The traditional Segment Anything Model （SAM） relies on manual prompts during segmentation of Meibomian gland images， making it difficult to handle issues such as dense glands， irregular shapes， and blurred boundaries. To address this， an improved model， namely ResSAM， was proposed. ResSAM eliminated the reliance on manual intervention by introducing an automatic prompt encoder. The backbone network was pruned and optimized to further enhance the model's segmentation efficiency. Focal Loss and Smooth IoU Loss were used for training optimization， and the SE （Squeeze-and-Excitation） and cross-attention mechanisms were integrated to reduce the impact of individual differences and blurred boundaries， thereby improving the model's segmentation accuracy. Experimental results on two self-built datasets， Lower Lid and Upper Lid， showed that ResSAM achieved the best performance in terms of the number of parameters and Giga FLoating-point OPerations （GFLOPs）； its segmentation results obtained the highest Dice scores （88.69% and 87.75%， respectively） and the highest Intersection-over-Union （IoU） values （79.69% and 78.58%， respectively）. The research results indicate that the ResSAM optimizes both efficiency and accuracy， supporting early prevention and clinical diagnosis of Meibomian Gland Dysfunction （MGD）.

Existing underwater small object detection methods are primarily based on deep learning algorithms， which face challenges in balancing lightweight design and detection accuracy， so that they unable to meet the requirements of real-time and resource-constrained platforms. Therefore， Graph-DETR， a lightweight underwater small object detection model based on RT-DETR （Real-Time DEtection TRansformer） and a graph Transformer， was proposed. The model used a lightweight MobileNetV4 backbone improved with the Large Separable Kernel Attention mechanism （LSKAttention） and the Context-Mixing dynamic convolutional block （CM block） to enhance feature extraction efficiency and reduce model complexity. Additionally， a hierarchical Graph Transformer Feature Pyramid Network （GTFPN） was proposed to strengthen multi-scale feature fusion， and the hybrid encoder was optimized via Wavelet Transform Convolution （WTConv）， Adaptive downsampling （Adown）， and path pruning， thereby achieving convolutional receptive field expansion of the CNN-based Cross-scale Feature Fusion （CCFF） module with low parameterization. Experimental results on the underwater public dataset URPC2020 show that， compared to RT-DETR， Graph-DETR reduces the parameters by 66.9% and the reasoning latency by 6.8 ms， achieving a mean Average Precision （mAP） of 53.2% and an Average Precision of 86.8% at an IoU threshold of 0.5 （AP@0.5）； on URPC2021， it has 81.3% recall， 54.1% mAP， 87.6% AP@0.5 with only 10.5 ms latency， outperforming the existing methods. Graph-DETR exhibits excellent performance in underwater small object detection and is practical for deployment on resource-constrained underwater platforms.

To address the problems posed by highly realistic AI-generated text， driven by the rapid development of Large Language Models （LLMs）， and the performance degradation of traditional detection methods， an adaptive multi-feature fusion detection method for AI-generated text was proposed. Firstly， a language style feature set covering text statistical features， language structural features， and language uncertainty features was constructed to capture differences between real and AI-generated texts； then， deep semantic features of texts were extracted using independent encoding technology. Based on these， a dual-path mapping feature-adaptive fusion strategy was designed： language-style features and deep semantic features were first fused at a primary level， and secondary fusion was then performed using deep learning to enhance the capability of adaptive feature fusion. Experimental results demonstrate that the proposed method achieves detection accuracies of 98.1% on the Chinese SocialAI-Detect dataset and 98.5% on the English TuringBench dataset； compared with the best-performing baseline， J-Guard （Journalism Guided adversarially robust detection of AI-generated news）， the improvements are 2.3 and 2.1 percentage points， respectively， verifying the effectiveness of the proposed method.

Distributed Denial of Service （DDoS） attacks， as a highly destructive type of cyber attacks， have become one of the most severe threats and challenges in the field of cybersecurity in recent years due to their low attack costs， high attack efficiency， and strong concealment. DDoS attacks employ a distributed control approach to mix malicious traffic with legitimate network requests， making it difficult for traditional security defense mechanisms such as Intrusion Detection System （IDS） and firewalls to identify and mitigate such attacks effectively. Consequently， the efficient detection and effective defense against DDoS attacks have become research hotspots and difficulties in the field of cybersecurity. Based on systematic survey of the existing research on DDoS attacks， the following was performed. Firstly， the classification methods of DDoS attacks were sorted out， and DDoS attacks were summed up from multiple perspectives， so as to provide a deeper understanding of DDoS attack mechanisms. Secondly， an analysis of the current development of DDoS attacks was conducted， with particular focuses on discussing the development trends in attack intensity， attack methods， and attack distribution， thereby providing support for the research on more efficient DDoS defense technologies. Thirdly， an in-depth analysis and evaluation of the status of DDoS attack defense technologies was conducted from both industrial and academic perspectives， which focused on DDoS detection and defense methods based on programmable switches and machine learning in the academic aspect， and compared and analyzed the defense architectures adopted by different participants in DDoS defense in the industrial aspect as well as summarized the technical characteristics， application scenarios， and the existing challenges of the architecture. Finally， based on a comprehensive analysis of the current DDoS attack situations， the future development directions， opportunities， and challenges of DDoS defense technology were prospected， providing new ideas and directions for researchers in the field of cybersecurity and promoting further innovation and development of DDoS defense technology.

For Expensive Multi-objective Optimization Problem （EMOP）， although numerous related algorithms have been proposed， most existing algorithms have not achieved satisfactory results. The primary reason is that the infill sampling criteria in these algorithms fail to balance the convergence， diversity and uncertainty of selected individuals. Therefore， a Two-stage Infill Sampling-based Expensive Multi-Objective Evolutionary Algorithm （TISEMOEA） was proposed. In the first stage， a convergence-based infill sampling criterion was proposed， so as to select individuals with both good convergence and diversity， and then balance convergence and diversity. In the second stage， a diversity-based infill sampling criterion was proposed， so as to select individuals with great uncertainty without damaging convergence， and then improve the accuracy of the model and the diversity of the population. Furthermore， an adaptive diversity enhancement strategy was proposed to adjust the frequency of selecting individuals using the diversity-based infill sampling criterion， thereby enhancing population diversity and balancing exploration and exploitation capabilities of the algorithm. TISEMOEA was compared with five state-of-the-art algorithms， MOEA/D-EGO （MOEA/D with the Gaussian process model）， HeE-MOEA （Heterogeneous Ensemble-based infill criterion for MOEA）， TISS-EMOA （Two-stage Infill Sampling-based Semi-supervised EMOA）， PCSAEA （Pairwise Comparison based Surrogate-Assisted Evolutionary Algorithm）， and SFA/DE （Evolutionary multiobjective optimization assisted by scalarization function approximation for high-dimensional expensive problems）， on the DTLZ and WFG test sets with 28 and 27 test problems， and the Inverted Generational Distance （IGD） metric was analyzed. The results show that TISEMOEA achieves the best results in 19 and 16 test problems， respectively.

When detecting small targets in multi-scale remote sensing images， target detection algorithms based on deep learning are prone to false detection and missed detection. One of the reasons is that the feature extraction module carries out multiple down-sampling operations. The second reason is the failure to pay attention to the contextual information required by different categories and different scales of targets. To solve this problem， a small object detection algorithm in remote sensing images integrating attention and contextual information ACM-YOLO （Attention-Context-Multiscale YOLO） was proposed. Firstly， to reduce the loss of small target feature information， fine-grained query aware sparse attention was applied， thereby avoiding missed detection. Secondly， to pay more attention to the contextual information required by different categories of remote sensing targets， the Local Contextual Enhancement （LCE） function was designed， thereby avoiding false detection. Finally， to strengthen multi-scale feature fusion capability of the feature fusion module on small targets in remote sensing images， the weighted Bi-directional Feature Pyramid Network （BiFPN） was adopted， thereby improving detection effect of the algorithm. Comparison experiments and ablation experiments were performed on DOTA dataset and NWPU VHR-10 dataset to verify effectiveness and generalization of the proposed algorithm. Experimental results show that on the two datasets， the proposed algorithm has the mean Average Precision （mAP） reached 77.33% and 96.12% respectively， and the Recall increases by 10.00 and 7.50 percentage points， respectively， compared with YOLOv5 algorithm. It can be seen that the proposed algorithm improves mAP and recall effectively， which reduces false detection and missed detection.

Both the Damped Least Squares （DLS） and Genetic Algorithm （GA） are applicable to automatic design of optical systems. Although DLS has a high search efficiency， it is susceptible to falling into local optima traps. Conversely， GA has strong global search capability in the parameter space of optical structures but weak local search capability. To address these challenges， a Correctable Reinforced Search GA （CRSGA） was proposed. Firstly， DLS was introduced after the GA crossover operation to enhance local search capability. Additionally， a correction strategy was introduced to rollback individuals with deteriorated fitness values before the next iteration， thereby achieving corrective evolutionary results. The improvement of two aspects to genetic algorithm enhanced strengths and compensated for weaknesses. Three typical optical system design experiments， including Double Gaussian （DG）， Reversed Telephoto （RT）， and Finite Conjugate Distance Imaging （FCDI）， were conducted to validate the effectiveness of CRSGA. CRSGA outperforms both DLS and GA， and its optimization outcomes are about 8.92%， 12.19%， and 9.39% respectively better than those of commercial optical design software Zemax DLS. In particularly， the optimization outcomes achieve a significant improvement， reaching 99.98%， 94.33%， and 88.45% respectively compared to the Zemax HAMMER algorithm. In conclusion， it is shown that the proposed algorithm is effective for optical system optimization and can be used for automatic optical system design.

Aiming at the problem of the existing drift detection methods in balancing the detection delay， false positives， false negatives， and spatiotemporal efficiency， a new stage transition threshold parameter was proposed， and a multi-stage weighting mechanism including “stable stage-warning stage-drift stage” was introduced in the concept drift detection to weight the instances in stages， and the mechanism was applied to the double sliding window. Then a Multi-Stage weighted Drift Detection Method （MSDDM） based on Hoeffding inequality was proposed. On artificial datasets， MSDDM detected abrupt and gradual concept drift faster than Fast Hoeffding Drift Detection Method （FHDDM）， Drift Detection Method based on Hoeffding’s bound （HDDM） and other drift detection methods， while maintained a low false detection rate and a false alarm rate. At the same time， MSDDM had the highest classification accuracy in most cases compared with other methods on real-world datasets. Experimental results show that MSDDM can detect concept drift in data streams with high drift detection performance and great spatiotemporal efficiency.

Aiming at the problem that the existing adversarial example generation methods require a lot of queries to the target model， which leads to poor attack effects， a Text Adversarial Examples Generation Method based on BERT （Bidirectional Encoder Representations from Transformers） model （TAEGM） was proposed. Firstly， the attention mechanism was adopted to locate the keywords that significantly influence the classification results without query of the target model. Secondly， word-level perturbation of keywords was performed by BERT model to generate candidate adversarial examples. Finally， the candidate examples were clustered， and the adversarial examples were selected from the clusters that have more influence on the classification results. Experimental results on Yelp Reviews， AG News， and IMDB Review datasets show that compared to the suboptimal adversarial example generation method CLARE （ContextuaLized AdversaRial Example generation model） on Success Rate （SR）， TAEGM can reduce the Query Counts （QC） to the target model by 62.3% and time consumption by 68.6% averagely while ensuring the SR of adversarial attacks. Based on the above， further experimental results verify that the adversarial examples generated by TAEGM not only have good transferability， but also improve the robustness of the model through adversarial training.

Aiming at the disadvantages of traditional Bald Eagle Search optimization algorithm （BES）， such as easy to fall into the local optimum and slow convergence， a BES with Golden Sine Algorithm （Gold-SA） and crisscross strategy （GSCBES） was proposed. Firstly， the position update formula based on inertia weight was set in the traditional BES search stage. Then， Gold-SA was introduced in the stage of predation. Finally， the crisscross strategy was introduced to modify the global optimum and population. The optimization ability of the proposed algorithm was evaluated by the simulation experiments on 11 Benchmark functions， CEC2014 functions and by using Wilcoxon rank sum test. The results show that the proposed algorithm converges faster. At the same time， the weights and thresholds of Back Propagation （BP） neural network were assigned by the proposed algorithm， and the optimized BP neural network model was used in the prediction of air quality， the values of Mean Absolute Error （MAE）， Root Mean Square Error （RMSE）， Mean Square Error （MSE）， and Mean Absolute Percentage Error （MAPE） are smaller than those of BP neural network model and Particle Swarm Optimization （PSO） based BP neural network model，and the prediction accuracy is improved.

Aiming at the problems of the mismatch between aspect words and irrelevant context and the lack of grammatical level features in Aspect-Based Sentiment Analysis （ABSA） at current stage， an improved ABSA model integrating match-Long Short-Term Memory （mLSTM） and grammatical distances was proposed， namely mLSTM-GCN. Firstly， the correlation between the aspect word and the context was calculated word by word， and the obtained attention weight and the context representation were fused as the input of the mLSTM， so that the context representation with higher correlation with the aspect word was obtained. Then， the grammatical distance was introduced to obtain a context which was more grammatically related to the aspect word， so as to obtain more contextual features to guide the modeling of the aspect word， and obtain the aspect representation through the aspect masking layer. Finally， in order to exchange information， location weights， context representations and aspect representations were combined， thereby obtaining the features for sentiment analysis. Experimental results on Twitter， REST14 and LAP14 datasets show that compared with Aspect-Specific Graph Convolutional Network （ASGCN）， mLSTM-GCN has the accuracy improved by 1.32， 2.50 and 1.63 percentage points， respectively， and has the Macro-F1 score improved by 2.52， 2.19 and 1.64 percentage points， respectively. Therefore， mLSTM-GCN can effectively reduce the probability of mismatch between aspect words and irrelevant context， and improve the classification effect.

To strengthen the control and management of local airspace routes， a route discovery method based on trajectory point clustering was proposed. Firstly， for the simulation data generated according to the distribution characteristics of the real data， the pre-processing module was used to weaken and remove the noise of the trajectory data. Secondly， a route discovery method including outlier elimination， trajectory resampling， trajectory point clustering， clustering center correction， and connecting clustering centers was proposed to extract the routes. Finally， the result of route extraction was visualized and the proposed method was validated using civil aviation data. The experimental results on the simulated data show that the node coverage and the length coverage of the proposed method is 99% and 94% respectively， under the noise intensity of 0.1° and the buffer area of 30 km. Compared with the rasterization method， the proposed method has higher accuracy and can extract the routes more effectively， achieving the purpose of extracting the common routes of aircraft.

Test architecture design is the basic and key issue of Integrated Circuit (IC) test, and the design of effective test architecture that meet the needs of IC is of great importance to reduce chip cost, improve product quality and increase product competitiveness. Therefore, a test architecture with several ring scan chains using the same test pin in the round robin manner was proposed, namely RRR Scan. In RRR Scan, the scan flip-flops were designed as multiple ring scan chains, which can work in stealth scan mode, ring shift scan mode and linear scan mode. The ring shift scan mode enables the reuse of test data, thus reducing the size of the test set; the stealth scan mode can shorten the test data shifting path, thus significantly reduing the test shifting power consumption, so that the architecture is a general test architecture with the characteristics of data reuse and low power consumption. In addition, in the architecture, the physically adjacent scan cells can be set into the same ring scan chain with little wiring cost. With stealth scan mode, both the shifting length and the delay of test data can be reduced. Experimental results show that the shifting power consumption can be reduced greatly by RRR Scan, and for S13207 circuit, the shifting power consumption is only 0.42% of that of the linear scan.

Image fusion technology plays an important role in computer-aided diagnosis. Detail extraction and energy preservation are two key issues in image fusion, and the traditional fusion methods address them simultaneously by designing the fusion method. However, it tends to cause information loss or insufficient energy preservation. In view of this, a fusion method was proposed to solve the problems of detail extraction and energy preservation separately. The first part of the method aimed at detail extraction. Firstly, the Non-Subsampled Shearlet Transform (NSST) was used to divide the source image into low-frequency and high-frequency subbands. Then, an improved energy-based fusion rule was used to fuse the low-frequency subbands, and an strategy based on the intuitionistic fuzzy set theory was proposed for the fusion of the high-frequency subbands. Finally, the inverse NSST was employed to reconstruct the image. In the second part, an intensity enhancement method was proposed for energy preservation. The proposed method was verified on 43 groups of images and compared with other eight fusion methods such as Principal Component Analysis (PCA) and Local Laplacian Filtering (LLF). The fusion results on two different categories of medical image fusion (Magnetic Resonance Imaging (MRI) and Positron Emission computed Tomography (PET), MRI and Single-Photon Emission Computed Tomography (SPECT)) show that the proposed method can obtain more competitive performance on both visual quality and objective evaluation indicators including Mutual Information (MI), Spatial Frequency (SF), Q value, Average Gradient (AG), Entropy of Information (EI), and Standard Deviation (SD), and can improve the quality of medical image fusion.

For Mixed-Variable Optimization Problem (MVOP) containing both continuous and categorical variables, a coevolution strategy was proposed to search the mixed-variable decision space, and a Coevolutionary Ant Colony Optimization Algorithm for MVOP (CACOA _MV) was developed. In CACOA _MV, the continuous and categorical sub-populations were generated by using the continuous and discrete Ant Colony Optimization (ACO) strategies respectively, the sub-vectors of continuous and categorical variables were evaluated with the help of cooperators, and the continuous and categorical sub-populations were respectively updated to realize the efficient coevolutionary search in the mixed-variable decision space. Furthermore, the ability of global exploration to the categorical variable solution space was improved by introducing a smoothing mechanism of pheromone, and a "best+random cooperators" restart strategy facing the coevolution framework was proposed to enhance the efficiency of coevolutionary search. By comparing with the Mixed-Variable Ant Colony Optimization (ACO _MV) algorithm and the Success History-based Adaptive Differential Evolution algorithm with linear population size reduction and Ant Colony Optimization (L-SHADE _ACO), it is demonstrated that CACOA _MV is able to perform better local exploitation, so as to improve approximation quality of the final results in the target space; the comparison with the set-based Differential Evolution algorithm with Mixed-Variables (DE _MV) shows that CACOA _MV is able to better approximate the global optimal solutions in the decision space and has better global exploration ability. In conclusion, CACOA _MV with the coevolutionary strategy can keep a balance between global exploration and local exploitation, which results in better optimization ability.

Aiming at the problems of large number of model parameters, large computational cost and low accuracy when using deep learning algorithms for pointer instrument recognition task, an intelligent detection and recognition system of circular pointer instrument based on the combination of improved pre-trained MobileNetV2 network model and circular Hough transform was proposed. Firstly, the Hough transform was used to solve the interference problem of non-circular areas in complex scene. Then, the circular areas were extracted to construct datasets. Finally, the circular pointer instrument recognition was realized by using the improved pre-trained MobileNetV2 network model. The average confusion matrix was used to measure the performance of the proposed model. Experimental results show that, the recognition rate of the proposed system in the recognition task of circular pointer instruments reaches 99.76%. At the same time, the results of comparing the proposed model with other five different network models show that the proposed model and ResNet50 both have the highest accuracy, but compared with ResNet50, the proposed network model has the model parameter number and model computational cost reduced by 90.51% and 92.40% respectively, verifying that the proposed model is helpful for the further deployment and implementation of industrial grade real-time circular pointer instrument detection and recognition in mobile terminals or embedded devices.

Considering the data sharing difficulties and the risk of privacy disclosure in grid cloud server based on blockchain, a Data Sharing model based on Double Consortium Blockchains in smart grid(DSDCB) was proposed. Firstly, the data of electricity was stored under-chain by Inter Planetary File System(IPFS), the IPFS file fingerprints were stored on-chain, and the electricity data was shared to other consortium blockchain based on the multi-signature notary technology. Secondly, with ensuring privacy from leakage, proxy re-encryption and secure multi-party computing were combined to share single-node or multi-node security data. Finally, fully homomorphic encryption algorithm was used to integrate ciphertext data reasonably without decrypting the electricity data. The 51% attack, sybil attack, replay attack and man-in-the-middle attacks were resisted by the single-node cross-chain data sharing model of DSDCB. It was verified that the security and privacy of data were guaranteed by the secure multi-party cross-chain data sharing model of DSDCB when the number of malicious participants was less than k and the number of honest participants was more than 1. The simulation comparison shows that the computational cost of the DSDCB model is lower than those of Proxy Broadcast Re-Encryption(PBRE) and Data Sharing scheme based on Conditional PBRE(CPBRE-DS), and the model is more feasible than the Fully Homomorphic Non-interactive Verifiable Secret Sharing(FHNVSS) scheme.

When the Log-Structured Merge-tree (LSM-tree) in the time series database is under high write load or resource constraints, file compaction not in time will cause a large accumulation of LSM C ₀ layer data, resulting in an increase in the latency of ad hoc queries of recently written data. To address this problem, a two-stage LSM compaction framework was proposed that realizes low-latency query of newly written time series data on the basis of maintaining efficient query for large blocks of data. Firstly, the file compaction process was divided into two stages:quickly merging of a small number of out-of-order files, merging of a large number of small files, then multiple file compaction strategies were provided in each stage, finally the two-stage compaction resource allocation was performed according to the query load of the system. By implementing the test of the traditional LSM compaction strategy and the two-stage LSM compaction framework on the time series database Apache IoTDB, the results showed that compared with the traditional LSM, the two-stage file compaction module was able to greatly reduce the number of ad hoc query reads while improving the flexibility of the strategy, and made the historical data analysis and query performance improved by about 20%. Experimental results show that the two-stage LSM compaction framework can increase the ad hoc query efficiency of recently written data, and can improve the performance of historical data analysis and query as well as the flexibility of compaction strategy.

Concerning the safety and monitoring problems of the blind people during traveling, the design of a networked cane system for blind people based on machine learning algorithms was proposed. Multiple functions were added to the system, such as obstacle avoidance, positioning, alarm and communication. First, infrared obstacle avoidance and ultrasonic ranging obstacle avoidance were designed as the basic functions of the system, which could be used to detect road conditions and obstacles for the daily travel of the blind and provide real-time voice and motor vibration reminders. Second, remote communication function for help was added to the system, which was able to send help text messages and phone calls to specific mobile numbers. In addition, Global Positioning System (GPS) function, accelerometer gyroscope attitude angle calculation function and abnormal attitude alarm function based on K-Nearest Neighbor (KNN) and Dynamic Time Warping (DTW) algorithms were also added, which were able to transfer all kinds of information data to the cloud server storage. Finally, the WeChat mini program was used to replace the native APP as the monitoring operation interface, and functions such as one-click alarm, weather query, blind safety information were provided. Test results show that the proposed system has the attitude recognition success rate reached 86%, and has the accuracy improved by nearly 31% compared to the attitude angle system. The networked cane system for blind people can greatly improve the security of the blind during traveling, so that the blind can ask for help in time when an accident occurs, and achieve the safe monitoring and positioning monitoring of the blind postures.

The MAVLink is a lightweight communication protocol between Unmanned Aerial Vehicle (UAV) and Ground Control Station (GCS). It defines a set of mutual bi-directional messages between UAV and GCS, including UAV states and GCS control commands. However, the MAVLink protocol lacks sufficient security mechanisms, and there are security vulnerabilities that may cause serious threats and hidden dangers. To resolve these problems, a security communication scheme for the UAV system based on the MAVLink protocol was proposed. First, the connection requests were broadcasted by the UAV constantly and alternately; then the public key was sent to the UAV by the GSC, and the DH algorithm was used by both sides to negotiate a shared key, and the AES algorithm was used to encrypt the communication on MAVLink message packages, achieving identity authentication. If the UAV did not receive the public key sent by the GCS within the specified time or a decryption error on MAVLink message package happened, the UAV would actively disconnect and update a new public key to rebroadcast the connection request. In addition, concerning the security problem of the UAV system being maliciously tampered with, the system firmware was self-checked during booting. Finally, based on the formal verification platform UPPAAL, it has been proved that the proposed scheme has the security properties of liveness, connectability and connection uniqueness. Results of the communication process between UAV PX4 1.6.0 and GCS QgroundControl 3.5.0 show that the proposed secure communication scheme of UAV system can prevent malicious eavesdropping, message tampering, man in the middle attack and other malicious attacks in the communication process between UAV and GCS, and solve the security vulnerabilities of MAVLink protocol well with little effect on UAV performance.

Aiming at the problem of neglecting some narrow roads due to the formation constraints in the multi-UAV （Unmanned Aerial Vehicle） cooperative trajectory planning, a Fast Particle Swarm Optimization method based on Adaptive Distributed Model Predictive Control (ADMPC-FPSO) was proposed. In the method, the formation strategy combining leader-follower method and virtual structure method was used to construct adaptive virtual formation guidance points to complete the cooperative formation control task. According to the idea of model predictive control, combined with the distributed control method, the cooperative trajectory planning was transformed into a rolling online optimization problem, and the minimum distance and other performance indicators were used as cost functions. By designing the evaluation function criterion, the variable weight fast particle swarm optimization algorithm was used to solve the problem. The simulation results show that the proposed algorithm can effectively realize the multi-UAV cooperative trajectory planning, can quickly complete the adaptive formation transformation according to the environmental changes, and has lower cost than the traditional formation strategy.

Although the existing image inpainting methods can recover the content of the missing area of the image, there are still some problems, such as structure distortion, texture blurring and content discontinuity, so that the inpainted images cannot meet people’s visual requirements. To solve these problems, an image inpainting method based on dilated convolution was proposed. By introducing the idea of dilated convolution to increase the receptive field, the quality of image inpainting was improved. This method was based on the idea of Generative Adversarial Network (GAN), which was divided into generative network and adversarial network. The generative network included global content inpainting network and local detail inpainting network, and gated convolution was used to realize the dynamical learning of the image features, solving the problem that the traditional convolution neural network method was not able to complete the large irregular missing areas well. Firstly, the global content inpainting network was used to obtain an initial content completion result, and then the local texture details were repaired by the local detail inpainting network. The adversarial network was composed of SN-PatchGAN discriminator, and was used to evaluate the image inpainting effect. Experimental results show that compared with the current image inpainting methods, the proposed method has great improvement in Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity (SSIM) and inception score. Moreover, the method effectively solves the problem of texture blurring in traditional inpainting methods, and meets people’s visual requirements better, verifying the validity and feasibility of the proposed method.

When the areas to be inpainted in the face image are large, there are some visual defects caused by the inpainting of the existing methods, such as unreasonable image semantic understanding and incoherent boundary. To solve this problem, an end-to-end image inpainting model of U-Net structure based on dense blocks was proposed to achieve the inpainting of semantic face of any mask. Firstly, the idea of generative adversarial network was adopted. In the generator, the convolutional layers in U-Net were replaced with dense blocks to capture the semantic information of the missing regions of the image and to make sure the features of the previous layers were reused. Then, the skip connections were adopted to reduce the information loss caused by the down-sampling, so as to extract the semantics of the missing regions. Finally, by introducing the joint loss function combining adversarial loss, content loss and local Total Variation (TV) loss to train the generator, the visual consistency between the inpainted boundary and the surrounding real image was ensured, and Hinge loss was used to train the discriminator. The proposed model was compared with Globally and Locally Consistent image completion(GLC),Deep Fusion(DF) and Gated Convolution(GC) on CelebA-HQ face dataset. Experimental results show that the proposed model can effectively extract the semantic information of face images, and its inpainting results have the boundaries with natural transition and clear local details. Compared with the second-best GC, the proposed model has the Structure SIMilarity index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) increased by 5.68% and 7.87% respectively, while the Frechet Inception Distance (FID) decreased by 7.86% for the central masks; and has the SSIM and PSNR increased by 7.06% and 4.80% respectively while the FID decreased by 6.85% for the random masks.

Aiming at the shortage of traditional matrix completion algorithm in image reconstruction, a completion algorithm based on NonLocal self-similarity and Low Rank Matrix Approximation (NL-LRMA) was proposed. Firstly, the nonlocal similar patches corresponding to the local patches in the image were found through similarity measurement, and the corresponding grayscale matrices were vectorized to construct the nonlocal similar patch matrix. Secondly, aiming at the low-rank property of the obtained similarity matrix, Low-Rank Matrix Approximation (LRMA) was carried out. Finally, the completion results were recombined to achieve the goal of restoring the original image. Reconstruction experiments were performed on grayscale and RGB images. The results show that the average Peak Signal-to-Noise Ratio (PSNR) of NL-LRMA algorithm is 4 dB to 7 dB higher than that of the original LRMA algorithm on a classic dataset; at the same time, NL-LRMA algorithm is better than IRNN (Iteratively Reweighted Nuclear Norm), WNNM (Weighted Nuclear Norm Minimization), LRMA (Low-Rank Matrix Approximation) and other traditional algorithms in the terms of visual effect and PSNR value. In short, NL-LRMA algorithm effectively make up for the shortcomings of traditional algorithms in natural image reconstruction, so as to provide an effective solution for image reconstruction.

The teaching of Chinese architecture history has building structures too complex, is limited to 2D planar teaching and is not easy for students to master and apply, therefore an implementation method of Chinese architecture history teaching system based on mixed reality technology was proposed. The wooden structure system of Baoguo Temple in Ningbo was taken as an example, and the mixed reality device Microsoft HoloLens was used as the teaching platform. Firstly, 3ds Max was applied to the 3D simulation modeling of the wooden structure system of Baoguo Temple based on the collected data, and a building model library was built. Then, the 3D human-computer interface of the virtual teaching system was constructed in unity3D, the key technologies were used including environment understanding and human-computer interaction based on C# scripts, and a Chinese architectural history teaching system using HoloLens was implemented with core functions of building structure recognition and cultural cognition. The results show that the system has good 3D visual effects and natural effective human-computer interaction, which can improve the efficiency of knowledge transfer and the initiative of students.

To solve the problem of low recognition rate of motor imagery ElectroEncephaloGram (EEG) signals, considering that EEG signals contain abundant time-frequency information, a recognition method based on Convolutional Neural Network (CNN) in time-frequency domain was proposed. Firstly, Short-Time Fourier Transform (STFT) was applied to preprocess the relevant frequency bands of EEG signals to construct a two-dimensional time-frequency domain map composed of multiple time-frequency maps of electrodes, which was regarded as the input of the CNN. Secondly, focusing on the time-frequency characteristic of two-dimensional time-frequency domain map, a novel CNN structure was designed by one-dimensional convolution method. Finally, the features extracted by CNN were classified by Support Vector Machine (SVM). Experimental results based on BCI dataset show that the average recognition rate of the proposed method is 86.5%, which is higher than that of traditional motor imagery EEG signal recognition method, and the proposed method has been applied to the intelligent wheelchair, which proves its effectiveness.

Griffin-Lim algorithm is widely used in end-to-end speech synthesis with phase estimation, which always produces obviously artificial speech with low fidelity. Aiming at this problem, a system for end-to-end speech synthesis based on WaveNet network architecture was proposed. Based on Seq2Seq (Sequence-to-Sequence) structure, firstly the input text was converted into a one-hot vector, then, the attention mechanism was introduced to obtain a Mel spectrogram, finally WaveNet network was used to reconstruct phase information to generate time-domain waveform samples from the Mel spectrogram features. Aiming at English and Chinese, the proposed method achieves a Mean Opinion Score (MOS) of 3.31 on LJSpeech-1.0 corpus and 3.02 on THchs-30 corpus, which outperforms the end-to-end systems based on Griffin-Lim algorithm and parametric systems in terms of naturalness.

In order to solve the problem that the traditional cascaded Convolutional Neural Network (CNN) has low accuracy of feature point localization in left ventricular ultrasound image, an improved cascaded CNN with region extracted by Faster Region-based CNN (Faster-RCNN) model was proposed to locate the left ventricular endocardial and epicardial feature points in ultrasound images. Firstly, the traditional cascaded CNN was improved by a structure of two-stage cascaded. In the first stage, an improved convolutional network was used to roughly locate the endocardial and epicardial joint feature points. In the second stage, four improved convolutional networks were used to fine-tune the endocardial feature points and the epicardial feature points separately. After that, the positions of joint contour feature points were output. Secondly, the improved cascaded CNN was merged with target region extraction, which means that the target region containing the left ventricle was extracted by the Faster-RCNN model and then was sent into the improved cascaded CNN. Finally, the left ventricular contour feature points were located from coarse to fine. Experimental results show that compared with the traditional cascaded CNN, the proposed method is much more accurate in left ventricle feature point localization, and its prediction points are closer to the actual values. Under the root mean square error evaluation standard, the accuracy of feature point localization is improved by 32.6 percentage points.

To deal with poor robustness in strong noise environment, weak adaptability to complex mixed noise that appear in the existing unsupervised image segmentation models, an improved noise-robust image segmentation model based on One-class SVM (Support Vector Machine) method was proposed. Firstly, a data outlier detection mechanism was constructed based on One-class SVM. Secondly, an outlier degree was introduced into the energy function, so that more accurate image information could be obtained by the proposed model under multiple noise intensities and the failure of weight-descend mechanism in strong noise environment was avoided. Finally, the segmentation contour was driven to the target edge by minimizing the energy function. In noise image segmentation experiments, the proposed model could obtain ideal segmentation results with different types and intensities of noise. Under F₁-score metric, the proposed model is 0.2 to 0.3 higher than LCK (Local Correntropy-based K-means) model, and has better stability in strong noise environments. The segmentation convergence time of the proposed model is only slightly longer than that of LCK model by about 0.1 s. Experimental results show that the proposed model is more robust to probabilistic, extreme values and mixed noise without significantly increase of segmentation time, and can segment natural images with noise.

By the combination of fuzzy particle swarm optimization algorithm based on receding horizon control and improved artificial potential field, an on-line path planning method for achieving fixed-wing Unmanned Aerial Vehicle (UAV) path planning in uncertain environment was proposed. Firstly, the minimum circumscribed circle fitting was performed on the convex polygonal obstacles. Then, aiming at the static obstacles, the path planning problem was transformed into a series of on-line sub-problems in the time domain window, and the fuzzy particle swarm algorithm was applied to optimize and solve the sub-problems in real time, realizing the static obstacle avoidance. When there were dynamic obstacles in the environment, the improved artificial potential field was used to accomplish the dynamic obstacle avoidance by adjusting the path. In order to meet the dynamic constraints of fixed-wing UAV, a collision detection method for fixed-wing UAV was proposed to judge whether the obstacles were real threat sources or not in advance and reduce the flight cost by decreasing the turning frequency and range. The simulation results show that, the proposed method can effectively improve the planning speed, stability and real-time obstacle avoidance ability of fixed-wing UAV path planning, and it overcomes the shortcoming of easy to falling into local optimum in traditional artificial potential field method.