by controlling the addressing of the four read/write pointers. Experimental results show that the designed elastic buffer can achieve the function of SKP addition and deletion, and its clock frequency can satisfy the protocol of Universal Serial Bus 3.0.
Skew cyclic codes over F q[u,v]/〈u2-1, v3-v, uv-vu〉
GUAN Yue, LIU Yan, SHI Minjia, LU Zhenyu, WU Bo
2017, 47(10): 862-868. doi: 10.3969/j.issn.0253-2778.2017.10.009
The skew cyclic codes over R=
A variable granularity-based mapping scheme
FAN Jin, TAN Shoubiao, CHEN Junning
2017, 47(10): 869-877. doi: 10.3969/j.issn.0253-2778.2017.10.010
Address mapping is the most important function of flash translation layer (FTL), and it should have both high performance and low memory footprint. Although demand-based address mapping scheme(DFTL) has relatively high performance and low memory footprint, it has problems. First, it is a page-level mapping scheme. Each cache slot stores only one mapping record, and each mapping record stores only one physical page number and the corresponding logical page number, so that the cache , at a fixed size, can only store a limited number of mapping records. Second, each mapping record itself cannot exploit the spatial locality of the request. Thus, in the DFTL scheme, since the cache hit ratio is low, DFTL has to frequently access the mapping pages in the flash memory to read the mapping records, which significantly reduces the performance of the system. A new scheme named VGFTL (a mapping scheme of variable granularity-based flash translation layer) was proposed, which could significantly increase the cache hit ratio. Experimental results show that the average hit ratio of cache has reached 89.85% in VGFTL scheme, which is much higher than that of the DFTL scheme, 45.46%. VGFTL can significantly reduce the number of block erasures and system response time compared to DFTL, and is close to pure page-level mapping scheme in performance.
Spice modeling of 4H-SiC MOSFET based on advanced mobility model
ZHOU Yuming, LI Yongjie, BAO Guanjia, LIU Hangzhi
2017, 47(10): 878-884. doi: 10.3969/j.issn.0253-2778.2017.10.011
Spice modeling of 4H-SiC MOSFET based on advanced mobility model has been developed. This modeling employs the Spice Level-1 model of MOSFET, but the constant mobility in the current equations has been replaced by the advanced mobility expressions, which can exactly reflect the effect of 4H-SiC/SiO2 interface features on the characteristics of 4H-SiC MOSFET. The transfer characteristics of the developed 4H-SiC MOSFET model have been verified by the production datasheet, and the dynamic characteristics have been experimentally verified in DC/DC Boost converter. Based on the developed 4H-SiC MOSFET model, the effect of 4H-SiC/SiO2 interface trap and surface roughness on the dynamic characteristics of 4H-SiC MOSFET has been discussed. The results show that the switching loss increases with the increase in interface trap density, but surface roughness exists little impact on switching characteristics. The achieved results are very helpful to device application and device process.