Traditional Hyper-Spectral Image （HSI） spatio-spectral fusion algorithms usually use static spectral dictionary， in which the dictionary learning and the image fusion are two separate processes， thereby giving poor performance when processing noisy spatio-spectral fusion tasks. To address this problem， a noisy HSI spatio-spectral fusion algorithm based on Dynamic Dictionary Learning （DDL） was proposed， which adopted an iterative strategy that updates dictionary atoms dynamically during the fusion process， thereby collaborating to complete the spatio-spectral fusion and noise removal tasks. Firstly， a coarse denoising was performed on the input HSI and the denoising result was utilized to initialize the spectral dictionary. Secondly， the sparse representation technique was employed to fuse the two input images with the above initialized dictionary， resulting an intermediate fusion image. Thirdly， the intermediate fusion image was fed back to the dictionary learning module to update the dictionary atoms continuously， thereby forming a dynamic spectral dictionary. Finally， by iterating the above process， the final output image was obtained. Simulation results on three remote sensing HSI datasets show that the proposed algorithm can remove noise effectively while improving spatial resolution of the images. At the same time， experimental results on real noisy image bands indicate that the proposed algorithm can improve visual quality of the fused images effectively. On Cuprite Mine dataset， the Peak Signal-to-Noise Ratio （PSNR） of the proposed algorithm is increased by 32.48% and 10.72% respectively compared to those of Generalized Tensor Nuclear Norm （GTNN） method and AL-NSSR method — the method of denoising first and then fusion， with Gaussian noise variance of 0.15 and amplification factor of 8.
In practical applications， multi-view metric learning has become an effective method for handling multi-view data. However， the incompleteness of multi-view data poses significant challenges for multi-view metric learning. Although some methods have attempted to address incomplete multi-view issue， they still have the following shortcomings： 1） most of the existing methods rely on k-Nearest Neighbors （kNN） of the existing samples to fill in missing data， and ignore unique characteristics of samples or views easily； 2） they only utilize the existing sample representations to calculate neighbors， and cannot fully express neighbor relationships between samples. To address these issues， a Dual imputation based Incomplete Multi-View Metric Learning method （DIMVML） was proposed. Firstly， latent features of each view were extracted using a deep autoencoder， and then missing samples were filled in by combining distribution information of samples and difference information between views. Secondly， the results were fused according to quality of the completed samples to obtain higher-quality completion results. Finally， intra-view and inter-view relationships were optimized through a loss function. Experimental results show that in clustering experiments， the proposed method achieves superior accuracy and F1 score on HandWritten， Caltech101-7， Leaves， and YouTubeFace10 datasets compared to advanced multi-view methods such as Subgraph Propagation and Contrastive Calibration （SPCC） and Latent Heterogeneous Graph Network （LHGN）； in classification experiments， the proposed method outperforms other multi-view methods significantly in accuracy on CUB， ORL， and HandWritten datasets.
Aiming at the problems of incomplete information dimension， sparse interaction data and redundant interaction information in recipe recommendation tasks， a Recipe recommendation model based on hierarchical learning of Flavor embedding heterogeneous graph （RecipeFlavor） was proposed. Firstly， the flavor molecule dimension was introduced， and a heterogeneous graph was constructed on the basis of users， foods， ingredients and flavor substances of ingredients to represent the connection among four kinds of nodes effectively. Then， a hierarchical learning module based on heterogeneous graph was constructed on the basis of information transmission mechanism， and combined with Squeeze Attention （SA） mechanism， different node relationships were regarded as different information channels， so that key interaction information between nodes was extracted and noise was suppressed. Finally， a Contrastive Learning （CL） module was constructed on the basis of feature-aware noise， and positive and negative sample discrimination tasks were introduced in model learning， thereby enhancing the information associations among users and recipe nodes and improving the model’s learning ability for features. Experimental results show that compared with HGAT （Hierarchical Graph ATtention network for recipe recommendation） model on Recipe 1M+ large dataset， RecipeFlavor has the Area Under the ROC Curve （AUC） increased by 1.44 percentage points， and the model Precision （Pre）， Hit Rate （HR）， Mean Average Precision （MAP）， and Normalized Discounted Cumulative Gain （NDCG） of Top-10 increased by 0.76， 6.11， 2.68， and 3.05 percentage points， respectively. It can be seen that the introduction of flavor molecule information expands the learning dimension of recipe recommendation， and RecipeFlavor can extract key information in heterogeneous graph effectively， and enhance correlation among users and recipes， and thus improving the precision of recipe recommendations.
Aiming at the shortcomings of the original Aquila Optimizer （AO）， such as insufficient local development ability， low optimization accuracy and slow convergence speed， a Multi-Strategy Improved AO （MSIAO） for robot path planning was proposed. Firstly， the Sobol sequence was introduced to initialize the Aquila population， which was conducive to diversity of the initial population and improved the convergence speed. Secondly， the local search method was improved by using golden sine operator and idea of self-learning and social learning of particle swarm， which enhanced exploitation ability of the algorithm and reduced the possibility of falling into the local optimum. Meanwhile， a non-linear balance factor was used as switching condition of the two stages， which made better communication among the populations， and was able to balance the global exploration and local exploitation more effectively. Finally， multiple experiments were carried out. Through the simulation on 12 benchmark functions and 10 CEC2017 complex functions， it can be seen that the proposed improvement strategies enhance the global optimization ability of MSIAO greatly. Results of applying MSIAO to robot path planning show that MSIAO can obtain shorter and more reliable moving paths. In 20×20 grid map， the average path of MSIAO is shortened by 2.53%， 3.83%， and 6.70% compared to those of Particle Swarm Optimization （PSO） algorithm， the original AO， and Butterfly Optimization Algorithm （BOA）， respectively； and in 40×40 grid map， the average path of MSIAO is shortened by 10.65%， 5.27%， and 14.88% compared to those of the above three algorithms， verifying that the path-finding of MSIAO is more efficient.
To address the problems of low multi-classification accuracy， poor generalization， and easy privacy invasion in traditional encrypted traffic identification methods， a multi-classification deep learning model that combines Attention mechanism （Attention） with one-Dimensional Convolutional Neural Network （1DCNN） was proposed， namely Attention-1DCNN-CE. This model consists of three core components： 1） in the dataset preprocessing stage， the spatial relationship among packets in the original data stream was retained， and a cost-sensitive matrix was constructed on the basis of the sample distribution； 2） based on the preliminary extraction of encrypted traffic features， the Attention and 1DCNN models were used to mine deeply and compress the global and local features of the traffic； 3） in response to the challenge of data imbalance， by combining the cost-sensitive matrix with the Cross Entropy （CE） loss function， the sample classification accuracy of minority class was improved significantly， thereby optimizing the overall performance of the model. Experimental results show that on BOT-IOT and TON-IOT datasets， the overall identification accuracy of this model is higher than 97%. Additionally， on public datasets ISCX-VPN and USTC-TFC， this model performs excellently， and achieves performance similar to that of ET-BERT （Encrypted Traffic BERT） without the need for pre-training. Compared to Payload Encoding Representation from Transformer （PERT） on ISCX-VPN dataset， this model improves the F1 score in application type detection by 29.9 percentage points. The above validates the effectiveness of this model， so that this model provides a solution for encrypted traffic identification and malicious traffic detection.
In recent years， with the rapid development of deep learning technology， entity and relation extraction has made remarkable progress in many fields. However， due to complex syntactic structures and semantic relationships of Chinese text， there are still many challenges in Chinese entity and relation extraction. Among them， the problem of overlapping triple in Chinese text is one of the important challenges. A Hybrid Neural Network Entity and Relation Joint Extraction （HNNERJE） model was proposed in this article to address the issue of overlapping triple in Chinese text. HNNERJE model fused sequence attention mechanism and heterogeneous graph attention mechanism in a parallel manner， and combined them with a gated fusion strategy， so that it could capture both word order information and entity association information of Chinese text， and adaptively adjusted the output of subject and object markers， effectively solving the overlapping triple issue. Moreover， adversarial training algorithm was introduced to improve the model’s adaptability in processing unseen samples and noise. Finally， SHapley Additive exPlanations （SHAP） method was adopted to explain and analyze HNNERJE model， which effectively revealed key features in extracting entities and relations. HNNERJE model achieved high performance on NYT， WebNLG， CMeIE， and DuIE datasets with F1 score of 92.17%， 93.42%， 47.40%， and 67.98%， respectively. The experimental results indicate that HNNERJE model can transform unstructured text data into structured knowledge representations and effectively extract valuable information.
Existing robotic grasping operations are usually performed under well-illuminated conditions with clear object details and high regional contrast. At the same time， for low-light conditions caused by night and occlusion， where the objects’ visual features are weak， the detection accuracies of existing robotic grasp detection models decrease dramatically. In order to improve the representation ability of sparse and weak grasp features in low-light scenarios， a grasp detection model incorporating visual feature enhancement mechanism was proposed to use the visual enhancement sub-task to impose feature enhancement constraints on grasp detection. In grasp detection module， the U-Net like encoder-decoder structure was adopted to achieve efficient feature fusion. In low-light enhancement module， the texture and color information was respectively extracted from local and global level， thereby balancing the object details and visual effect in feature enhancement. In addition， two low-light grasp datasets called low-light Cornell dataset and low-light Jacquard dataset were constructed as new benchmark dataset of low-light grasp and used to conduct the comparative experiments. Experimental results show that the accuracies of the proposed low-light grasp detection model are 95.5% and 87.4% on the benchmark datasets respectively， which are 11.1， 1.2 percentage points higher on low-light Cornell dataset and 5.5， 5.0 percentage points higher on low-light Jacquard dataset than those of the existing grasp detection models， including Generative Grasping Convolutional Neural Network （GG-CNN）， and Generative Residual Convolutional Neural Network （GR-ConvNet）， indicating that the proposed model has good grasp detection performance.
Domain adaptation algorithms are widely used for cross-corpus speech emotion recognition. However， many domain adaptation algorithms lose the discrimination of target domain samples while pursuing the minimization of domain discrepancy， resulting in their presence at the decision boundary of the model in a high-density form， which degrades the performance of the model. Based on the above problem， a Decision Boundary Optimized Domain Adaptation （DBODA） method based cross-corpus speech emotion recognition was proposed. Firstly， the features were processed by using convolutional neural networks. Then， the features were fed into the Maximum Nuclear-norm and Mean Discrepancy （MNMD） module to maximize the nuclear norm of the sentiment prediction probability matrix of the target domain while reducing the inter-domain discrepancy， thereby enhancing the discrimination of the target domain samples and optimize the decision boundary. In six sets of cross-corpus experiments set up on the basis of Berlin， eNTERFACE and CASIA speech databases， the average recognition accuracy of the proposed method is 1.68 to 11.01 percentage points ahead of those of the other algorithms， indicating that the proposed model effectively reduces the sample density around the decision boundary and improves the prediction accuracy.
With the development of high-throughput sequencing technology， massive genome sequence data provide a data basis to understand the structure of genome. As an essential part of genomics research， splice site identification plays a vital role in gene discovery and determination of gene structure， and is of great importance for understanding the expression of gene traits. To address the problem that existing models cannot extract high-dimensional features of DNA （DeoxyriboNucleic Acid） sequences sufficiently， a splice site prediction model consisted of BERT （Bidirectional Encoder Representations from Transformers） and parallel Convolutional Neural Network （CNN） was constructed， namely BERT-splice. Firstly， the DNA language model was trained by BERT pre-training method to extract the contextual dynamic association features of DNA sequences and map DNA sequence features with a high-dimensional matrix. Then， the DNA language model was used to map the human reference genome sequence hg19 data into a high-dimensional matrix， and the result was adopted as input of parallel CNN classifier for retraining. Finally， a splice site prediction model was constructed on the basis of the above. Experimental results show that the prediction accuracy of BERT-splice model is 96.55% on the donor set of DNA splice sites and 95.80% on the acceptor set， which improved by 1.55% and 1.72% respectively， compared to that of the BERT and Recurrent Convolutional Neural Network （RCNN） constructed prediction model BERT-RCNN. Meanwhile， the average False Positive Rate （FPR） of donor/acceptor splice sites tested on five complete human gene sequences is 4.74%. The above verifies that the effectiveness of BERT-splice model for gene splice site prediction.
To study the performance and application prospects of novel intelligent optimization algorithms， six bionic intelligent optimization algorithms proposed in the past few years were analyzed， concluding Harris Hawks Optimization （HHO） algorithm， Equilibrium Optimizer （EO）， Marine Predators Algorithm （MPA）， Political Optimizer （PO）， Slime Mould Algorithm （SMA）， and Heap-Based Optimizer （HBO）. Their performance and applications in different constrained engineering optimization problems were compared and analyzed. Firstly， the basic principles of six optimization algorithms were introduced. Secondly， the optimization tests were performed on ten standard benchmark functions for six optimization algorithms. Thirdly， six optimization algorithms were applied to solve three engineering optimization problems with constraints. Experimental results show that the convergence accuracy of PO is the best for the optimization of unimodal and multimodal test functions and can reach the theoretical optimal value zero many times. The EO and MPA are better for solving constrained engineering problems with fast optimization speed， high stability and standard deviation of a small order of magnitude. Finally， the improvement methods and development potentials of six optimization algorithms were analyzed.
Aiming at the problems of network training difficulty and low utilization rate of feature information caused by increasing network layers in super-resolution restoration technology， an image super-resolution restoration algorithm based on dual attention Information Distillation Network （IDN） was designed and implemented. Firstly， by taking the advantage of the low computational complexity of IDN and the advantage of the information distillation module by which more features were extracted， the weights of the features were readjust adaptively by introducing the Residual Attention Module （RAM） and considering the interdependence of image channels， so as to further improve the reconstruction ability of high-resolution details of images. Then， a new mixed loss function sensitive to edge information was designed to refine the image and accelerate the convergence of the network. Test results on Set5， Set14， BSD100 and Urban100 public datasets show that the visual effect and Peak Signal-to-Noise Ratio （PSNR） of the proposed method are superior to those of the current mainstream algorithms.
In data layer, the course model and resource model were built based on Markov chain and vector space model, and the teacher model was built based on teachers' personal registration information and nodes of course model. In off-line layer, the content features of course model and resource model were extracted via Term Frequency-Inverse Document Frequency (TF-IDF) algorithm, and the course model and resource model of data layer were initialized and optimized. Then relations between any two resources or recourse and course were calculated using association rules mining and similarity measure, and intermediate recommendation results were given using teacher model and course model. A weighted hybrid recommendation algorithm was proposed to generate recommendation list in on-line layer. The proposed system has been successfully applied in a real education resources sharing platform which consists of 600 thousand teaching resources.
