In order to address the low accuracy and speed of detection by manual and traditional automation methods for the weld seam surface of traction seat， a lightweight weld seam quality detection algorithm YOLOv5s-G2CW was proposed for the weld seam surface of traction seat. Firstly， the GhostBottleneckV2 module was applied as a replacement for the C3 module in YOLOv5s to reduce the number of parameters used in the model. Then， the CBAM （Convolutional Block Attention Module） was introduced into the Neck of the YOLOv5s model for integration of the weld features in two dimensions： channel and space. Also， the positioning loss function of the YOLOv5s model was improved into Wise-IoU， focusing on the predictive regression of ordinary quality anchor frames. Finally， the 13 × 13 feature layer used for the detection of large-sized objects in the YOLOv5s model was removed to further reduce the number of parameters used in the model. Experimental results show that， compared with the YOLOv5s model， the size of YOLOv5s-G2CW model reduces by 53.9%， the number of frames transmitted per second increases by 8.0%， and the mAP （mean Average Precision） value increases by 0.8 percentage points. It can be seen that the model is applicable to meet the requirements for real-time and accurate detection of the weld seam surface for traction seat.
With the rapid development of Artificial Intelligence （AI） in recent years， the demand for Deep Neural Network （DNN） from devices with limited resources such as embedded devices and mobile devices has increased sharply. The problem of how to compress neural networks without affecting the effect of DNNs has great theoretical and practical significance， and is a hot research topic in deep learning now. Firstly， aiming at the problem that DNN is difficult to be ported to resource-limited devices such as mobile devices due to their large models and large computational cost， the experimental performance of existing DNN compression algorithms in terms of memory usage， running speed， and compression effect was deeply analyzed， so that the influence factors of the DNN compression algorithm were explored. Then， the knowledge transfer structure composed of student network and teacher network was designed， the knowledge distillation， structural design， network pruning， and parameter quantization mechanisms were fused together， and a DNN optimization and compression model based on hybrid mechanism was proposed. Experimental comparison and analysis were conducted on mini-ImageNet dataset using AlexNet as the Benchmark. Experimental results show that the capacity of compressed AlexNet is reduced by 98.5% with 6.3% loss of accuracy， which verify the effectiveness of the proposed algorithm.
To improve the learning process of multi-agent system and the robustness of the system to external disturbances， an iterative learning consensus control algorithm with feedback control was proposed. Firstly， the learning process of agents was improved by sharing the control input information among agents， and the robustness of the system was improved by designing a feedback controller when there were non-iterative repetitive disturbances outside the system. Then， by using the contraction mapping method， the system consensus was analyzed， and the convergence condition of the algorithm was derived strictly. Finally， the correctness and effectiveness of the algorithm was verified through simulations. Compared with the P-type algorithm， the improved algorithm has higher convergence speed and smoother convergence curve in the presence of external disturbances.
In order to solve the problem of high computational complexity of traditional multi-user mmWave relay system beamforming methods， a Singular Value Decomposition （SVD） method based on Deep Learning （DL） was proposed to design hybrid beamforming for the optimization of the transmitter， relay and receiver. Firstly， DL method was used to design the beamforming matrix of transmitter and relay to maximize the achievable spectral efficiency. Then， the beamforming matrix of relay and receiver was designed to maximize the equivalent channel gain. Finally， a Minimum Mean Square Error （MMSE） filter was designed at the receiver to eliminate the inter-user interference. Theoretical analysis and simulation results show that compared with Alternating Maximization （AltMax） and the traditional SVD method， the hybrid beamforming method based on DL reduces the computational complexity by 12.5% and 23.44% respectively in the case of high dimensional channel matrix and many users， and has the spectral efficiency improved by 2.277% and 21.335% respectively with known Channel State Information （CSI）， and the spectral efficiency improved by 11.452% and 43.375% respectively with imperfect CSI.
Unmanned Aerial Vehicle （UAV） is flexible and easy to deploy， and can assist Mobile Edge Computing （MEC） to help wireless systems improve coverage and communication quality. However， there are challenges such as computational latency requirements and resource management in the research of UAV-assisted MEC systems. Aiming at the delay problem of UAV providing auxiliary calculation services to multiple ground terminals， a Twin Delayed Deep Deterministic policy gradient （TD3） based Task Offloading Algorithm for Delay Minimization （TD3-TOADM） was proposed. Firstly， the optimization problem was modeled as the problem of minimizing the maximum computational delay under energy constraints. Secondly， TD3-TOADM was used to jointly optimize terminal equipment scheduling， UAV trajectory and task offloading ratio to minimize the maximum computational delay. Simulation analysis results show that compared with the task offloading algorithms based on Actor-Critic （AC）， Deep Q-Network （DQN） and Deep Deterministic Policy Gradient （DDPG）， TD3-TOADM reduces the computational delay by more than 8.2%. It can be seen that TD3-TOADM algorithm has good convergence and robustness， and can obtain the optimal offloading strategy with low delay.
In order to mine and sense the geometric features of point clouds and further improve the semantic segmentation effect of point clouds by feature enhancement， a point clouds semantic segmentation network based on feature enhancement was proposed. Firstly， the Geometric Feature Sensing Of Point cloud （GFSOP） module was designed to make the network capable of sensing the local geometric structure of point clouds， semantic representations were enhanced by capturing spatial features between points， and multi-scale features were obtained by the idea of hierarchical extraction of features. At the same time， spatial attention and channel attention were fuseed to predict semantic labels of point clouds， and the segmentation performance was improved by strengthening spatial correlation and channel dependence. Experimental results on the indoor dataset S3DIS （Stanford large-scale 3D Indoor Spaces） show that compared with PointNet++， the proposed network improves the mean Intersection over Union （mIoU） by 5.7 percentage points and the Overall Accuracy （OA） by 3.1 percentage points， and has stronger generalization performance and more robust segmentation effect on point clouds with problems of noise， uneven point cloud density and unclear boundaries.
Aiming at the problems of limited representation of spectrogram feature correlation information and unsatisfactory denoising effect in the existing speech enhancement methods， a speech enhancement method of Double Complex Convolution and Attention Aggregating Recurrent Network （DCCARN） was proposed. Firstly， a double complex convolutional network was established to encode the two-branch information of the spectrogram features after the short-time Fourier transform. Secondly， the codes in the two branches were used in the inter- and and intra-feature-block attention mechanisms respectively， and different speech feature information was re-labeled. Secondly， the long-term sequence information was processed by Long Short-Term Memory （LSTM） network， and the spectrogram features were restored and aggregated by two decoders. Finally， the target speech waveform was generated by short-time inverse Fourier transform to achieve the purpose of suppressing noise. Experiments were carried out on the public dataset VBD （Voice Bank+DMAND） and the noise added dataset TIMIT. The results show that compared with the phase-aware Deep Complex Convolution Recurrent Network （DCCRN）， DCCARN has the Perceptual Evaluation of Speech Quality （PESQ） increased by 0.150 and 0.077 to 0.087 respectively. It is verified that the proposed method can capture the correlation information of spectrogram features more accurately， suppress noise more effectively， and improve speech intelligibility.
Aiming at the problems of low accuracy and poor real-time performance of Noctiluca scintillans red tide extraction in the field of satellite remote sensing， a Noctiluca scintillans red tide extraction method from Unmanned Aerial Vehicle （UAV） images based on deep learning was proposed. Firstly， the high-resolution RGB （Red-Green-Blue） videos collected by UAV were used as the monitoring data， the backbone network was modified to VGG-16 （Visual Geometry Group-16） and the spatial dropout strategy was introduced on the basis of the original UNet++ network to enhance the feature extraction ability and prevent the overfitting respectively. Then， the VGG-16 network pre-trained by using ImageNet dataset was applied to perform transfer learning to increase the network convergence speed. Finally， in order to evaluate the performance of the proposed method， experiments were conducted on the self-built red tide dataset Redtide-DB. The Overall Accuracy （OA）， F1 score， and Kappa of the Noctiluca scintillans red tide extraction of the proposed method are up to 94.63%， 0.955 2， 0.949 6 respectively， which are better than those of three traditional machine learning methods — K-Nearest Neighbors （KNN）， Support Vector Machine （SVM） and Random Forest （RF） as well as three typical semantic segmentation networks （PSPNet （Pyramid Scene Parsing Network）， SegNet and U-Net）. Meanwhile， the red tide images of different shooting equipment and shooting environments were used to test the generalization ability of the proposed method， and the corresponding OA， F1 score and Kappa are 97.41%， 0.965 9 and 0.938 2， respectively， proving that the proposed method has a certain generalization ability. Experimental results show that the proposed method can realize the automatic accurate Noctiluca scintillans red tide extraction in complex environments， and provides a reference for Noctiluca scintillans red tide monitoring and research work.
Considering the seasonal， nonlinear and non-stationary characteristics of air passenger demand series， an air passenger demand forecasting model based on a dual decomposition and reconstruction strategy was proposed. Firstly， the air passenger demand series was decomposed twice by Seasonal and Trend decomposition using Loess （STL） and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise （CEEMDAN） methods， and the components were reconstructed based on the feature analysis results of complexity and correlation. Then， Seasonal AutoRegressive Integrated Moving Average （SARIMA）， AutoRegressive Integrated Moving Average （ARIMA）， Kernel based Extreme Learning Machine （KELM） and Bidirectional Long Short-Term Memory （BiLSTM） network models were selected by model matching strategy to predict each reconstructed component respectively， among which the hyperparameters of KELM and BiLSTM models were determined by the Adaptive Tree of Parzen Estimators （ATPE） algorithm. Finally， the prediction results of the reconstruction components were linearly integrated. The air passenger demand data collected from Beijing Capital International Airport， Shenzhen Bao’an International Airport and Haikou Meilan International Airport were taken as research subjects for one-step and multi-step ahead prediction experiments. Experimental results show that compared with the single decomposition ensemble model STL-SAAB， the proposed model has the Root Mean Square Error （RMSE） improved by 14.98% to 60.72%. It can be seen that guided by the idea of “divide and rule”， the proposed model combines model matching and reconstruction strategies to extract the inherent development pattern of the data， which provides a new thinking to scientifically predict the change of air passenger demand.
Focused on the issue that real interaction targets cannot be matched with the virtual interaction targets one by one when providing passive haptics for redirected walking users in a vast Virtual Reality （VR） space， a method with two physical proxies acting as haptic proxies to provide haptic feedback for multiple virtual targets was proposed， in order to meet the user’s passive haptic needs alternately during the redirected walking process based on Artificial Potential Field （APF）. Aiming at the misalignment of virtual and real targets caused by the redirected walking algorithm itself and inaccurate calibration， the position and orientation of the virtual target were designed and haptic retargeting was introduced in the interaction stage. Simulation experimental results show that the design of the virtual target position and orientation can reduce the alignment error greatly. User experiments prove that haptic retargeting further improves the interaction accuracy and can bring users a richer and more immersive experience.
During the clustering process of high?dimensional data with feature missing， there are problems of the curse of dimensionality caused by data high dimension and the invalidity of effective distance calculation between samples caused by data feature missing. To resolve above issues， a K?Nearest Neighbor （KNN） imputation subspace clustering algorithm for high?dimensional data with feature missing was proposed， namely KISC. Firstly， the nearest neighbor relationship in the subspace of the high?dimensional data with feature missing was used to perform KNN imputation on the feature missing data in the original space. Then， multiple iterations of matrix decomposition and KNN imputation were used to obtain the final reliable subspace structure of the data， and the clustering analysis was performed in that obtained subspace structure. The clustering results in the original space of six image datasets show that the KISC algorithm has better performance than the comparison algorithm which clusters directly after interpolation， indicating that the subspace structure can identify the potential clustering structure of the data more easily and effectively； the clustering results in the subspace of six high?dimensional datasets shows that the KISC algorithm outperforms the comparison algorithm in all datasets， and has the optimal clustering Accuracy and Normalized Mutual Information （NMI） on most of the datasets. The KISC algorithm can deal with high?dimensional data with feature missing more effectively and improve the clustering performance of these data.
The onion router （Tor） anonymous communication system is used by criminals to engage in criminal activities on the dark networks， which brings severe challenges to social security. Tor website traffic is captured and analyzed by Tor website traffic analysis technology and therefore illegal behaviors hidden on the internet are timely discovered to conduct network supervision. Based on this， a Tor website traffic analysis model based on Self-Attention and Hierarchical SpatioTemporal （SA-HST） features was proposed on the basis of self-attention mechanism and spatiotemporal features. Firstly， attention mechanism was introduced to assign different weights to the network traffic features to highlight the important features. Then， Convolutional Neural Network （CNN） with multi-channel parallel structure and Long Short-Term Memory （LSTM） network were used to extract the spatiotemporal features of input data. Finally， Softmax function was used to classify data. SA-HST can achieve 97.14% accuracy in closed world scenario， which is 8.74 percentage points and 7.84 percentage points higher compared to CUMUL（CUMULative sum fingerprinting） model and deep learning model CNN. In open world scenario， SA-HST has the evaluation indicators of confusion matrix above 96% stably. Experimental results show that self-attention mechanism can achieve efficient feature extraction under lightweight model structure. By capturing important， multi-view spatiotemporal features of anonymous traffic for classification， SA-HST has certain advantages in terms of classification accuracy， training efficiency and robustness.
Aiming at the dynamic average consensus tracking problem of multi-agent systems with time-varying reference inputs， a proportional-integral consensus tracking algorithm was proposed. In the scenario of communication data between multi-agents being quantized， the average consensus tracking problem based on quantization was studied. Firstly， on the basis of the integral algorithm， a proportional link was introduced to make agents to track the average value of the reference inputs better by communicating with neighborhood agents under the constraints of the control agreement. Then， under the fixed， strongly connected and balanced topology structure， sufficient conditions for the multi-agent system asymptotically tracking to the average value of time-varying reference inputs without and with quantized information transmission data were obtained by using matrix analysis and Routh criteria respectively. Finally， numerical simulations verify the accuracy of the results and confirm the effectiveness of the proposed algorithm.
In order to retrieve images of traffic violations from a large number of road traffic images， a semantic retrieval method based on specific object self-recognition was proposed. Firstly， road traffic domain ontology as well as road traffic rule description were established by experts in traffic domain. Secondly， traffic image features were extracted by Convolutional Neural Network （CNN）， and combined with the strategy for classifying image features which is based on the proposed improved Support Vector Machine based Decision Tree （SVM-DT） algorithm， the specific objects and the spatial positional relationship between the objects in the traffic images were automatically recognized and mapped into the association relationship （rule instance） between the corresponding ontology instance and its objects. Finally， the image semantic retrieval result was obtained by reasoning based on ontology instances and rule instances. Experimental results show that the proposed method has higher accuracy， recall and retrieval efficiency compared to keyword and ontology traffic image semantic retrieval methods.
Aiming at the sharp increasing of data on the cloud caused by the development and popularization of cloud native technology as well as the bottlenecks of the technology in performance and stability， a Haystack-based storage system was proposed. With the optimization in service discovery， automatic fault tolerance and caching mechanism， the system is more suitable for cloud native business and meets the growing and high-frequent file storage and read/write requirements of the data acquisition， storage and analysis industries. The object storage model used by the system satisfies the massive file storage with high-frequency reads and writes. A simple and unified application interface is provided for business using the storage system， a file caching strategy is applied to improve the resource utilization， and the rich automated tool chain of Kubernetes is adopted to make this storage system easier to deploy， easier to expand， and more stable than other storage systems. Experimental results indicate that the proposed storage system has a certain performance and stability improvement compared with the current mainstream object storage and file systems in the situation of large-scale fragmented data storage with more reads than writes.
As a "killer" application in cloud computing, cloud gaming is leading the revolution of way of gaming. However, the high latency between the cloud and end devices hurts user experience. Aiming at the problem, a low latency cloud gaming system deployed on edge nodes, called Gaming@Edge, was proposed based on edge computing concept. To reduce the overhead of edge nodes for improving the concurrency, a cloud gaming running mechanism based on compressed graphics streaming, named GSGOD (Graphics Stream based Game-on-Demand), was implemented in Gaming@Edge system. The logic computing and rendering in the game running were separated and a computing fusion of edge nodes and end devices was built by GSGOD. Moreover, the network data transmission and latency were optimized through the mechanisms such as data caching, instruction pipeline processing and lazy object updating in GSGOD. The experimental results show that Gaming@Edge can reduce average network latency by 74% and increase concurrency of game instances by 4.3 times compared to traditional cloud gaming system.
In the recent years, the computer image understanding has wide and profound applications in intelligence traffic, satellite remote sensing, machine vision, image analysis of medical treatment, Internet image search and etc. As its extension, the image holistic scene understanding is more complex and integrated than basic image scene understanding task. In this paper, the basic framework for image understanding, the researching implication and value, typical models for image holistic scene understanding were summarized. The four typical holistic scene understanding models were introduced, and the model frameworks were thoroughly compared. At last, some research insufficiency and future direction in image holistic scene understanding were presented, which pointed out some new insights for the further research in this area.
