As smart mobile devices become increasingly popular，mobile applications such as object detection in videos are greatly limited by the computing and storage capacity of mobile devices． Traditional cloud computing can not meet the requirements of applications for network delay，network jitter，security and other issues． For this purpose，this paper designed the quality of experience ( QoE) as a comprehensive measure of object detection accuracy and energy consumption based on end-cloud collaborative system． Regarding the quality of experience as the optimization goal，this paper then proposed an adaptive scheduling algorithm for video image detection and tracking． The algorithm schedules object detection and object tracking by predicting network bandwidth and calculating transmission delay． Experimental results on the KITTI video dataset show that the detection-tracking adaptive scheduling algorithm can achieve a higher QoE value，significantly reduce the energy loss，and achieve an detection accuracy of 78. 3% ．

For the detection of non-uniform and non-periodic irregular texture surface defects，it is difficult to achieve conventional image reconstruction as the background texture is a non-stationary signal． It is also difficult to obtain sufficient information about the defects such as shape，gray scale and other image features in advance． And the visual features of similar defects can present great dispersion． The existing detection methods based on target features are inapplicable． Therefore，this paper proposed an unsupervised learning surface defect detection method based on image inpainting． Firstly，the method used a small number of normal texture samples as training set to train the network model． Then，the missing area was set artificially in the sample image and the network model was used to predict the content of the missing areas during the testing stage． By combining the structural similarity evaluation and residual error of the reconstructed image and the image to be tested，the defect detection and separation were realized． The experimental results show that the proposed method can not only effectively detect the defects on the regular texture surface but also detect the defects of irregular texture surfaces effectively，showing good practicability and adaptability．

In view of the problems of sparsity and underutilization of the heterogeneity of auxiliary information faced by current collaborative filtering methods and the advantages of heterogeneous information networks ( HIN) in modeling complex heterogeneous information，a HIN based tightly coupled recommendation model ( HTCRec) was proposed in this paper． It utilizes the heterogeneous information network embedding and a tightly coupled collaborative filtering framework to carry out personalized recommendation． Firstly，it aggregates meta-paths in a HIN and their corresponding path instances． Then it uses the attention mechanism to represent the auxiliary information of the target users or items in terms of the embedding of the respective aggregation meta-paths． At last，the meta-path is explicitly incorporated into the tightly coupled interaction model for personalized recommendation． The experimental results of the real data sets show that compared with the state-of-the-art recommendation models，the HTCRec model has better recommendation performance and effectively alleviates the problem of data sparsity．

The optimal rapid expansion randomized tree ( RRT* ) is an asymptotically optimal path planning method for mobile robots． Quick-RRT* reduces the initial path length of RRT* and increases the path convergence speed． In order to further improve the convergence speed of Quick-RRT* ，this paper proposed a dual-tree Quick-RRT* ( Quick-RRT* -Connect) algorithm． Firstly，two random trees were generated at the start and end points respectively based on the Quick-RRT* algorithm． Two trees grew in turn and they were connected with greedy method． Then， the probability completeness and asymptotic optimality of the proposed algorithm were analyzed and testified． Finally，based on the Matlab platform，Quick-RRT* -Connect was compared with RRT* ，Quick-RRT* and RRT* - Connect in three environments． The results show that the improved algorithm can not only find initial path and suboptimal path in a shorter time，but also reduce the initial path length．

As a common implementation of recommender system，collaborative filtering can bring personalized recommendation service experience to users． Traditional collaborative filtering models do not mine and analyze the attention level of different explicit attributes of users and items，leading to the critical level of different explicit attributes not paid attention by the model． Therefore，based on the coupled collaborative filtering model based on convolutional neural network，an attention mechanism was introduced in the paper to deeply mine the critical degree of explicit attributes and enhance the parameter learning gradient on critical attributes． And a new method of calculating the coupling degree was proposed to ensure the flushness of parameters and to improve the recommendation performance of the model． The experimental results show that the recommendation accuracy rate of the model proposed in the paper is better than that of traditional collaborative filtering methods and coupled collaborative filtering models，and the cumulative gain of topK@ 10 hit ratio and normalized discount cumulative gain reach 0. 8508 and 0. 5850，respectively．

A vehicle recognition algorithm based on improved CenterNet was proposed to solve the problem of low target recognition rate in vehicle recognition system． Firstly，the algorithm used ResNet18 as the basic network to reduce network parameters． Secondly，in view of the problem that the CenterNet model has the demerits of inaccurate positioning of vehicles，it was proposed to use the distance intersection over union loss to replace offset loss and width-high loss; meanwhile，single-scale spatial feature fusion ( SASFF) method and adaptive hierarchical feature fusion ( AHFF) method were employed to fuse several feature maps of the network． The experimental results show that on the Vehicle data set，mean average precision of the improved CenterNet models increases by 1. 9% ; on the BDD100K and Pascal VOC data sets，average precision at 0. 5 of the intersection over union between the predicted boxes and the true boxes increase by 5. 2% and 2. 5% ，respectively，and the inference speed on GTX1080Ti can reach 149 f /s． The improved CenterNet proposed in this paper can significantly improve vehicle recognition accuracy．

Building energy consumption prediction needs both time series features and categorical features, but traditional models can only deal with one of the features. Aiming at this problem, a new neural network integrating one-dimensional convolutional kernel and word embedding was proposed in this paper. The one-dimensional convolutional kernel can extract the continuous time series features, and the word embedding model can embed the discrete categorical features, based on which a building energy consumption prediction model is established that can simultaneously process both time series features and categorical features. By comparing with the gradient boosting decision regression tree and the long short time memory network, it is proved that the proposed model has good performance in efficiency and accuracy. In terms of hyperparameter adjustment, the automatic hyperparameter optimization algorithm based on Bayesian optimization was adopted, and the algorithm can find the optimal hyperparameter in the search space. Compared with manual optimization, the automatic hyperparameter optimization algorithm can improve the perfor-mance of the model in a relatively short time. Finally, simulation studies were conducted and the results demonstrate that the proposed model is better in performance than the ensemble learning model and the long short-time memory network.

In order to establish a permanent and stable task distribution link in the perceptual service process, firstly, an intimacy quantification method based on social attributes(IQSA) was proposed; secondly, a community detection algorighm based on information entropy similarity（CDIES）was proposed by combining the information entropy theory with social relationship; finally, IQSA algorithm was compared with two popular models by experiments, and the accuracy and validity of CDIES algorithm was assessed according to the result of community devision, modularity and time cost. The experimental results show that, compared with the content-based friend recommendation and relationship-based two typical recommendation algorithms, the IQSA algorithm has best comprehensive performance in accuracy, recall, and f1-score. And the modularity and time cost of the result of community devision of CDIES algorithm outperforms that of GN algorithm and FN algorithm. 

Aiming at the poor matching effect of ill-conditioned regions and excessive model parameters in the stereo matching network based on deep learning, an end-to-end stereo matching network based on multi-level feature fusion and recurrent cost aggregation（MFRANet）was proposed. Firstly, in order to take into account both the low-level detail information and high-level semantic information of the image, a multi-stage feature fusion module, which uses a phased and step-by-step feature fusion strategy to effectively fuse multi-level and multi-scale features, was proposed. Secondly, a recurrent mechanism was proposed in the cost aggregation stage to optimize the aggregation of the matching cost volume in a recurrent manner, and it can improve the aggregation effect while avoid introducing too many parameters. Finally, the disparity calculation module based on the Soft Argmin algorithm was used to calculate the image disparity. And through the two public datasets of KITTI 2012/2015 and SceneFlow, the network was trained and tested, and a comparative study with other end-to-end stereo matching networks was caaried out. Experimental results show that, for the two public datasets of SceneFlow and KITTI 2015, MFRANet has more accurate matching results than other end-to-end stereo matching networks; for the SceneFlow dataset, the end-point error is reduced to 0.92 pixels; for the KITTI 2015 dataset, the mismatching rate is reduced to 2.21%.

Due to the limited samples of footprint data and its high similarity between types and large gap within a type, there is no effective method to express footprint data and classify footprint. In order to solve the problem of bimodal footprint classification, the multiple-module relation network (MulRN) based on few-shot learning was proposed in this paper. Multiple modules were applied in the algorithm to improve the ability of extraction and mea-surement of characters. Inception module and MRFB module which possess a multi-branch structure were used to improve the character extraction ability. Spatial Attention Module (SAM) and Channel Attention Module (CAM) were adopted to extract the character of footprint with high discrimination for accurate classification. Also, experiments were carried out on few-shot data sets such as miniImageNet, Omniglot and bimodal 2D footprint data sets. Experimental results show that the proposed method is effective for few-shot data sets and bimodal 2D footprint data sets. It is worth mentioning that the accuracy of 5-way 5-shot experiment on bimodal data sets of right foot is up to 95.41%.

A fast point feature histogram descriptor algorithm combined with point cloud texture information is proposed to improve the feature extraction efficiency and matching accuracy in the point cloud matching and recognition process. Firstly, a shape feature histogram was constructed based on the fast point feature histogram descriptor and a texture feature histogram was constructed by using CIELab color space and multiple point-to-texture attribute metrics. Then the two feature histograms were connected to obtain a fast point feature histogram descriptor combined with point cloud texture information. The feature histogram descriptor was verified by using public point cloud data sets and real spot cloud data. and the feature matching test and point density change test were carried out for this feature descriptor and multiple existing descriptors. The test results show that the comprehensive performance of the descriptor is the best when the CIE00 color difference is used as the point-to-texture attribute metrics. The algorithm has a good feature description performance and high feature extraction efficiency and matching efficiency and it has strong robustness when the point cloud density changes.

Most score prediction studies are based on the assumption that the missing values are random． However， the missing data of the score matrix of the actual on-line recommendation system is non-random． Incorrect assumptions about the missing data can lead to biased parameter estimation and prediction． In order to improve the accuracy of non-random missing score matrix filling，the internal principle of user and item score matrix was analyzed in this paper． It presents a method to transform the score matrix of user and object into the equivalent bilateral block diagonal matrix by row or column transformation． Then the matrix decomposition method was applied to different blocks to decompose and predict the score，making local data update and decomposition become a reality． The experimental results on the public test dataset show that the proposed method can improve the score filling effect，solve the problem of non-random score missing effectively，and improve the prediction accuracy of the recommendation system． The improved block matrix also has a better speedup ratio in the distributed processing experiment，which shows that the proposed method has better scalability．

The designing thought of verification equations in PGHR protocol is singular，that is，a verification equation targets only one constraint，ignoring their joint effect． In this paper，a new ZK-SNARK protocol——— CPGHR protocol———was obtained by using the indivisible properties of additional constant coefficient factors and the combined effect of verification equations to achieve effective compression of the PGHR protocol． At the same time，a strict verification of the security of the new protocol was given，and the validity of the protocol was theoretically analyzed and experimentally verified． The results show that the amount of evidence of the new protocol is reduced by about 75% ，and the calculation efficiency is improved by about 33% ．

Image and point cloud are the common data formats for 3D object detection，for images have a superior object recognition capability and point clouds contain accurate spatial information． In order to utilize the above mentioned advantages of both images and point clouds，a 3D object detection method named Bird-PointNet based on bird's eye view of point cloud remapping approach was proposed． First，point cloud was encoded into bird's eye view format for object recognition and rough positioning． Then the results from bird's eye view detection was remapped into the point cloud's space for precise detection． Experiments on the BEV detection benchmark and the 3D detection benchmark of KITTI dataset have demonstrated that the proposed Bird-PointNet method has a higher accuracy of 3D detection，compared with the baseline method that only with bird's eye view coding of point cloud．

For the classic perspective-three-point ( P3P) problem，when the Z-axis coordinates of the three-dimensional control points are randomly distributed in a large range，there are still problems of poor numerical stability， degradation caused by increased image noise，and low computational efficiency． Therefore，a fast and stable algebraic solution method was proposed in this article． Firstly，when the proposed solution directly estimates the rotation and position of a calibrated camera from three 3D to 2D point correspondences，an intermediate coordinate frame is introduced between the world coordinate frame and the camera coordinate frame to reduce the number of unknown parameters，and the rotation matrix is normalized to simplify the calculation process and improve the calculation efficiency． Secondly，the midpoint between the two control points was chosen as the origin point of the intermediate coordinate frame，so as to improve the anti-noise performance of the P3P problem in degenerate configurations． Finally，the P3P problem was transformed into a biquadratic equation with one unknown parameter by using a Grbner basis，then a closed solution to the P3P problem was obtained． Experimental results show that the proposed algorithm can achieve better numerical stability and anti-noise performance compared with other three classic algorithm of the P3P problem.

A drug interaction extraction model combining category key words with attention mechanism was proposed to enhance the discrimination among different categories of data and improve the performance of classifier． Firstly，the keywords of each class were selected based on the chi-square test and document frequency． Then，the position coding of keywords and drug pairs was added into the pre-trained model BERT，in order to make the difference of the samples more salient． The distribution information of keywords and other words in the sentence was learned through the attention mechanism to improve the performance of the model． Aiming at the problem of too much negative samples in the drug interaction extraction experiment，a negative sample filtering method based on rules and patterns was proposed to effectively reduce the proportion of positive and negative samples． Compared with other DDI models based on CNN，LSTM，and BERT，KA-BERT model can better improve performance on DDI data，which proves the effectiveness of KA-BERT model． The results of the test on chemical protein relation extraction show that the precision，recall and F1 score of KA-BERT model are enhanced significantly，which further proves the validity and universality of KA-BERT model．

Since the outbreak of the covid-19 epidemic，related rumors have spread rampantly． Traditional rumor identification models have difficulties in epidemic rumor identification because the size of epidemic rumors is not large enough to train a good classification and identification model． Therefore，it is an urgent task to build a rumor identification model based on a small amount of epidemic rumor data． To deal with the problem of insufficient training data，text enhancement and generative adversarial networks ( GAN) methods were used to generate a large amount of information similar to epidemic rumors and to improve the identification effect of epidemic rumors． First， the textual characteristics was analyzed to extract keyword of epidemic rumors． Second，epidemic rumor generation model was constructed based on the idea of GAN，and historical rumors which do not contain epidemic rumor features were textually enhanced by the epidemic rumor feature thesaurus，and a large amount of new rumor data containing epidemic rumor features were generated． Finally，the newly generated rumor data are combined with the epidemic rumor data to train a more accurate classification model of the epidemic rumor． Experiment results show that the rumor identification effect is improved by 3% after using the GAN extended training set． The new model is evidently much better than the traditional machine learning and deep learning algorithms，and it provides a new way for the identification of rumors in public health emergency．

As an important content of judicial information disclosure，case documents should be disclosed to the public after the trial． Some case documents involving juvenile are likely to cause the disclosure of juvenile personal privacy information． In order to conduct targeted privacy protection processing，the first step is to accurately identify documents involving juvenile information from a large number of case documents． At the same time，in order to solve the problem of difficulty in effective supervised learning due to the lack of labeled samples in the real data set，this paper proposed a juvenile case documents recognition method based on semi-supervised learning． Firstly， the corpus text of the case document was pre-processed，and then the features of the text were extracted with Word2Vec and BEＲT-wwm-ext． After the above processing，the long corpus text was converted into the data format that can be used as the input for the classification model． Then the classification model was trained with the PU learning method，and an effective classifier was constructed with a large number of unlabeled samples under the condition of few positive examples． Then，based on the prediction results of the classification model，active learning method was employed to obtain keywords and screen the prediction results，so as to further improve the prediction effect． Finally，the case documents recognition method proposed in this article achieves a recall of 98. 67% and a precision of 81. 02% on the test set constructed based on the proportion of real scenes．