Search

Road network plays an important role in the fields of navigation, urban planning, and transportation. Extracting road network data from imagery based on machine learning models is an efficient and economical method for obtaining road network data. In order to save labor costs, crowdsourced data can be employed to automatically acquire the labels for model training. In response to the current challenges in road extraction, such as the limited number of labeled samples, low precision of sample labels generated from crowdsourced data, and difficulty in obtaining accurate road label data, which lead to low-quality, incomplete, and inaccurate road extraction, this study proposes a label refining framework based on a road matching and integrate algorithm. Labels are generated from OpenStreetMap (OSM) vector data, and roads are extracted from very high resolution orthoimage using the U-net model. The extracted roads are then matched and integrated with the original data to generate refined labels, which are employed for further model training and road extraction. Experimental results demonstrate that this process can overcome the poor quality of samples directly generated from the OSM data, i.e., the label refining framework led to significant improvements with respect to the completeness, accuracy, and quality of the road network extraction results.

Electric vehicle (EV) charging spots require frequent inspections to identify potential safety hazards, and charging piles in charging spots exhibit a significant failure rate, necessitating maintenance. The efficiency and quality of inspection-maintenance are crucial components in the operation and management of charging spots, directly affecting the operation cost, service level, and user experience. The inspection-maintenance of EV charging spots requires two different types of personnel, and the impact on customer satisfaction cannot be ignored. In addition, the large-scale charging spots have further increased the difficulty of routing. A novel routing optimization problem for large-scale inspection-maintenance of EV charging spots is proposed in this paper, considering multiple types of personnel and composite customer satisfaction. While inspections can be carried out by general personnel, maintenance requires professional engineers. Composite customer satisfaction is with respect to the waiting time for maintenance of faulty charging spots and the duration of absence of professional engineers in the working station. A bi-objective routing optimization model is proposed, minimizing the cost of arranging personnel and traveling between charging spots, and the loss of composite customer satisfaction simultaneously. Due to the large scale of the problem caused by the enormous number of charging spots, an effective multi-stage solution approach is proposed. Finally, numerical experiments, managerial insights, and a case study are provided. The purpose of this research is to help the operators to decide the routes for inspection-maintenance of EV charging spots and offer valuable guidance in balancing operational efficiency with customer satisfaction.

Extracting building construction time is crucial for effective land resource management and sustainable urban development, particularly in fast-growing urban areas. However, acquiring building construction time remains challenging due to limited observations with multiple changes. To address this issue, we applied a monthly time series of remote sensing images and the LandTrendr change detection algorithm to extract building construction times. We identified the sensitive index of short wavelength infra-red (SWIR) from satellite observations for detecting changes in building construction, demolition, and reconstruction. Comparing composite results at different temporal intervals revealed that monthly data is more effective in accurately characterizing building changes compared to daily and yearly intervals. Additionally, our improved algorithm in Google Earth Engine identified the maximum change time as the construction turning point at the pixel level. We then revealed Beijing's construction time from 1990 to 2020 by overlaying building footprint data with extracted year information from Landsat images. Our results achieved an 82.32% agreement on identified construction time of buildings with a two-year tolerance strategy using 560 randomly collected building samples. Our results outperformed traditional methods such as annual time series composition with the same LandTrendr algorithm, historical surveying map, and building change time of social big data monitoring, with derived overall accuracies of 68.75%, 74.64%, and 67.47%, respectively, suggesting the good performance of the adopted approach. This study offered a potential avenue for detailed monitoring of urban building changes at a fine-grained spatial scale, with far-reaching implications for sustainable urban development practices.

Infrared object detection is one of the critical technologies for maritime search and rescue. However, it is still challenging due to the strong background clutter interference and the lack of small object information. We proposed a temporal-spatial pyramid network for infrared maritime object detection. We proposed a nested temporal pyramid to represent the temporal features through motion differences maps and energy accumulation maps to distinguish the wave clutter and objects. We proposed a dense spatial pyramid to learn the spatial features and the differences between temporal maps and then to clarify and locate objects. For training, we designed a scale-related composite loss function with correlated location description and weighted confidence loss. Finally, based on the ablation and comparison experiments, the proposed method performs better on maritime infrared sequences.

The device performance of indium zinc oxide (IZO) thin-film transistors (TFTs) is optimized through process optimizations. By jointly adjusting the annealing condition, the channel thickness and the sputtering atmosphere, the roughness and oxygen vacancies (Vos) are precisely regulated. The optimized IZO TFTs can achieve the highest field effect mobility of ~71.8 cm2/Vs with a threshold voltage of ~-0.6 V. Reliability of IZO TFTs under positive/negative bias stress is also examined. The interface quality and the Vo are two key factors influencing the device performance and reliability, confirmed by X-ray photoelectron spectroscopy and atomic force microscopy analysis.

The information sharing among internet-connected vehicles is close, but there may be the issue of vehicles providing false information. This paper proposes a new incentive mechanism scheme for VANET to provide a decentralized and trusted service management. We introduce a trust-based message evaluation system, incorporating reputation assessment and negative feedback functions to evaluate the credibility of messages and assessments. A reward and punishment mechanism based on repeated games is designed to incentivize vehicles to proactively share information, thereby preventing false information and defamation attacks. Additionally, a blockchain-based distributed trust management scheme is designed to ensure the consistency of trust management storage through the consensus mechanism of the blockchain, reducing reliance on centralized institutions. Simulation results demonstrate that the establishment of reputation and incentive mechanisms among vehicles enhances the enthusiasm for information sharing and improves the accuracy of identifying false messages.

Blood is vital to sustain human life, to ensure that hospitals or blood centers have sufficient blood reserves, this paper is based on Long- and Short-Term Temporal Patterns with Deep Neural Networks (LSTNET) and Attention Mechanism, uses Gaussian Process Bayesian Optimization (GPBO) to optimize the model, and propose GPBO-LST-Attention network model. Based on the clinical blood transfusion data of a medical institution in Harbin in the past 10 years, the clinical blood transfusion volume was modeled and the blood usage in the future was predicted. The model proposed in this paper was subjected to ablation experiments and compared with other models, using the Root Mean Square Error (RMSE), the Mean Absolute Error (MAE) and the Mean Absolute Percentage Error (MAPE) analyses to evaluate the model performance, and Akaike Information Criterion corrected (AICc) was used to analyze the model performance. It is shown that when the attention module and the GPBO algorithm are used in the GPBO-LST-Attention model proposed in this paper, the RMSE is reduced by 0.789, the MAE by 0.592, and the MAPE by 0.49. The prediction error values of error indexes and the AICc value are lower than the other models. It shows that the model established in this paper can accurately predict blood consumption in a future time and has excellent model performance, which can provide data reference for the blood reserve of hospitals or blood centers.

In low signal-to-noise ratio (SNR) conditions, to address the problem of poor speech enhancement effect of traditional neural networks, this paper combines Convolution Recurrent Neural Network (CRN) with Gated Linear Units (GLU) to extract speech characteristics. Meanwhile, considering the advantages of the Empirical Mode Decomposition (EMD) algorithm, this paper put forward a model for speech enhancement that is based on Adaptive Mean Median Empirical Modal Decomposition and Multilayer Gated Convolutional Recurrent Neural Networks (ME-MGCRN). The model first uses the improved EMD algorithm to split the original noisy speech into low-frequency and high-frequency components. After that, the noisy speech is denoised using noise correlation and completes the feature extraction process. Finally, the extracted feature information is fed into the MGCRN network to realize speech enhancement further. This paper analyzes the performance of the Adaptive Mean Median Empirical Modal Decomposition algorithm using the LibriSpeech ASR dataset. This paper compares and investigates the enhanced performance of the ME-MGCRN model with the baseline model and the traditional model in terms of evaluation metrics such as the perceptual evaluation of speech quality (PESQ) and short-time objective intelligence (STOI). The research indicates that the ME-MGCRN model proposed in this paper has improved in the evaluation indexes, such as PESQ and STOI, respectively, by 0.22% and 5.6% than the baseline model and the traditional model, and the speech enhancement effect is better when Huber is selected as the loss function.

In the financial field, texts such as news and commentaries, as carriers of public opinion, have the function of reflecting investor sentiment, influencing investment decisions and market trends, and extracting positive or negative sentiment from them in a timely manner is very important to the investment decisions of fintech companies and investors. However, financial sentiment analysis is challenging due to issues such as unclear sentiment polarity of financial texts, high context-dependency, and highly specialised and specific expressions of the linguistic features of financial texts. In order to overcome these challenges, we design a hybrid topic feature and pre-trained model financial text sentiment analysis model, and we design the method of improved attention mechanism to optimise the model iteratively, the improved attention mechanism reduces the noise caused by the improper allocation of attention to a single word through adaptive attention threshold, attention weight masking mechanism, so as to capture more contextual information and avoid the loss of information, thus improving the stability of the model. This improves the stability of the model. The introduction of thematic features enables the model to capture potential semantic structures and long-distance word associations in the text to enhance the understanding of text context. The experimental results show that the method has an improvement of 2.05%-7.27% in F1 value compared with the baseline method, which is suitable for financial text sentiment analysis.

The primary cause of high carbon emissions due to inefficient transportation is the lack of cooperation among various logistics companies during distribution. Joint distribution mode represents an effective means to address this issue. However, research on the joint distribution mode primarily focuses on the centralized joint distribution mode. In the context of the decentralized joint distribution mode, logistics companies can autonomously choose their distribution routes based on the orders they are dispatched. If logistics companies unilaterally aim to reduce their own costs, it can easily lead to an increase in carbon emissions from the alliance, which can have adverse effects. This paper focuses on the decentralized joint distribution alliance and considers reducing carbon emissions through the allocation of the alliance’s order. A two-stage approach is proposed, based on routing optimization and worst-case estimation. The mathematical models for each stage are established and algorithms are designed. Subsequently, management recommendations are provided through numerical analysis. This paper aims to explore methods to address the challenges faced by decentralized joint distribution alliances.

As a crucial maritime search and rescue method, infrared object detection is critical in influencing the success rate. Research on infrared maritime images is limited, and the problems of smaller object sizes, more substantial noise, and less detailed information still need to be solved. To tackle these problems, we proposed an infrared maritime object detection network with feature enhancement and adjacent fusion. A spatial feature enhancement module and a semantic feature enhancement module are designed to enhance the location information of dim small targets and the deep semantic information, respectively. We designed a feature adjacent fusion network to fully use multiscale feature information. We built a maritime infrared dataset and compared the proposed method with existing advanced traditional and learning methods. The proposed method achieves better detection results.

In this paper, an integrated device which realized the dual-mode integration of power amplifier (PA) and radio frequency (RF) switch based on GaN dual-gate (DG) structure is designed and fabricated. The integrated device provides two working modes and meets the performance requirements of PA and RF switch. In the transmit (Tx) mode, the integrated device is used as PA. At the frequency of 3.6 GHz, compared with the conventional GaN HEMT (high electron mobility transistor) as PA, due to the field modulation effect of the dual-gate structure, the current collapse is effectively suppressed, and the output power density $(P_{out})$ of the integrated device is increased from 6.90 W/mm to 7.85 W/mm, and the power added efficiency (PAE) is increased from 44.3% to 51.1%. In the receive (Rx) mode, the integrated device is used as RF switch. Compared with conventional RF switch, the isolation of integrated devices is significantly improved. At the frequency of 3.6GHz, the isolation is improved from 19.4 dB to 41.0 dB; At the frequency of 40 GHz, the isolation is improved from 3.0 dB to 29.7 dB. The integrated device provides a novel function integration scheme for RF frontends, with great potential for improving the integration and dimension of future communication systems.

In modern underwater warfare, detecting and locating enemy submarines is the primary task, but there are still many difficulties in detecting submarines. The acoustic signals may distort in the ocean channel due to the complexity of the ocean, which affects the detection of sonar. That is to say, the impact of environmental factors needs to be considered in submarine detection. Therefore, this article proposes an underwater collaborative detection method based on complex marine environments, constructs a sonar underwater detection capability analysis system to improve detection performance, analyzes the impact of marine environmental factors on sonar detection capability, evaluates sonar performance using sonar quality factor, calculates sonar detection distance, establishes polar coordinates, and explores sonar detection capability; Besides, the concept of collaborative detection is also introduced, and a collaborative detection probability distribution surface is proposed to fuse the detection capabilities of different platforms, presenting the collaborative detection capabilities of multiple platforms in the form of probability contour lines. The detection performance of different deployment methods is explored by changing the placement position and number of platforms. At last, the research has shown that the underwater detection capability of sonar is closely related to the marine environment. The square layout in underwater collaborative detection has better detection performance. The research results of this article can provide references for the layout of platforms in underwater detection tasks.

The effective disposal of healthcare waste is a highly concerned issue. Healthcare waste collection poses transport risks and temporary storing risks, and the deliberation undertaken by decision-makers concerning waste collection entails potential scenario wherein some hospitals have not yet submitted their service requests and may submit their requests during the healthcare waste collection procedure. A routing optimisation problem for healthcare waste collection with temporary storing risks and sequential uncertain service requests is introduced. A two-stage decision-making is proposed and mathematical models corresponding to each stage are developed. Different solution algorithms are developed for different stages or different scales of instances, including the improved $\varepsilon $ -constraint method and Non-dominated Sorting Genetic Algorithm-II for the solution procedure in Stage 1, and the Compare-choose-move algorithm for the solution procedure in Stage 2. Finally, the models and algorithms are tested by numerical instances and several suggestions for healthcare waste collection have been proposed based on sensitivity analysis.

Spectral information and backscatter information are both exclusively important bases for land cover classification, and these two kinds of information are found in multispectral images and SAR images, respectively. Therefore, the fusion of complementary information of multispectral and SAR images can effectively improve land cover classification accuracy. However, the existing fusion methods of multispectral and SAR images generally have some problems, such as insensitivity to edge information, serious interference by speckle noise, and unreasonable settings of fusion rules, which lead to unsatisfactory results of land cover classification. To solve this issue, a fusion method based on adaptive multiscale Gaussian co-occurrence filtering decomposition is proposed. Gaussian filtering and adaptive co-occurrence filtering are applied to the original image to smooth out speckle noise and interference edges within the textures while preserving edge information between textures. Through multi-scale spatial decomposition, the separation of detail information, edge information and basic information is realized, while multi-layer fusion of image features is performed. Finally, the fused image with low noise interference, clear boundary and uniform pixel convergence is generated. Experimental results show that the proposed method generally performs the best in eight evaluation indexes compared with ten other methods. The overall accuracy, average accuracy and Kappa coefficient of land cover classification are increased by 7.674%, 6.776%, and 0.098, respectively, compared with those of the original multispectral image in Area 1, and by 6.904%, 7.649%, and 0.089, respectively, compared with those of the original multispectral image in Area 2.

In silicon carbide (SiC) planar insulated-gate bipolar transistor (IGBT), a large distance between neighboring p-bodies is beneficial to enhance the on-state conductivity modulation, but will expose the gate oxide to high electric field in off-state. With p-bodies placed closer, the gate oxide field is reduced, but the conductivity modulation is suppressed. In this work, a new SiC planar IGBT with oxide shield is proposed and studied by TCAD simulations. The proposed SiC IGBT achieves improved trade-off between on-state voltage drop $(V_{\mathrm{ ON}})$ and maximum gate oxide electric field $(E_{\text {ox-m}})$ . When a quite larger distance between neighboring p-bodies is adopted in the proposed SiC IGBT, a low $V_{\mathrm{ ON}}$ is obtained, while the $E_{\text {ox-m}}$ can be kept at a small value with the oxide shielding structures protecting the gate oxide. Switching characteristics are also studied, and the proposed SiC-IGBT delivers much better trade-off between turn-off energy loss $(E_{\mathrm{ OFF}})$ and $V_{\mathrm{ ON}}$ than the conventional SiC planar IGBT.

In this work, high performance InAlN/GaN HEMT based on the n+GaN regrown ohmic contact with n+GaN contact ledge structure is proposed. The regrown ohmic contact of InAlN/GaN HEMT is formed by MBE n+GaN regrowth and self-stopping etching, which makes the total ohmic contact resistance between the 2DEG channel and the ohmic metal decrease to $0.12 \Omega \cdot $ mm and forms n+GaN contact ledge structure. Owing to the n+GaN contact ledge on the InAlN barrier, with the increasing of drain-source voltage (VDS), an additional current path comes into being between the n+GaN contact ledge on the InAlN barrier and the 2DEG channel, which can “shorten” the device effective drain-source distance, thus further reducing the parasitic resistance. Compared with regrown InAlN/GaN HEMT without n+GaN contact ledge structure, the peak transconductance (Gm.max) of regrown InAlN/GaN HEMT with n+GaN contact ledge structure increases from 747 mS/mm to 874 mS/mm, and the saturation current density (ID.max) increases from 2.6 A/mm to 2.9 A/mm. Besides, the self-stopping etching on the access region does not induce extra defects, and negligible current collapse is obtained. As the results of low parasitic resistance, high output current density, low knee voltage and negligible current collapse, power-added-efficiency (PAE) of 44% together with output power density (Pout) of 2.5 W/mm is achieved at 30 GHz and $V_{\mathrm{ DS}}$ of 10 V, which indicates regrown InAlN/GaN HEMT with n+GaN contact ledge structure has great potential for millimeter-wave low voltage applications. Additionally, the transfer and schottky gate characteristics show negligible degradation after OFF/ON-state electrical stress tests.

A solar cell defect detection method with an improved YOLO v5 algorithm is proposed for the characteristics of the complex solar cell image background, variable defect morphology, and large-scale differences. First, the deformable convolution is incorporated into the CSP module to achieve an adaptive learning scale and perceptual field size; then, the feature extraction capability of the model is enhanced by introducing the ECA-Net attention mechanism; finally, the model network structure is improved and one tiny defect prediction head is added to improve the accuracy of target detection at different scales. To further optimize and improve the YOLO v5 algorithm, this paper uses Mosaic and MixUp fusion data enhancement, K-means++ clustering anchor box algorithm, and CIOU loss function to enhance the model performance. The experimental results show that the improved YOLO v5 algorithm achieves 89.64% mAP for the model trained on the solar cell EL image dataset, which is 7.85% higher than the mAP of the original algorithm, and the speed reaches 36.24 FPS, which can complete the solar cell defect detection task more accurately while meeting the real-time requirements.

In this paper, the impact of ohmic structure on channel-to-channel (C2C) coupling effect in InAlN/GaN double channel (DC) HEMTs is systematically analyzed and studied. For the un-recessed ohmic structure, the electrons in the upper channel can easily inject into the bottom channel due to the ultra-thin InAlN back barrier layer. Therefore, the maximum drain current and transconductance peak of the bottom channel significantly increase. For recessed ohmic structure, the reduced vertical electric field strength of the upper channel can effectively weaken the coupling effect between the two GaN channels. Benefiting from the suppressed vertical transport of electrons in the upper channel, higher drain current and transconductance of the upper channel are obtained in the recessed ohmic structure. In addition, the transmission electron microscope (TEM) microstructural analysis of the DC HEMTs with recessed ohmic structure was also performed. This work shows that the recessed ohmic structure can modulate the electron transport mode in the InAlN/GaN DC HEMTs. The coupling effect of the two channels will play a major role in influencing the characteristics of current gain cutoff frequency ( $\,f_{\mathrm{ T}}$ ) / maximum power gain cutoff frequency ( $\,f_{\mathrm{ max}}$ ) versus $V_{\mathrm{ GS}}$ and have a significant effect on large-signal characteristics, which is quite attractive for the fabrication of power microwave GaN-based HEMTs with wide gate swing.

The robust bounded control problem for the permanent magnet linear motors with inequality constraints is studied. Firstly, the dynamical model of the system is built, through the state transformation is used to satisfy the inequality constraint of the position output. Thus, the controller after the state transformation can ensure that the control output of the linear motor stays the desired range. Selecting the appropriate boundary function, and then the upper and lower bounds of the control output can be set according to our control requirements. The control scheme can guarantee uniform boundedness and uniform ultimate boundedness. The results of experiments and simulation show that the proposed algorithm can assure the control output within the desired range regardless of the uncertainty.

The difficulties in video-based face recognition, such as dramatic pose variations and low quality, can be alleviated by leveraging the rich complementary information between the frames. However, limited by the mini-batch training strategy, the current deep learning methods only utilizes the frames in each batch during training, which ignore the context of the entire video. In this paper, we propose a context-aware feature aggregation scheme to aggregate complementary information between different frames. Firstly, a two-branch structure is designed as the Context-aware feature Aggregation Network (CAN). Secondly, a context-aware training strategy using a context bank is proposed, which alleviates the limitation of mini-batch samples by using the context of the entire video or several images belonging to the same ID and thus achieves global contribution estimation result. Comparative studies on benchmark datasets, such as IJB-C, YouTube Face (YTF), PaSC and COX, confirm that the proposed approach can achieve state-of-the-art level. Meanwhile, qualitative analysis on Multi-PIE dataset indicates that the contribution learned by the CAN is reasonable and beneficial to video face recognition.

Long-tailed distribution generally exists in large-scale face datasets, which poses challenges for learning discriminative feature in face recognition. Although a few works conduct preliminary research on this problem, the value of the tail data is still underestimated. This paper addresses the long-tailed problem from the perspective of maximally exploiting the tail data. We propose a Joint Alternating Training (JAT) framework to learn discriminative feature from both the long-tailed data and the tail data by using alternating training strategy. JAT consists of two branches: 1) the long-tailed data branch is adopted to learn the universal discrimination information from the whole long-tailed data with instance-balanced sampling. 2) the tail data branch is designed to exploit the discriminative information in the tail data with class-balanced sampling. To compensate the insufficient samples and lack of intra-class variations, we apply data augmentation (DA) to the tail data. We further propose margin-based mixup (MarginMix) for data augmentation, which can deal with the nonlinearity of margin-based softmax loss and stabilize the training process in mixup. Furthermore, we obtain the best combination of strategies (i.e., JAT+DA+ MarginMix) for long-tailed face recognition, which can maximally exploit the discriminative information in the tail data while retaining the universal discrimination learned from the long-tailed data. Extensive experiments on 8 face datasets demonstrate that our proposed methods and combination of strategies can effectively address the long-tailed problem in face recognition.

Transportation in controlled industrial sites provides a conducive environment for technologies of Connected Automated Vehicles (CAV). Recent studies show that safe and efficient road sharing between CAVs and pedestrians is challenging. Besides safety issues, a significant loss of time occurs when pedestrians cross a stream of CAVs. Currently, many techniques have been employed to improve the coordination between CAVs and pedestrians. They focus on pedestrian detection, display of the intention of CAVs, and cooperative collision avoidance. However, one of the most significant sources of information that the pedestrian uses for her/his decision-making is the speed profile of CAVs. This paper aims to provide a safe and efficient pedestrian crossing at industrial sites through communicative crossing behavior. To this end, a suitable speed profile of the CAV is designed by assuming that pedestrians and CAV play a cooperative game to move as close as possible to their desired speed. First, a system analysis is proposed to derive the optimal decision and trajectory for each agent. Then, Deep Reinforcement Learning (DRL) is used to control the longitudinal speed of the CAVs. Compared with Model Predictive Control approach, DRL allows coping with unforeseeable pedestrian behaviors (e.g. long reaction time, varying ideal speed, stop in the middle, etc.). Simulations and experiments with real human testers based on immersive hamlet are performed. Results show that the proposed speed profile outperforms significantly the collision avoidance approach.

Recently, remote sensing (RS) technology are becoming an increasingly important technology in estimating forest growing stem volume (GSV), and the saturation issue of spectral variables from various optical sensors severely hinders the improvement of mapping forest GSV, especially in the planted forest with high GSV. Forest canopy height is widely considered as one of the major factors to increase the saturation levels in mapping GSV. However, it is rather difficult to invert the forest canopy height without precisely external DEM for large regions. In this study, the canopy height model (CHM) was derived from ZY-3 stereo images with subtracting open-sourced external DEM and the response of saturation levels was analyzed by adding forest height in the planted coniferous forest (Larch and Chinese pine). To further describe the relationships between the forest height and saturation levels, five datasets with five estimation models (Linear, MLR, SVR, KNN, and RF) and three methods of variable selection (Stepwise, LASSO, and Pearson) were applied to estimate the forest GSV using corrected CHM and 49 alternative variables extracted from ZY-3 multispectral images. Meanwhile, a spherical model was employed to quantitatively describe the saturation levels of combined variables. The results showed that values of relative root means square error were decreased from 29.3% to 25% for Larch and from 26.5% to 22.2% for Chinese pine after adding the corrected CHM, respectively. Meanwhile, the saturation level of each combined variable set was successfully determined by the spherical model. The results illustrated that the saturation levels of GSV were significantly increased by adding corrected CHM from open-sourced external DEM. Specially, the averaged saturation levels were increased from 220 m3/ha to nearly 300 m3/ha for Chinese pine and from 150 m3/ha to 220 m3/ha for Larch, respectively. It is proved that ZY-3 stereo and multispectral images have great potential for accurate estimation of forest GSV by delaying the saturation levels using extracted CHM with open-sourced external DEM.

Face masks bring a new challenge to face recognition systems especially against the background of the COVID-19 pandemic. In this paper, a method used for mitigating the negative effects of mask defects on face recognition is proposed. Firstly, a low-cost, accurate method of masked face synthesis, i.e. mask transfer, is proposed for data augmentation. Secondly, an attention-aware masked face recognition (AMaskNet) is proposed to improve the performance of masked face recognition, which includes two modules: a feature extractor and a contribution estimator. Therein, the contribution estimator is employed to learn the contribution of the feature elements, thus achieving refined feature representation by simple matrix multiplications. Meanwhile, the end-to-end training strategy is utilized to optimize the entire model. Finally, a mask-aware similarity matching strategy (MS) is adopted to improve the performance in the inference stage. The experiments show that the proposed method consistently outperforms on three masked face recognition datasets: RMFRD, COX and Public-IvS. Meanwhile, a qualitative analysis experiments using CAM indicate that the contribution learned by AMaskNet is more conducive to masked face recognition.

To fully learn the latent temporal dependencies from post-disturbance system dynamic trajectories, deep learning is utilized for short-term voltage stability (STVS) assessment of power systems in this paper. First of all, a semi-supervised cluster algorithm is performed to obtain class labels of STVS instances due to the unavailability of reliable quantitative criteria. Secondly, a long short-term memory (LSTM) based assessment model is built through learning the time dependencies from the post-disturbance system dynamics. Finally, the trained assessment model is employed to determine the systems stability status in real time. The test results on the IEEE 39-bus system suggest that the proposed approach manages to assess the stability status of the system accurately and timely. Furthermore, the superiority of the proposed method over traditional shallow learning-based assessment methods has also been proved.

Task deployment has become a research hotspot for load balancing in joint “cloud-edge” datacenter. In view of the problem that most of the hosts are overloaded in the current joint “cloud-edge” datacenter, which may cause unbalanced load in the center, existing research mainly pay attention to the problem of unilateral load balancing of cloud computing center or edge computing center. In order to realize efficient deployment of “cloud-edge” tasks and overall load balancing, on the basis of the deployment mode of joint “cloud-edge”, this paper proposes a resource management and task deployment strategy JCETD (Joint Cloud-Edge Task Deployment) based on pruning algorithm and deep reinforcement learning. The main idea consists of two parts: firstly, the set of “cloud-edge” hosts is pruned according to the attribute value of the physical host. Then, there will be a non-dominated set of joint hosts which reduces the computational complexity of the whole algorithm and improve the computational efficiency of the system. Secondly, the problem of task deployment is simulated as a deep reinforcement learning process under the “cloud-edge” model. Through the continuous exploration and utilization of the system environment, the tasks are reasonably and efficiently deployed in the cloud computing center and edge computing center. Finally, the “cloud-edge” system can achieve an efficient computing performance and overall load balancing. The experimental results show that the proposed algorithm significantly reduces the total completion time and average response time compared with the existing research, which effectively optimizes the service ability and realizes the load balancing of the joint “cloud-edge” system.

A discrete GaN power transistor’s substrate is typically connected to its source electrode. However, on the GaN-on-Si power IC platform, the high-side transistor (HS-transistor) and low-side transistor (LS-transistor) share a common substrate that cannot be simultaneously connected to both source electrodes of the two transistors. Thus, the termination of the common substrate remains an undecided issue. In this work, comprehensive TCAD simulations are exploited to reveal the influences of various substrate termination schemes. It is found the common substrate inevitably leads to severe degradation in the dynamic ${R} _{\mathrm{ ON}}$ due to back-gating effects. The mechanisms for the degradations vary with the substrate termination scheme, and will be discussed in detail. To address these issues, we propose a new GaN power IC platform on an engineered bulk silicon substrate, and study the new platform with TCAD simulations. The proposed platform provides a local electrical substrate (a p+ island) for each GaN power transistor. The source electrode of each GaN transistor is connected to its local electrical substrate, while all devices share a common mechanical substrate. The junctions between the local substrates and the underlying n-layer provide an effective isolation between GaN transistors. The back-gating effects are completely suppressed for the GaN integrated bridge circuit.

As one road network is often the backbone in a geo-spatial dataset, capturing and/or updating the road networks using remote sensing imagery play an important role in traffic management, urban planning, vehicle navigation, and emergency management. Along with the progress of remote sensing launching technologies and the successful applications of deep learning in the field of computer vision, it has become more and more efficient and economical to employ deep learning methods for road-features extractions from very high-resolution (VHR) remote sensing imagery. Meanwhile, as one of the most significant and popular volunteer geographic information data sources, the OpenStreetMap (OSM) including the complete road networks in the wide world has been accumulated in the past decades. In this article, a generic and automatic approach for extracting road networks from VHR remote sensing images has been proposed based on fully convolutional neural network, in which the road centerlines from OSM have been employed to construct the labels for the model training and validation. In the conducted experiments on various VHR image datasets with two different spatial resolutions of 0.3 and 1 m, the proposed model has demonstrated quite satisfactory results - the overall completeness and correctness of the roads extraction from VHR remote sensing images exceed 94.0% and 98.0%, respectively.

In this work, low voltage RF power capability on AlGaN/GaN and InAlN/GaN HEMTs is analyzed from the perspective of DC and pulse characteristics, for terminal applications whose operating voltage is usually in the range of 3 to 15 V. Device fabrication is performed on mature AlGaN/GaN heterojunction as well as strongly polarized InAlN/GaN heterojunction, to make a comparison of low voltage RF power capability between two devices. Although it suffers from relatively severe RF dispersion, InAlN/GaN HEMT delivers higher output power density (Pout) than its opponent, benefiting from the lower parasitic resistance and knee voltage as well as higher output current density. At 8 GHz, Pout of 1.62 W/mm and 1.10 W/mm are achieved for InAlN/GaN and AlGaN/GaN HEMT, respectively, both of which are biased at class AB operation and Vds of 9 V. However, the tremendous degradation of power added efficiency (PAE) occurs as the higher drain voltage is applied on InAlN/GaN HEMT, as the result of the severe gate leakage. What is more, a higher PAE is more necessary than Pout for terminal applications. Either InAlN/GaN HEMT or AlGaN/GaN HEMT has its own specific voltage range to deliver higher PAE. Concretely, InAlN/GaN HEMT is more suitable to applications with operating voltage not exceeding 6 V, and AlGaN/GaN HEMT is preferred for ones with relatively higher voltage, accompanied by the decent PAE as high as 62% to 66% and moderate Pout to meet the demand of various low voltage terminal applications.

As ground-level O3 has become one of the major pollutants affecting air quality in recent years, monitoring and revealing the spatial distribution pattern of O3 are of great significance to study air-pollution characteristics. Based on the multilayer backpropagation neural network, one O3 estimation model is proposed to obtain the continuous spatial distribution of O3 concentrations, where the Landsat 8 images, meteorological parameters, and air quality data have been integrated together as the input for the model training and validation. In order to enhance the estimation accuracy, the proposed model has been optimized with respect to the influencing factors and spatial extent. In the test areas of Beijing, Tianjin, and Shijiazhuang of China, the proposed O3 estimation model has demonstrated quite satisfactory performance—with the average coefficient of determination (R2) larger than 0.90 and root-mean-square error smaller than 19.0 μg/m3. It is worth mentioning that all the data employed in this research are freely available and can be applied nationwide in the mainland of China. Taking advantage of the generic nature and the positive O3 estimation results with high accuracy and spatial resolution, the proposed model can be expected to be a new way for studying air-pollution characteristics and, thus, support the decision making for environmental governance.

Diabetic Retinopathy (DR) may lead to blindness in diabetic patients, which is one of the most severe eye diseases. Therefore, using automatical technology to detect DR at the early phase has very vital clinical significance. In order to detect the microaneurysms (MAs) and hard exudates (HEs) of DR, a novel detection method based on deep symmetric convolutional neural network is proposed in this paper. The symmetric convolutional structure is used to improve the effectiveness of feature extraction. The proposed method also can overcome the imbalance of positive and negative samples to avoid overfitting by increasing the width and depth of the network. Furthermore, different network structures (convolution, pooling) are used to achieve different feature filtering in the stage of feature extractions. According to the experimental results, the proposed method is superior to the state-of-the-art approach on the public dataset DIARETDB1 (DB1). The detection accuracy of the objects is 92.0%, 93.2%, 93.6%, when using different filtering structures (convolution, max-pooling, ave-pooling) respectively. The detection of microaneurysms is much improved by using ave-pooling layer for feature filtering, and the max-pooling layer can improve the detection of hard exudates.

The inverse piezoelectric effect and trap effect in GaN HEMTs under the high electric field by reverse-bias step stress at cryogenic temperature (CT, 77K) have been investigated. It is found that the inverse piezoelectric effect is suppressed while the trap effect is enhanced at CT. Due to the lower tensile stress in the as-grown AlGaN barrier at CT, the amplitude of the critical voltage related to inverse piezoelectric effect increases from 75V at 300K to 100V at CT, mitigating the irreversible degradation of the devices. The inverse piezoelectric effect dominates the degradation of reverse gate leakage, which is weakened at CT. However, the devices suffer from more degradation of Id (decreases by 94.38%) and Gm,max (decreases by 95.15%) at CT in the dark, induced by the serious trap effect due to the longer emission time. UV-light-assisted stress measurements have been utilized to identify the contribution of the two mechanisms to the degradation of device characteristics at CT. The degradation of Id(Gm, max) is only 0.82% (3.31%) for the piezoelectric effect and 93.56% (91.84%) for trap effect at CT, respectively.

As a consequence of expedited urban construction, large quantities of wastes are shipped every day. Because of the excessive pursuit of economic benefits by muck truck drivers and unreasonable transportation route choices, the accident rates of these drivers continue to increase. It is very important to optimize the safe transportation routing of muck trucks. In this paper, a safe transportation routing model was established to minimizes path safety risks and single driving time as the bi-objective, and two different situations for the optimal path of the model were analyzed. For the situation in which there are several short paths in the network, one of which is the path with the least safety risks, an exact algorithm, A*, was designed, and time complexity was analyzed. For the situation in which there is a single path or several shortest paths in the network but the path with the minimal path safety risks is not included in the set of shortest paths, an approximation algorithm, GA*, was designed, and time complexity and the approximation ratio were analyzed. Our experimental results demonstrated the validity of our developed model and approaches. Moreover, a case study is carried out on a real-case application related to the network in the Xi'an International Port District.

For the purpose of addressing the multi-objective optimal reactive power dispatch (MORPD) problem, a two-step approach is proposed in this paper. First of all, to ensure the economy and security of the power system, the MORPD model aiming to minimize active power loss and voltage deviation is formulated. And then the two-step approach integrating decision-making into optimization is proposed to solve the model. Specifically speaking, the first step aims to seek the Pareto optimal solutions (POSs) with good distribution by using a multi-objective optimization (MOO) algorithm named classification and Pareto domination based multi-objective evolutionary algorithm (CPSMOEA). Furthermore, the reference Pareto-optimal front is generated to validate the Pareto front obtained using CPSMOEA; in the second step, integrated decision-making by combining fuzzy c-means algorithm (FCM) with grey relation projection method (GRP) aims to extract the best compromise solutions which reflect the preferences of decision-makers from the POSs. Based on the test results on the IEEE 30-bus and IEEE 118-bus test systems, it is demonstrated that the proposed approach not only manages to address the MORPD issue but also outperforms other commonly-used MOO algorithms including multi-objective particle swarm optimization (MOPSO), preference-inspired coevolutionary algorithm (PICEAg) and the third evolution step of generalized differential evolution (GDE3).

In order to make Thermostatically Controlled Loads (TCLs) better meet the scheduling requirements, a day-ahead scheduling of equivalent energy storage model that takes into account of the minimum-on-off time is established. By considering the minimum-on-off time, the charging and discharging power, as well as the energy storage are modified, and the relationship between heat exchange power and energy storage are developed. By this way, the equivalent energy storage model more accurately reflects the real thermodynamic characteristic of TCLs and enables TCLs to exert the actual potential to participate in the scheduling. Finally, the simulation results verify the feasibility of the proposed scheduling method.

In this paper, an in-fiber composite modal interferometer (CMI) is proposed and experimentally demonstrated based on cascaded no-cladding fiber (NCF) and thin-core fiber (TCF) structure, which respectively corresponds to multimode interferometer (MMI) and Mach-Zehnder interferometer (MZI). The self-image effect of MMI is deeply analyzed and the interference fringes distribution of CMI is accurately designed by the selective the length of NCF and TCF. The comprehensive tests are then conducted and the experimental results show that interference fringes present similar temperature response but the apparent differences in strain and refractive index (RI) tests owing to the asymmetrical structure. MMI has a nearzero intensity sensitivity of strain and red-shifts with the increased RI. But MZI has the sensitivity of -19.3nm/RIU and a 4.94 dB increment in the range of 300-1100 με. Thus high-discrimination simultaneous triple-parameter measurement is achieved by means of joint wavelength and intensity demodulation. The proposed CMI has the merits of high sensitivity, low cost and ease of fabrication, and is very promising and potential for the engineering monitoring and sensing applications.

In this article, a novel feedforward control method is proposed to control the hysteretic nonlinearity and resonance in piezoelectric-driven mechanism. A third-order rate-dependent Rayleigh model is established according to voltage dependence and rate dependence tests using sinusoidal and triangular waveform signals. In order to verify the accuracy of this model, the tracking errors of the Rayleigh model are analyzed and a comparison of the 3D Rayleigh model and the experimental data is visualized. The modeling accuracy of Rayleigh model in minor loops is also analyzed quantitatively. The hysteresis compensation Rayleigh model is then derived based on the energy compensation method. To control the mechanical resonance in piezoelectric-driven mechanism, a triangular input signal trajectory optimization method is developed based on minimum-acceleration trajectory planning theory. The turning parts of the triangular waveform signal are replaced with smooth curves but the linear parts are retained. Experiments are conducted to demonstrate the effectiveness of the proposed control method.

To detect the shape of a small magnetic target in the shallow underground layer, this article proposes a recognition method based on Mask-RCNN. Firstly, using COMSOL software and MATLAB software to establish the database of magnetic targets model under different shapes and orientations, which greatly enriched the diversity of the training data set. Then, the ${G}_{\mathrm {zz}}$ component of the magnetic gradient tensor matrix is selected to highlight the shape features of the magnetic target, and the contour image is generated. The experimental data set is created by using the deep learning annotation tool Labelme. Finally, Resnet101 is used as the backbone network and feature pyramid network (FPN) structure is used to extract features. The regional recommendation network (RPN) is trained end-to-end to create regional recommendations for each feature map. The detection results of 200 test images show that the average detection accuracy of the method is 97%, and the recall rate is 94%. The simulation results show that the recognition accuracy and robustness of the method are improved.

Traditional non-negative matrix factorization methods cannot learn the subspace from the high-dimensional data space composed of binary codes. One hopes to discover a compact parts-based representation composed of binary codes, which can uncover the intrinsic information and simultaneously respect the geometric structure of the original data. For this purpose, we introduce discrete hashing methods and propose a novel non-negative matrix factorization to generate binary codes from the original data. In this paper, we construct an affinity graph to encode the geometrical structure of the original data, and the learned binary code subspace achieved by matrix factorization respects the structure. The proposed problem can be formulated as a mixed integer optimization problem. Therefore, we transform it into several sub-problems including an integer optimization problem, two convex problems with the non-negative constraint and a quadratic programming problem. Optimizing each sub-problem alternately until we achieve a local optimal solution. Image clustering and retrieval on image datasets show the excellent performance of our method in comparison to other dimensional reduction methods.

The influential factors of green logistics development have been an important research topic for academics. However, little attention has been devoted to studying the issue of the prioritization and the hierarchical structure of these factors. The purpose of this study is to identify the key factors influencing the development of green logistics, and to explore the prioritization and the hierarchical structure of the influential factors. An indicator system for influential factors of green logistics development is first developed based on the theory of triple bottom line, including sixteen indicators or factors. The comprehensive influences and the prioritization of these sixteen factors are then analyzed using the DEMATEL method and the multi-level hierarchical structure of these factors is constructed by applying the ISM method. The results indicate that government policies, public supports, laws and regulations, green education, green technology innovations, consumer green demands, quantity and quality of green logistics talents are more influential than the other factors. Social factors including laws and regulations and government policies, public supports and green education are the underlying factors affecting the development of green logistics. Economic factors are at the middle levels and environmental factors are at the surface-level of the hierarchical structure. Recommendations are finally given to governmental agencies and business firms. These findings will be beneficial to promoting the sustainable development of green logistics.

This paper presents a modulation constellation design technique for polarization modulation communication in wireless communication. Such constellation can be used for confidential polarization modulation communication on the depolarization channel. Information theory is innovatively applied to constellation design, which allows us to take the channel depolarization effects into account during the constellation design process. Specifically, we derive the security capacity of the communication system under depolarized channel conditions and use it as the objective function for secure constellation optimization. We use a heuristic genetic algorithm to solve the above optimization problem. The numerical results show that compared to the quasi-optimal geometry-based constellation, the resulting constellation can have an advantage of secure capacity and have better secure performance.

It is a challenging problem to detect the apple in natural environment using traditional object recognition algorithms due to occlusion, fluctuating illumination and complex backgrounds. Deep learning methods for object detection make impressive progress, which can automatically extract the number, pixel position, size and other features of apples from the images. In this paper, four deep learning recognition models, Faster RCNN based on AlexNet, Faster RCNN based on ResNet101, YOLOv3 based on DarkNet53 and improved YOLOv3 were employed to carry out recognition experiments on red and green apple under three illumination and two image sharpness conditions, with the transfer learning to accelerate the training process. The results showed that improved YOLOv3 model had the best recognition effect among the four detection models. F1 value of red apple recognition was 95.0%, 94.6% and 94.1% for normal, insufficient and excessive illumination, respectively, and F1 value of green apple recognition was 94.9%, 94.0% and 91.1%. There were F1 value of 92.8% and 92.1% for red and green apple recognition in blurred images, respectively. Moreover, improved YOLOv3 algorithm still had the better performance for occlusion, spot, overlap and incomplete apples, with a recognition recall rate higher than 88.5%. It can be concluded that improved YOLOv3 algorithm can provide a more efficient way for apple detection in natural environment.

Recently, convolutional neural networks (CNNs) showed excellent performance in many tasks, such as computer vision and remote sensing semantic segmentation. Especially, the ability to learn high-representation features of CNN draws much attention. And random forest (RF) algorithm, on the other hand, is widely applied for variables selection, classification, and regression. Based on the previous fusion models that fused CNN with the other models, such as conditional random fields (CRFs), support vector machine (SVM), and RF, this article tested a method based on the fusion of an RF classifier and the CNN for a very high resolution remote sensing (VHRRS) based forests mapping. The study area is located in the south of China and the main purpose was to precisely distinguish Lei bamboo forests from the other subtropical forests. The main novelties of this article are as follows. First, a test was conducted to confirm if a fusion of CNN and RF make an improvement in the VHRRS information extraction. Second, based on RF, variables with high importance were selected. Then, a test was again conducted to confirm if the learning from the selected variables will further give better results.

Accurate and efficient identification of the urban vegetation abundance is of great importance for urban planning and management. A lot of efforts have been made to estimate the urban fractional vegetation cover (FVC) using multispectral images by the pixel-based mixture analysis method. However, urban FVC maps comprising various meaningful landscapes have wider applications. Compared with other moderate spatial resolution multispectral imagery (e.g., SPOT, Landsat 8), the Sentinel-2 multispectral instrument (MSI) imagery has higher resolution, larger coverage, and shorter revisit time. So it may provide higher accuracy for urban FVC mapping. This article derives an accurate object-based urban FVC map for Changsha city, China, from the 10-m resolution Sentinel-2 data acquired in 2017. For producing the urban FVC maps, the mixture analysis methods were applied on segmental image objects instead of pixels. The results demonstrate that the object-based mixture analysis method achieved a higher FVC estimation accuracy than the pixel-based mixture analysis did, and it effectively removed the “salt and pepper” phenomena. The object-based linear model fully constrained least squares and achieved the best estimation accuracy (R2 = 0.92, RMSE = 0.0956). The red-edge band reflectance information of the MSI imagery can improve the accuracy of the FVC maps, but not significantly. The object-based urban FVC maps would be a good alternative to the traditional pixel-based maps.

Paddy rice is a major stable food, accounting for about 20% world's food supply. And the rice paddy, an important artificial wetland type, plays an important role in the regional ecological environment. This study proposes a stacked generalization and spectral mixture approach to map paddy rice using coarse spatial resolution images [Moderate Resolution Imaging Spectralradiometer, (MODIS)]. By this method, the time series MODIS enhanced vegetation index images, phenological variables, land surface water index, elevation, and slope images are all employed to produce the optimal feature combination, which is then used to map paddy rice by the stacking algorithm. The validation experiment using the data of the Dongting Lake area showed that the proposed method can improve the overall accuracy of single classifiers, including the support vector machine, random forest, k-nearestneigbor (kNN), extreme gradient boosting (XGB), and decision tree. Stacking (XGB) achieves the highest overall accuracy (90.3%) and Kappa coefficient (0.86), which are 2.8% and 0.03 higher than that of using the single kNN classifier. Furthermore, its user accuracies for distinguishing double-cropping rice and single-season rice are 92.5% and 90.0%, respectively. In terms of the paddy rice classification accuracy, the stacking model is also superior to single classifiers. Moreover, the MODIS-derived rice map obtained by the stacked generalization approach and the spectral mixture method area has a large determination coefficient (R2 = 0.9975) with the government statistic data. The results demonstrate the potential of the proposed method in using coarse spatial resolution images for large-scale paddy rice mapping.

NAND flash memory is widely used in various computing systems. However, flash blocks can sustain only a limited number of program/erase (P/E) cycles, which are referred to as the endurance. On one hand, in order to ensure data integrity, flash manufacturers often define the maximum P/E cycles of the worst block as the endurance of flash blocks. On the other hand, blocks exhibit large endurance variations, which introduce two serious problems. The first problem is that the error correcting code (ECC) is often over-provisioned, as it has to be designed to tolerate the worst case to ensure data integrity, which causes longer decoding latency. The second problem is the underutilized block's lifespan due to conservatively defined block endurance. Raw bit error rate (RBER) of most blocks have not arrived the allowable RBER based on the nominal endurance point, which implies that the conventional P/E cycle-based block retirement policies may waste large flash storage space. In this paper, to exploit the storage capacity of each flash block, we propose an RBER-aware lifetime prediction scheme based on machine learning technologies. We consider the problem that the model can lose prediction effectiveness over time and use incremental learning to update the model for adapting the changes at different lifetime stages. At run time, trained data will be gradually discarded, which can reduce memory overhead. For evaluating our purpose, four well-known machine learning techniques have been compared in terms of predictive accuracy and time overhead under our proposed lifetime prediction scheme. We also compared the predicted values with the tested values obtained in the real NAND flash-based test platform, and the experimental results show that the support vector machine (SVM) models based on our proposed lifetime prediction scheme can achieve as high as 95% accuracy for flash blocks. We also apply our proposed lifetime prediction scheme to predict the actual endurance of flash blocks at four different retention times, and the experimental results show that it can significantly improve the maximum P/E cycle of flash blocks from 37.5% to 86.3% on average. Therefore, the proposed lifetime prediction scheme can provide a guide for block endurance prediction.

In order to revamp the cleaning contract from the head-count basis into a performance basis, a fair and unbiased cleanliness classification is necessary. However, the perception of cleanliness is very subjective to the observer. Hence, it is not an easy task to quantify the cleanliness. This paper presents an application of principal component analysis (PCA) in conjunction with convolutional neural networks (CNN) to identify the cleanliness of a restroom up to three levels; namely, dirty, average, and clean. The proposed method includes an application specific data augmentation algorithm and a PCA-based feature analysis schema to select the best suited CNN model for our dataset. Since this study focused on a specific application, we benchmark the performances of the proposed method performances with the state-of-the-art computer vision algorithms on our dataset. Moreover, our study shows a machine learning approach toward automating the inspection process of a restroom.

In this paper, a data-driven adaptive optimal control strategy is proposed for a class of linear systems with structured time-varying uncertainty, minimizing the upper bound of a pre-defined cost function while maintaining the closed-loop stability. An off-policy data-driven reinforcement learning algorithm is presented, which uses repeatedly the online state signal on some fixed time intervals without knowing system information, yielding a guaranteed cost control (GCC) law with quadratic stability for the system. This law is further optimized through a particle swarm optimization (PSO) method, the parameters of which are adaptively adjusted by a fuzzy logic mechanism, and an optimal GCC law with the minimum upper bound of the cost function is finally obtained. The effectiveness of this strategy is verified on the dynamic model of a two-degree-of-freedom helicopter, showing that both stability and convergence of the closed-loop system are guaranteed and that the cost is minimized with much less iteration than the conventional PSO method with constant parameters.

Outsourcing image search services to public clouds is an ever-increasing trend. However, directly outsourcing image datasets to untrusted clouds introduces privacy concerns. Several secure image retrieval schemes have been proposed recently. However, most of them require participation from image owners when building secure indexes, which wastes many computational resources of the image owners. Several schemes are proposed to solve this problem, but they suffer from low search accuracy on large datasets. In this paper, we propose the first secure image retrieval scheme that simultaneously solves these two problems. To obtain higher search accuracy, we extract image features via fine-tuned convolutional neural networks. Then, the image features are encrypted by using the secure k-Nearest Neighbor algorithm. To improve search speed and reduce the cost of image owners, we let cloud servers locally build a secure hierarchical index graph by using the encrypted image features. Besides, the secure index can be built and updated in parallel. We provide security analysis for the proposed scheme. Performance evaluations on the CIFAR-10 dataset show that the proposed scheme is practical. Moreover, compared with a recent scheme, our scheme can save more index construction time and cost of image owners when building secure indexes.