15 October 2021, Volume 47 Issue 10
-
Select all|
-
Computer Engineering. 2021, 47(10): 1-15. https://doi.org/10.19678/j.issn.1000-3428.0061710The human lumen images observed by the doctor through the endoscope is shown in a two-dimensional style, and fail to present the three-dimensional relationships between the lesions, blood vessels and adjacent tissues in the human lumen.The three-dimensional human lumen reconstruction technology and visualization methods can clearly and comprehensively display the three-dimensional images of lesions and tissues, assisting doctors in precise operations.In this paper, the three-dimensional reconstruction technologies for human lumen are divided into the active measurement methods and passive measurement methods.On this basis, this paper summarizes the reconstruction technologies based on structured light, Time Of Flight(TOF), binocular stereo vision, and monocular vision in order, and discusses their development.Then focusing on the reconstruction technology using Simultaneous Localization and Mapping(SLAM), the paper analyzes and compares the development, methods and characteristics of the feature point detection and matching in human lumen.Finally, the difficulties and future development trends of the three-dimensional human lumen reconstruction technology are prospected.
-
Computer Engineering. 2021, 47(10): 16-25. https://doi.org/10.19678/j.issn.1000-3428.0060683Path planning is one of the key technologies for autonomous navigation of mobile robots.It aims at planning a collision free optimal path from the current position to the destination in real time.This paper introduces the path planning techniques that are based on Reinforcement Learning(RL) and common methods, and categorizes the methods based on RL into two types:the value-based methods and the strategy-based methods.Then the paper compares value-based representation methods(including Timing Difference(TD), Q-Learning, etc.) and the strategy-based representation methods(including Strategy Gradient(SG) and Imitation Learning(IL), etc.), and analyzes the development status of its fusion strategy and Deep Reinforcement Learning(DRL).On this basis, the paper summarizes the advantages, disadvantages and application scenarios of the RL-based methods.Finally, the future development trends of the path planning techniques based on RL are discussed.
-
Computer Engineering. 2021, 47(10): 26-33. https://doi.org/10.19678/j.issn.1000-3428.0061371When computing tasks are transferred to the Mobile Edge Computing(MEC) servers, the service cache effectively reduces the real-time delay and bandwidth cost of acquiring and initializing the service application.In addition, Quality of Experience(QoE) is a key factor driving offload decisions, and effective utilization of limited computing resources can keep users satisfied.This paper considers a scenario where a single edge server is used to help mobile users perform a series of computing tasks.On this basis, a Mixed Integer Nonlinear Programming(MINLP) is established, and a Deep Deterministic Policy Gradient(DDPG) algorithm is proposed to jointly optimize the service cache location, the offload decision and the resource allocation, so as to improve the user's QoE of services and maximize the cost saved by users using computing resources.Simulation results show that the proposed method achieves higher QoE and lower cost than the algorithms using non-cache strategy, random-choice strategy and non-cache random-choice strategy.
-
Computer Engineering. 2021, 47(10): 34-42,51. https://doi.org/10.19678/j.issn.1000-3428.0061105Vehicular Ad-hoc Network(VANET) plays an important role in the construction of intelligent transportation systems.The message authentication schemes can ensure the reliability and security of VANET in practical applications, but most of the existing authentication schemes are limited in the computational efficiency.To address the problem, an authentication scheme based on Chebyshev chaotic map for VANET is proposed.With the semi-group nature of Chebyshev polynomial, the proposed scheme securely constructs the symmetric key to finish the key agreement phase between vehicle nodes and the Road-Side Unit(RSU).Then the vehicle nodes use the temporary shared key distributed by RSU to complete the anonymous message authentication phase.There is no need to verify a large revocation list for each signature, and the revocation of vehicles will not affect the performance of the group.The analysis results show that the proposed scheme satisfies the security requirements of VANET for resisting a variety of security attacks, and provides conditional privacy-preserving at the same time.In the key agreement phase and message authentication phase, the scheme exhibits an improved computational effiency and reduced communication overhead.
-
Computer Engineering. 2021, 47(10): 43-51. https://doi.org/10.19678/j.issn.1000-3428.0060802In order to improve the trust evaluation mechanism of the Internet of Vehicles(IoV) and establish the trust relationship between vehicles, a blockchain-based secure communication strategy for IoV is proposed.The elliptic curve encryption algorithm is used to realize identity registration and verification of vehicles.Then by using the direct trust value based on Beta distribution and the recommendation trust value based on the PageRank algorithm, the comprehensive trust value of a vehicles obtained.The smart contract technology is used to control and manage the comprehensive trust values and interaction information of vehicles, implementing the signing up, modification and update of comprehensive values of vehicles.Simulation results show that the proposed strategy can assist in defending against internal and external attacks on IoV, ensure the security of data transmission, and meet the communication requirements of rapidly moving IoV vehicle nodes for high throughput and low delay.
-
Computer Engineering. 2021, 47(10): 52-60. https://doi.org/10.19678/j.issn.1000-3428.0059629The existing subspace clustering methods are only applicable to single-layer networks, or just average the clustering results of each layer in the multi-layer network.They fail to consider the different amounts of information contained in each layer, which causes a reduction in the subspace clustering performance.To address the problem, a sparse subspace clustering method for multi-layer networks is proposed.Firstly, a distance regularization term and a non-negative constraint are jointly integrated into the framework of sparse subspace clustering, which enables the method to simultaneously exploit the global and local information of data for graph learning during clustering.In addition, the sparse constraint is introduced to provide the learned graph with a clear clustering structure, and an iterative algorithm is designed to optimize the solution.Experimental results on multiple real datasets show that the proposed method can mine the complementary information of different layers of the network, obtain an accurate consistent joint sparse representation, and effectively improve the community clustering performance of multi-layer networks.
-
Computer Engineering. 2021, 47(10): 61-66. https://doi.org/10.19678/j.issn.1000-3428.0059375The existing pruning algorithms for Convolutional Neural Network(CNN) models exhibit a low accuracy in evaluating the importance of parameters by relying on their own parameter information, which would easily lead to mispruning and affect the performance of model.To address the problem, an improved pruning method for CNN models is proposed.By training the model with sparse regularization, a deep convolutional neural network model with sparse parameters is obtained.Structural pruning is performed by combining the sparsity of the convolution layer and the BN layer to remove redundant filters.Experimental results on CIFAR-10, CIFAR-100 and SVHN datasets show that the proposed pruning method can effectively compress the network model scale and reduce the computational complexity.Especially on the SVHN dataset, the compressed VGG-16 network model reduces the amount of parameters and FLOPs by 97.3% and 91.2%, respectively, and the accuracy of image classification only loses 0.57 percentage points.
-
Computer Engineering. 2021, 47(10): 67-74. https://doi.org/10.19678/j.issn.1000-3428.0059083Semantic graph summarization is to extract key information from semantic graphs, and generate a summarized model of the original data set to solve problems in understanding, querying, and using large-scale semantic graphs.In order to improve the efficiency of current summarization algorithms, this paper proposes an approach of generating summaries based on ontology partition.The ontology partition algorithm is used to divide the semantic graph into sub-graphs.Then for each sub-graph, its partially ordered lattices (also named characteristic set lattices) of elements are generated using formal concept analysis.On this basis, the characteristic set lattices of all sub-graphs form the summarization of the original semantic graphs.The approach is tested on the Linked Open Data(LOD) dataset and the Berlin SPARQL Benchmark datasets.Results show that the proposed approach exhibits excellent scalability, and a significantimprovement in summarization efficiency.
-
Computer Engineering. 2021, 47(10): 75-81. https://doi.org/10.19678/j.issn.1000-3428.0059354To make full use of noisy speech features to improve the speech enhancement performance of deep neural networks, a speech enhancement method based on the fusion of time-domain and frequency-domain features is proposed.First, by using the waveform of noisy speech as the training feature and the log power spectrum of clean speech as the training target, the mapping from the time-domain features of noisy speech to the frequency-domain features of clean speech is designed.On this basis, the waveform and log power spectrum of noisy speech are used as training features to construct a speech enhancement network that integrates the time-domain and frequency-domain features of noisy speech.Experimental results show that compared with the methods using only frequency-domain features, the proposed method can significantly improve the quality and intelligibility of enhanced speech, and has better speech enhancement performance.
-
Computer Engineering. 2021, 47(10): 82-88. https://doi.org/10.19678/j.issn.1000-3428.0059116TextRank uses a co-occurrence window instead of PageRank Web hyperlinks to determine the relationships between words.However, the vocabulary graph under the co-occurrence window mechanism is an undirected graph, and in most cases, there is no cognitive directional link between the words in the actual Chinese texts and the words in the co-occurrence window.Under this mechanism, the relationship between the words is sharply different from the hyperlink relationship of PageRank.To address the problem, a keyword extraction method, S-TextRank, is proposed integrating semantic features.Based on TextRank, S-TextRank employs dependency relationships instead of co-occurrence windows to determine the relationships between words to simulate directional PageRank hyperlinks.In addition, different part-of-speech words are assigned with corresponding weight coefficients to simulate the importance of different types of Web pages.Finally, a non-keyword list is constructed by using the IDF method and Chinese grammar rules to exclude the influence of irrelevant words on the extraction results.Experimental results show that the accuracy of the S-TextRank method achieves 74% on the test set, 19.4 percentage points higher than that of the TextRank method.
-
Computer Engineering. 2021, 47(10): 89-96,102. https://doi.org/10.19678/j.issn.1000-3428.0058349In the Internet that produces massive data, knowledge is mutually related.A rough domain-specific knowledge graph can display the structured information of the knowledge in this domain, but usually fails to present the potential relationships between entities.To implement smooth extension of relationships between entitites in domain-specific knowledge graphs, a relationship discovery method based on inter-entity association rule analysis and topic analysis is proposed.By utilizing the entity-related data in a specific domain, the potential relationships between domain-specific entities are obtained by analyzing the association rules and the similarity of topic distribution among entity-related data sets.Then the newly discovered relationships are merged into the roughly constructed knowledge graphs to realize the potential relationship extension of the domain-specific knowledge graphs.The experimental results show that the proposed method can discover the commonalities between entities of different sectors, and thus mine the potential relationships between these entities.It improves the efficiency of relationship discovery, and smoothes the extension of domain-specific knowledge graphs.
-
Computer Engineering. 2021, 47(10): 97-102. https://doi.org/10.19678/j.issn.1000-3428.0058964The existing semantic text matching methods are mostly based on a simple attention mechanism for interactions, and less consideration is given to the structural information of the text itself and the original information interactions between the texts.To address the semantic matching between two Chinese texts, a text matching model, MAII, is proposed based on multi-angle information interactions.The model calculates the deep-level semantic interaction matrix of the two texts at the granularity level, local level and global level respectively.At the same time, the interactions between the word order information, the interactions between structural information, and the internal dependencies of the two texts are also considered.On this basis, a semantic vector with rich information is obtained, and then the semantic matching degree between the two texts is calculated through semantic reasoning.Experimental results show that compared to DSSM, ESIM and DIIN models that perform well on the English data sets, the MAII model exhibits better performance on the Chinese data set of CCKS 2018 Question Matching Competition with the accuracy reaching 77.77%.
-
Computer Engineering. 2021, 47(10): 103-110. https://doi.org/10.19678/j.issn.1000-3428.0059355Traffic replay can provide realistic traffic data for the cyber range, and support new technology verification and network security evaluation.To meet the needs of interactive user behavior simulation for complex virtual networks, an architecture for user behavior emulation is designed based on interactive traffic links.The architecture adopts a distributed traffic emulation strategy based on cloud platform to achieve diversified and scalable loading of user behavior emulation for complex target networks.The delay repair and compensation strategy in the process of traffic replay is further studied to improve the timing fidelity of interactive user behavior emulation.Results of emulation experiments show that this method can realize interactive large-scale user behavior emulation with the accuracy of traffic timing ensured.It has certain advantages in the diversity, scale and fidelity of behavior emulation over traditional methods such as ITRM and Tcpreplay, providing effective support for security evaluation.
-
Computer Engineering. 2021, 47(10): 111-115,124. https://doi.org/10.19678/j.issn.1000-3428.0059258Most of the existing secret sharing algorithms are constructed based on the SHAMIR algorithm, and involve complex polynomial calculations, which slows down data processing.In order to improve computational efficiency, a new secret sharing algorithm is proposed based on XOR operation.The number of clues to be generated is calculated according to the input values of k and n.Then all the clues required by the algorithm are generated by continuous XOR operation on a random binary sequence, and the clues are grouped in permutation and combinations to form a shadow secret.Experimental results show that the algorithm can realize arbitrary (k, n)-threshold secret sharing, and greatly improves the processing speed compared with the SHAMIR algorithm, while avoding the security risk of partial secret information disclosure.
-
Computer Engineering. 2021, 47(10): 116-124. https://doi.org/10.19678/j.issn.1000-3428.0061434Among identity verification methods, the single-mode verification methods rely on single features, and are vulnerable to forged authentication and attacks.To solve the problem, an implicit authentication method of multi-modal feature fusion based on user footprints is proposed.The data of user behavior when using the mobile devices, including the touch pressure, the track of finger movement, and the acceleration of user movement, is collected from sensors.Then the feature selection technique is used to extract the features of touch screen interactions, movement mode, and physical location.The extracted features are subsequently trained and fused.On this basis, the multi-modal feature fusion model is used to realize user identity authentication.Experimental results show that the proposed method achieves an authentication accuracy of over 98% in both the feature-level fusion mode and the strategy-level fusion mode.It is less vulnerable to forged authentication and attacks, and displays higher authentication accuracy and stability.
-
Computer Engineering. 2021, 47(10): 125-131. https://doi.org/10.19678/j.issn.1000-3428.0058958The existing code clone detection methods usually have the problem of single mark representation and complex abstract syntax tree structure.To address the problem, a code clone detection method is proposed based on hierarchical features.The method employs two-layer Bi-directional Long Short-Term Memory(Bi-LSTM) networks to extract deeper semantic information at the line level and global code level respectively.On this basis, the semantic features of the target code are mined.Then the attention mechanism is introduced to adjust the influence weight of important tokens and code lines, and thus enhance the performance of code clone detection for complex semantics.Finally, the softmax classifier is used to determine whether the target code is cloned.Experimental results show that the proposed method displays recall rate of 91% and precision of 97%, providing better performance than the NICAD, CCIS and CCLearner methods in code clone detection for complex semantics.
-
Computer Engineering. 2021, 47(10): 132-139,146. https://doi.org/10.19678/j.issn.1000-3428.0060480In the field of privacy and security of smart phones, implicit authentication has been widely studied as it can provide high security and user-friendly interactions.Still, the existing implicit authentication schemes suffer from difficulty in behavior feature collection and high complexity of authentication models.Given the limitations, a hierarchical implicit authentication scheme is proposed based on dynamic trust value.The scheme employs machine learning algorithms to train the model, extracting the features of the scrolling behavior as front-level authentication data.The output probability receives a trust value detection, and the result is taken as the input of back-level authentication data to output the final authentication result.Based on the stability and continuity of real user history authentication, this scheme calculates the average authentication probability value in a certain time window as the dynamically updated trust value, making the trust value ranging within the real user's authentication results.The experimental results demonstrate that the proposed scheme can achieve a classification accuracy of 98.63% and an equal error rate of 3.43%.Compared to the methods with only front-level authentication scheme, the proposed scheme can improve the accuracy of authentication and effectively prevent the impostors from illegally using the phones.
-
Computer Engineering. 2021, 47(10): 140-146. https://doi.org/10.19678/j.issn.1000-3428.0059210The security of encrypted data is affected by encryption algorithms and encryption devices.The reliability of the encryption devices can be tested by using multiple types of attacks, such as energy analysis.Among different attack methods, the method of side channel attacks based on deep learning has been widely concerned since it was proposed. This paper proposes a side channel attack method based on a deep learning network, LSTM.The method employs Correlation Power Analysis(CPA) to determine the interest points of the side channel power consumption data.Then based on the position of the interest points, an appropriate interest interval is selected as the feature vector to build the neural network model.The experimental results show that the LSTM model has higher efficiency in implementing side channel attacks than MLP and CNN.
-
Computer Engineering. 2021, 47(10): 147-152. https://doi.org/10.19678/j.issn.1000-3428.0059152In C-V2X communications, Mode 4 uses the Sensing Based Semi-Persistent Scheduling(SB-SPS) algorithm for resource allocation.This algorithm transmits messages with the maximum power, which will reduce the reliability of the system in the high-density traffic flow state.To optimize the SB-SPS algorithm, a joint resource allocation and power control algorithm based on Deep Reinforcement Learning(DRL) is proposed.After sensing the channel, the vehicle selects the sub-channel with the least interference and adjusts the transmission power adaptively according to the channel state.Then, it solves the optimal sub-channel selection scheme and power control scheme by interactive learning with the environment.The simulation results show that compared with the existing SB-SPS optimization algorithms, the proposed algorithm can improve the packet reception ratio by 5% in high-density highway scenarios, effectively improving the reliability of vehicle-to-vehicle communication.
-
Computer Engineering. 2021, 47(10): 153-159. https://doi.org/10.19678/j.issn.1000-3428.0059439Vehicular Ad Hoc Networks(VANET) are characterized by high-speed mobile vehicle nodes and dynamic network topology, which increases the transmission delay of communication links between vehicles and reduces connection time.For the maintenance of VANET connectivity, this paper proposes an improved AODV routing protocol, AODV-CMIRP, which introduces a selection algorithm based on dual cluster heads to reduce the influence of dynamic changes of the global network topology.The relative mobility and relative speed of nodes are introduced as the index of cluster head selection, and the auxiliary cluster head node is selected to ensure the overall lifetime of VANETs.The simulation results show that the AODV-CMIRP protocol can ensure the network connectivity and stability while exhibiting a lower average end-to-end delay and higher packet delivery fraction than CBDRP and AODV protocols.The proposed protocol can effectively prolong the cluster head lifetime and improve network stability.
-
Computer Engineering. 2021, 47(10): 160-165,173. https://doi.org/10.19678/j.issn.1000-3428.0059207Based on Device-to-Device(D2D) network and relay heterogeneous cellular network, resource reuse can be used to improve system performance, but it also complicates interference in the networks.To address the problem, a Power and Resource Allocation Game(PRAG) algorithm is proposed, which performs interference coordination in D2D network and relay heterogeneous cellular network through power control and resource allocation.Optimal transmitting power of D2D and relay links is derived in maximizing a utility function based a cost parameter.Then the optimal transmission power of D2D and relay links is determined.On this basis, the generated utility matrix is used in the game to choose suitable cellular users for resource reuse.Simulation results show that the proposed algorithm enables higher system throughput with less power compared with the Equal Power Allocation Random(EPAR) algorithm.
-
Computer Engineering. 2021, 47(10): 166-173. https://doi.org/10.19678/j.issn.1000-3428.0058910Ubiquitous electric IoT is characterized by massive access nodes and limited device resources.When providing reliable transmission service, ubiquitous electric IoT tend to suffer from congestion, resulting in loss of power information flow and excessive delay.To solve this problem, a link-stability-based congestion control algorithm, L-CoCC, is proposed based on the Constrained Application Protocol(CoAP) for the application layer in ubiquitous electric IoT.By using the Round-Trip Time(RTT) of strong, weak and failed messages, the algorithm determines the state of the network environment and smoothly estimates the Retransmission Timeout(RTO).Based on the number of retransmission times and the fluctuation value of message RTT, a method for limiting the lower bound RTO is introduced, and the aging concept is updated to avoid unnecessary retransmissions.The experimental results show that compared with the CoCoA++ algorithm and the CoAP algorithm, the proposed algorithm improves the throughput and the success rate, and reduces the average delay.It can effectively alleviate network congestion.
-
Computer Engineering. 2021, 47(10): 174-179,185. https://doi.org/10.19678/j.issn.1000-3428.0059114The Flying Ad-hoc Network(FANET) are highly dynamic, and the nodes are energy-limited, which makes it difficult for traditional routing protocols to adapt to UAV networks.In response to this problem, a routing algorithm, UAV-OLSR, for UAV networks is proposed based on the OLSR protocol.The algorithm realizes the status perception of UAV clusters based on link changes, and comprehensively considers node energy, node location and other factors to realize node quality assessment.In addition, the algorithm adopts the multi-path idea, selecting a data forwarding path that is optimal according to specific path measurement criteria.The simulation results show that UAV-OLSR has lower average packet transmission delay and higher packet delivery rate than OLSR and AODV.The proposed algorithm can extend the UAV networks lifetime.
-
Computer Engineering. 2021, 47(10): 180-185. https://doi.org/10.19678/j.issn.1000-3428.0059668The data transmission systems of remote sensing satellites for earth observation generally adopt Linux Virtual Server Cluster (LVS)architecture to transmit remote sensing data.However, when the LVSs for data transmission in the cluster try to resume transmission at breakpoints, the control information of the old servers is not synchronized with that of new servers, leading to duplicated data slices.To address the problem, a data processing mechanism based on distributed system architecture for remote sensing satellites, called DPM, is designed.The mechanism employs the Kafka message queue to quickly store packets, and uses a new code module to enable Spark Streaming to submit message offset value accurately.Then efficient data transmission rate and progress statistics method are used to record the running state of the DPM mechanism in real time.Experimental results show that the proposed mechanism can accurately record and submit the offset value of messages, and resume data transmission.It can ensure the accurate and stable transmission of remote sensing satellite data.
-
Computer Engineering. 2021, 47(10): 186-193. https://doi.org/10.19678/j.issn.1000-3428.0059234In the object detection task, there is a large size difference between different objects, which makes it difficult to effectively detect object with multiple size.Based on YOLOv3, the Bidirectional FPN Atrous Reception YOLOv3 (BR-YOLOv3) target detection network is proposed.Using atrous convolution can effectively improve the receptive field size of the network layer, using different numbers, convolution kernel size, and dilation rate convolution to build a multi-layer parallel Atrous Receptive Module(ARM), and by using Bidirectional Feature Pyramid Structure Network(BiFPN) realizes bidirectional fusion of shallow and deep features, improving the classified ability of shallow prediction branch, and enhancing the ability of deep prediction branch's target positioning.By using the LOSSGIOU positioning loss function, the target regression process is integrated, and the target miss rate is reduced.Experimental results show that the improved RB-Yolov3 on the Pascal VOC test set has a mean average precision of 79.24%, which is an increase of 4.65% on the basis of the original network.It is superior to mainstream target detection networks such as SSD and Faster RCNN in detection accuracy.
-
Computer Engineering. 2021, 47(10): 194-200. https://doi.org/10.19678/j.issn.1000-3428.0058761The existing image description models face the challenges of low training efficiency, low level of the decoder, and the poor grammar coherence and content diversity of the generated descriptive sentences.To address the problem, a deep image description model, Deep-NIC, based on Independent Recurrent Neural Network(IndRNN) is proposed.The deep decoder unit is built using both independent recurrent neuron and the Batch Normalization(BN) method.Then based on the stacked multiple layers of decoder units, the deep decoder is established.Finally, the Google inception V3 has been used as the encoder to build a deep image description model.Experimental results on the data set MS COCO2014 show that compared to the baseline model NIC, the Deep-NIC model delivers a performance improvement of 3.2% under the BLEU-4 scoring standards, 10.3% under METEOR, and 8.18% under CIDER.The proposed model is easier to train, and can provide better fitting performance.
-
Computer Engineering. 2021, 47(10): 201-206. https://doi.org/10.19678/j.issn.1000-3428.0059224The images collected in low-illumination environment are limited in the contrast, and suffer from detail loss and noise interference.To address the problem, a Retinex-based method is proposed to improve illumination map estimation and realize low-illumination image enhancement.The maximum values in the three color channels of R, G and B are calculated, and the illumination is approximated with L2 norm.Then an improved model based on the relative total variation is used to smooth and refine the bright channel, and implement adaptive Gamma correction.Finally, the image is enhanced by using the Retinex model.The MATLAB simulation platform is used for experiments of low-illumination image enhancement.The results show that compared with Retinex-Net, SRIE and other typical algorithms, the proposed algorithm can effectively improve image contrast and clarity, enhance image details, and make image colors more vivid and natural to improve visual effects.
-
Computer Engineering. 2021, 47(10): 207-213. https://doi.org/10.19678/j.issn.1000-3428.0059225Most of the existing face alignment methods are not end-to-end, and require frequent manual intervention, which leads to a reduction in their stability.To address the problem, an end-to-end face alignment method based on deep learning is proposed.The network required by this method is constructed based on the sub-modules of the MobileNet series in a structure similar to VGG.Taking the entire image as the input, the depth-wise separable convolution module is used for feature extraction, and the method employs an improved inverted residual structure to avoid the disappearance of gradients in the network training process while reducing the loss of features.The distance between eyes is taken as the basis for normalization.The designed network is tested on the 300W face dataset and compared with CDM, DRMF methods. The experimental results show that the proposed algorithm displays excellent accuracy and real-time performance.
-
Computer Engineering. 2021, 47(10): 214-220. https://doi.org/10.19678/j.issn.1000-3428.0059464The existing foggy image processing methods can achieve good dehazing effect, but some details are often lost, and noise amplification is easy to occur in the noisy areas.In order to solve these problems, a new variational dehazing model, H-TVBH, is proposed based on dark channel priori and the variational model that uses Total Variation(TV) and Bounded Hessian(BH) rule terms.The initial transmittance of the image is estimated according to the dark channel prior principle.At the same time, the atmospheric light value is estimated by quadtree decomposition.Then the obtained initial transmittance and atmospheric light value are applied to the proposed model.After that, the auxiliary variables and Bregman iteration parameters are introduced, and the split Bregman algorithm as well as fast Fourier transform is adopted to solve the optimized transmittance and dehazing image through alternate iterations.Experimental results show that the proposed algorithm can enhance the image contrast while effectively suppressing the noise in the image, retain the image texture details, and make the image clearer and more natural.
-
Computer Engineering. 2021, 47(10): 221-225,235. https://doi.org/10.19678/j.issn.1000-3428.0059009In order to reduce the coding complexity of the new generation of high-efficiency video coding standard H.266/VVC, a fast prediction mode decision method is proposed for intra coding.The distribution characteristic of transform residuals is first analyzed according to the rate-distortion theory.The coded bit-rate prediction model that considers the influence of scalar quantization is derived.Then combing the principle of intra prediction, the coded distortion prediction model is constructed by using the current residuals, the reference residuals and the reference distortion.Finally the optimal intra-prediction mode is determined based on the rate-distortion cost.Experimental results show that the proposed method could save 31% coding time on average while keeping a good coding efficiency.
-
Computer Engineering. 2021, 47(10): 226-235. https://doi.org/10.19678/j.issn.1000-3428.0059128Deep learning has improved the accuracy and efficiency of scene classification for optical remote sensing images, but some scenes are easily misclassified due to the rich semantic information in the images.At the same time, the hardware requirements and time overhead are increasing with the size of scaling network models, which restricts the application of deep learning models.To address the problem, a method for classifying the scenes in optical remote sensing images is proposed based on a lightweight network model.The method employs EfficientNet to extract image features, and then new features with richer information are generated from the extracted features.Multiple sub-classifiers are used to construct an ensemble learning module to analyze the new features, and get the pre-classification results.Finally, the pre-classification results are weighted to obtain the final classification results.The experimental results show that even if only 20% of data samples are used for training, the proposed method still exhibits an accuracy of 94.32% on the AID data set and 93.36% on the NWPU-RESISC45 data set.Compared with D-CNNs, CNN-CapsNet and other methods, the proposed method provides better classification performance with the number of parameters and amount of floating operations greatly reduced.
-
Computer Engineering. 2021, 47(10): 236-241. https://doi.org/10.19678/j.issn.1000-3428.0059168The existing target detection algorithms based on convolutional neural networks have achieved a high accuracy, but the accuracy gain comes at the cost of detection speed, making it difficult for the algorithms to implement real-time detection with limited computing power.To solve this problem, a series of lightweight methods are adopted based on the YOLO target detection algorithm.The methods employ Mobilenetv1 to replace the basic network of Darknet53, and depthwise separable convolutions to replace the 3×3 standard convolutions in the YOLO head part.On this basis, the convolution layer filter is sorted and pruned according to sensitivity.Finally, C++ inference algorithms are deployed on the embedded GPU TX2 platform.The test results on the VOC data set show that the improved algorithm provides an acceleration of 2.4 times while the accuracy is reduced by only 0.75 percentage points.Additionally, the memory occupied by the improved model is only 21.5% of that occupied by the original model.
-
Computer Engineering. 2021, 47(10): 242-251. https://doi.org/10.19678/j.issn.1000-3428.0059166To reduce the difficulty in extracting features of an occluded face, a dual-channel Convolutional Neural Network (CNN) model with occlusion perception is proposed.The model is constructed by integrating newly designed occlusiondecision units into VGG16 network, which aims at extractingexpression-related features of the areas that are less occluded.The model employs the transfer learning algorithm to pre-train the parameters of the convolutional layer, which means to alleviate the over-fittingproblem.At the meantime, the expression-related features of the whole facial image are extracted by the modified residual network.Finally, the outputs of theperceptive neural network and residual network arefused in a weighted manner.The experimental results show that the proposed model achieves an accuracy of 97.33% on CK+, 86% on RAF-DB, and 61.06%on SFEW.Compared with traditional OPCNN, ResNet, and VGG16 models, the proposed model exhibits a significant improvement in the accuracy of recognizing the expression of an occluded face.
-
Computer Engineering. 2021, 47(10): 252-259,268. https://doi.org/10.19678/j.issn.1000-3428.0059122Wireless Capsule Endoscope(WCE) technology can detect gastrointestinal abnormalities.However, the performance of computer-aided diagnosis based on WCE images is reduced due to the small amount of labeled image data, intra-class variation and inter-class similarity.To address the problem, an attentional relational network-based WCE image classification method is proposed.The method combines the relational network, the attention mechanism and the meta-learning training strategy.On this basis, an embedded module based on the attention mechanism is built to extract features of WCE images, and then the extracted features are input into the relation module after feature mapping cascade.The category of the samples is judged according to the similarity score output by the relation module, and the network is trained by using the meta-learning training strategy.The experimental results show that the classification accuracy of the proposed method is higher than that of RelationNet, MAML and other small sample classification methods, reaching up to 90.28% on the WCE dataset.
-
Computer Engineering. 2021, 47(10): 260-268. https://doi.org/10.19678/j.issn.1000-3428.0060938Human activity recognition is a deep learning-based technology, which uses deep learning network models to automatically extract deep features of data.The traditional machine learning algorithms rely heavily on manual intervention during feature extraction, and exhibit a poor generalization ability.To address the problem, a deep learning model, CLT-net, is proposed based on space-time feature fusion for human activity recognition.CLT-net employs Convolution Neural Network (CNN) to extract the deep hidden features of human activity data automatically.Also, Long Short-Term Memory (LSTM) network is used to construct the time series model to learn the long-term dependence of human activity features on the time series.Finally, the softmax classifier is used to classify different human activities.The experimental results based on the public dataset, DaLiAc, show that CLT-net achieves an accuracy of 97.6% in the recognition of 13 kinds of human activities, outperforming the traditional models based on CNN, LSTM and BP.CLT-net has better classification performance of human activity recognition.
-
Computer Engineering. 2021, 47(10): 269-275,282. https://doi.org/10.19678/j.issn.1000-3428.0059096The YOLO algorithm for object detection is limited by the inaccurate positioning of the boundary box and the low detection accuracy for small objects.To address the problem, an improved YOLO algorithm, dcn-YOLO, is proposed based on deformable convolution for object detection.The algorithm employs the K-means++ to cluster anchor boxes that are more in line with the size of data set, so as to reduce the impact of initial points on clustering results and speed up the convergence of network training.Then, a residual deformable convolution module, res-dcn, is constructed.Two improved dcn-YOLO algorithms are derived by embedding res-dcn in the first YOLO feature extraction head module or replacing three YOLO feature extraction head modules with res-dcn, so the network can adaptively learn the receptive field of feature points and extract more effective features for objects of different sizes and shapes, increasing the detection accuracy.Experimental results on VOC data sets show that the propose algorithm can effectively improve the object detection accuracy.Its mAP reaches 82.6%, which is 2.1 percentage points higher than that of YOLO, 5.2 percentage points higher than that of SSD and 9.4 percentage points higher than that of Faster R-CNN.
-
Computer Engineering. 2021, 47(10): 276-282. https://doi.org/10.19678/j.issn.1000-3428.0059043Generally, the speed gain of traditional target detection models comes at the cost of accuracy, and vice versa.To address the problem, a new pedestrian detection model, PD-CenterNet, is proposed based on CenterNet by improving its network structure and loss function.In terms of network structure, a feature fusion module based on attention mechanism is given in the up-sampling path to fuse low-level features and high-level features.In terms of the loss function, three factors α、γ and δ are designed to increase the loss of positive samples and reduce the loss of negative samples, balancing the loss of the samples.Experimental results show that compared with the CenterNet model, the proposed model improves the accuracy of network structure by 5.1% and the accuracy of the loss function by 9.81%.
-
Computer Engineering. 2021, 47(10): 283-289,297. https://doi.org/10.19678/j.issn.1000-3428.0059305In order to achieve fast and accurate detection of surface ship targets, this paper proposes a ship target detection algorithm based on improved YOLOv3-Tiny.Firstly, in network structure, the features of shallow layers of the network is enhanced and reconstructed according to the characteristics of ship targets to reduce the miss detection rate of small targets, and the improved residual network is introduced to improve the depth of the network while reducing the calculation of network parameters.Moreover, the pyramid network is used for multi-scale feature fusion to balance the detection capability between large ship targets and small ship targets in images.Secondly, in the network training, transfer learning strategy is employed to train the designed network model to alleviate the limitation of known ship image samples.Finally, in video detection, a video frame selection method for forward computation of the network model based on structure similarity of inter frames is proposed to improve the detection frame rate.The experimental results show that the proposed algorithm has precision rate up to 92.4%, with an increase of 7% compared with YOLOV3-Tiny, recall rate up to 84%, and detection frame rate up to 12 frames/s on CPU platform.
-
Computer Engineering. 2021, 47(10): 290-297. https://doi.org/10.19678/j.issn.1000-3428.0059404Traditional vision-based landing schemes cannot cope with the complicated environmental changes during landing of Unmanned Aerial Vehicles(UAV), and thus fail to process images in real time using UAV-borne processors.To address the problem, a lightweight and efficient Onboard-YOLO algorithm is proposed for UAV-borne processors.The algorithm employs separable convolution instead of conventional convolution kernels to improve the calculation speed.Then the attention mechanism is used for the automatic learning of channel feature weights to improve the accuracy of the model.The landing algorithm is tested in various cases of interference, including motion blur, occlusion, target going beyond the visual field, illumination and scale changes.The test results show that compared with the advanced real-time detection algorithms, the proposed Onboard-YOLO algorithm can deal with the complicated environmental changes better during landing.Its calculation speed reaches 18.3 frames per second on the airborne processor, which is 4.3 times faster than that of the original YOLO algorithm, and 25.7 times faster than that of Faster-RCNN.Additionally, the accuracy of the algorithm reaches 0.91, which is 8.9 percentage points higher than that of Mobilenet-SSD.Onboard-YOLO enables autonomous real-time precise landing based on the airborne processor, bringing the success rate of landing to 95%.
-
Computer Engineering. 2021, 47(10): 298-305,313. https://doi.org/10.19678/j.issn.1000-3428.0058733The existing methods for checking the wear of safety helmets suffer from complex background and strong interference, and display poor performance on small targets.To address the problem, an improved SSD algorithm is proposed for detecting the wear of safety helmets.The algorithm employs the lightweight MobileNet to construct the MobileNet-SSD algorithm, which improves the detection speed.Then the transfer learning strategy is introduced to address the difficulties in model training.Additionally, as the existing data sets of safety helmets are small-sized, which leads to the underfitting of the network, samples of safety helmets are collected from the actual building work videos to construct a suitable sample set.The experimental results show that the proposed algorithm provides a detection speed that is 10.2 times higher than that of the SSD algorithm with the cost of a minor loss in accuracy.
-
Computer Engineering. 2021, 47(10): 306-313. https://doi.org/10.19678/j.issn.1000-3428.0059312The existing road segmentation methods generally suffer from gradient disappearance, and are limited in feature utilization and semantic segmentation accuracy.To address the problem, a U-Net-based semantic segmentation model is proposed for satellite road images.The dense connection module is used to reduce gradient disappearance.Then the Atrous Spatial Pyramid Pooling(ASPP) structure is introduced to retain more image features.Finally, the attention monitoring mechanism is used when learning deep-level feature information, so as to more effectively extract the feature information of road elements.The test results on a road data set from satellite images show that compared with FCN, SegNet and U_Net algorithms, the proposed algorithm improves the accuracy to 96.3%, recall rate to 96.9% and precision to 96.6%.This algorithm can segment the road elements accurately.
-
Computer Engineering. 2021, 47(10): 314-320. https://doi.org/10.19678/j.issn.1000-3428.0058993Due to the limited reception domain, the existing multi-scale Convolutional Neural Networks(CNN) often fail to model space targets with super-scale variation.In order to solve this problem, a Hyper-Scale Self-Guided Attention Networks(HSSGAN) recognition framework for remote sensing ships is proposed.The framework employs a lightweight super-scale subspace module connected by groups to capture the super-scale feature and scale invariance of the ship.Then the super-scale feature map is refined gradually by using the self-guided attention, and a long-term dependency relationship is established between the super-scale local and global semantics adaptively to enhance the difference of the feature maps between classes.In addition, irrelevant information is ignored while relevant features are aggregated, so the identifiability of the target ship can be enhanced.The experimental results show that the HSSGAN method exhibits improved recognition performance with the F1 value reaching 0.966 78.
2026
Volume 52 No 3
Volume 52 No 3
(Started in 1975, Monthly)
Superintended by:
China Electronics Technology Group Corporation
Sponsored by：
East-China Institute of Computer Technology
Shanghai Computer Society
Edited & Published by：
Editorial Office of Computer Engineering
Chief Editor：Jiang Bo
Executive Chief Editor:
Hu Bing
Tel：021-67092217
E-mail: ecice06@ecict.com.cn
CN: 31-1289/TP
ISSN: 1000-3428
Superintended by:
China Electronics Technology Group Corporation
Sponsored by：
East-China Institute of Computer Technology
Shanghai Computer Society
Edited & Published by：
Editorial Office of Computer Engineering
Chief Editor：Jiang Bo
Executive Chief Editor:
Hu Bing
Tel：021-67092217
E-mail: ecice06@ecict.com.cn
CN: 31-1289/TP
ISSN: 1000-3428