Memory Occupation Research Articles

RaNAS: Resource-Aware Neural Architecture Search for Edge Computing

Neural architecture search (NAS) for edge devices is often time-consuming because of long-latency deploying and testing on edge devices. The ability to accurately predict the computation cost and memory requirement for convolutional neural networks (CNNs) in advance holds substantial value. Existing work primarily relies on analytical models, which can result in high prediction errors. This paper proposes a resource-aware NAS (RaNAS) model based on various features. Additionally, a new graph neural network is introduced to predict inference latency and maximum memory requirements for CNNs on edge devices. Experimental results show that, within the error bound of ±1%, RaNAS achieves an accuracy improvement of approximately 8% for inference latency prediction and about 25% for maximum memory occupancy prediction over the state-of-the-art approaches.

Simulations of X-ray focusing by zone plates in rotationally symmetric optical field utilizing the matrix-free Finite Difference Beam Propagation Method

We present the use of a finite difference method based on Crank-Nicholson scheme and recurrence scheme for computationally efficient simulation of the X-ray propagation through a zone plate. By introducing boundary and central conditions and by avoiding large matrix operations, the method achieves considerable speed, little memory occupation and low background noise. Accommodating refractive index profiles of arbitrary shape, it can be applied to assist optimizing X-ray zone plates and understanding focusing mechanism.

LSTM Gate Disclosure as an Embedded AI Methodology for Wearable Fall-Detection Sensors

In this paper, the concept of symmetry is used to design the efficient inference of a fall-detection algorithm for elderly people on embedded processors—that is, there is a symmetric relation between the model’s structure and the memory footprint on the embedded processor. Artificial intelligence (AI) and, more particularly, Long Short-Term Memory (LSTM) neural networks are commonly used in the detection of falls in the elderly population based on acceleration measures. Nevertheless, embedded systems that may be utilized on wearable or wireless sensor networks have a hurdle due to the customarily massive dimensions of those networks. Because of this, the algorithms’ most popular implementation relies on edge or cloud computing, which raises privacy concerns and presents challenges since a lot of data need to be sent via a communication channel. The current work proposes a memory occupancy model for LSTM-type networks to pave the way to more efficient embedded implementations. Also, it offers a sensitivity analysis of the network hyper-parameters through a grid search procedure to refine the LSTM topology network under scrutiny. Lastly, it proposes a new methodology that acts over the quantization granularity for the embedded AI implementation on wearable devices. The extensive simulation results demonstrate the effectiveness and feasibility of the proposed methodology. For the embedded implementation of the LSTM for the fall-detection problem on a wearable platform, one can see that an STM8L low-power processor could support a 40-hidden-cell LSTM network with an accuracy of 96.52%.

An efficient cross-view image fusion method based on selected state space and hashing for promoting urban perception

In the field of cross-view image geolocation, traditional convolutional neural network (CNN)-based learning models generate unsatisfactory fusion performance due to their inability to model global correlations. The Transformer-based fusion methods can well compensate for the above problems, however, the Transformer has quadratic computational complexity and huge GPU memory consumption. The recent Mamba model based on the selection state space has a strong ability to model long sequences, lower GPU memory occupancy, and fewer GFLOPs. It is thus attractive and worth studying to apply Mamba to the cross-view image geolocation task. In addition, in the image-matching process (i.e., fusion of satellite/aerial and street view data.), we found that the storage occupancy of similarity measures based on floating-point features is high. Efficiently converting floating-point features into hash codes is a possible solution. In this study, we propose a cross-view image geolocation method (S6HG) based purely on Vision Mamba and hashing. S6HG fully utilizes the advantages of Vision Mamba in global information modeling and explicit location information encoding and the low storage occupancy of hash codes. Our method consists of two stages. In the first stage, we use a Siamese network based purely on vision Mamba to embed features for street view images and satellite images respectively. Our first-stage model is called S6G. In the second stage, we construct a cross-view autoencoder to further refine and compress the embedded features, and then simply map the refined features to hash codes. Comprehensive experiments show that S6G has achieved superior results on the CVACT dataset and comparable results to the most advanced methods on the CVUSA dataset. It is worth noting that other floating-point feature-based methods (4096-dimension) are 170.59 times faster than S6HG (768-bit) in storing 90,618 retrieval gallery data. Furthermore, the inference efficiency of S6G is higher than ViT-based computational methods.

A method to shorten the time of high-dimensional matrix operations and convolution operations

Serverless computing, or Function-as-a-Service (FaaS), is a cloud computing execution model where the cloud provider dynamically manages the allocation and provisioning of servers which is a new way to solve the problem that the need of client.By serverless computing clients can easily solve their problems anytime,anywhere.However, the complex convolution operations and high-dimensional matrix operations can still consume a lot of resources such as computation time and memory occupancy in serverless computing. In this paper, firstly, the author selects two frequently used models in serverless computing, and then divide them into different numbers of partitions respectively to study the impact of partitioning methods on the computation time and memory occupancy. The author show that different partitioning methods can make an obviously impact on the computation time and memory occupancy,whitch shows that the proper partitioning of high-dimensional matrices and tensors can make a significantly reduction of calculation time and memory occupancy.

Data-driven offline reinforcement learning approach for quadrotor’s motion and path planning

Non-learning based motion and path planning of an Unmanned Aerial Vehicle (UAV) is faced with low computation efficiency, mapping memory occupation and local optimization problems. This article investigates the challenge of quadrotor control using offline reinforcement learning. By establishing a data-driven learning paradigm that operates without real-environment interaction, the proposed workflow offers a safer approach than traditional reinforcement learning, making it particularly suited for UAV control in industrial scenarios. The introduced algorithm evaluates dataset uncertainty and employs a pessimistic estimation to foster offline deep reinforcement learning. Experiments highlight the algorithm’s superiority over traditional online reinforcement learning methods, especially when learning from offline datasets. Furthermore, the article emphasizes the importance of a more general behavior policy. In evaluations, the trained policy demonstrated versatility by adeptly navigating diverse obstacles, underscoring its real-world applicability.

MaP: Increasing node capacity of programmable cloud gateways

Virtual routing table lookup is a crucial operation in multi-tenant cloud gateways, the essential devices to enable the elastic management of Virtual Private Clouds (VPCs). However, software-based cloud gateways face performance bottlenecks when handling ever-increasing traffic. To address this challenge, a recent trend involves accelerating virtual routing table lookups using programmable switches, with Sailfish being the state-of-the-art solution. Nonetheless, Sailfish’s table scale on a single programmable switch is limited by the ASIC memory sizes. We introduce a novel scheme, named MaP (Merge and Partition), to significantly increase the memory efficiency on a single Tofino 1.0 programmable switch. MaP exploits the similarities in routing table entries across various VPCs for table compression and leverages the hardware architecture of programmable chips to minimize the memory occupancy. Performance evaluations using real-world virtual routing tables demonstrate that our approach reduces the memory cost from 25% to 47.7% compared to Sailfish, without negatively impacting the throughput and latency. With MaP, a single programmable switch can manage a massive table containing over 2 million entries for nearly 1 million VPCs, while only occupying 59.0% of memory and 63.0% of stages.

Machine Learning Algorithms for Processing and Classifying Unsegmented Phonocardiographic Signals: An Efficient Edge Computing Solution Suitable for Wearable Devices.

The phonocardiogram (PCG) can be used as an affordable way to monitor heart conditions. This study proposes the training and testing of several classifiers based on SVMs (support vector machines), k-NN (k-Nearest Neighbor), and NNs (neural networks) to perform binary ("Normal"/"Pathologic") and multiclass ("Normal", "CAD" (coronary artery disease), "MVP" (mitral valve prolapse), and "Benign" (benign murmurs)) classification of PCG signals, without heart sound segmentation algorithms. Two datasets of 482 and 826 PCG signals from the Physionet/CinC 2016 dataset are used to train the binary and multiclass classifiers, respectively. Each PCG signal is pre-processed, with spike removal, denoising, filtering, and normalization; afterward, it is divided into 5 s frames with a 1 s shift. Subsequently, a feature set is extracted from each frame to train and test the binary and multiclass classifiers. Concerning the binary classification, the trained classifiers yielded accuracies ranging from 92.4 to 98.7% on the test set, with memory occupations from 92.7 kB to 11.1 MB. Regarding the multiclass classification, the trained classifiers achieved accuracies spanning from 95.3 to 98.6% on the test set, occupying a memory portion from 233 kB to 14.1 MB. The NNs trained and tested in this work offer the best trade-off between performance and memory occupation, whereas the trained k-NN models obtained the best performance at the cost of large memory occupation (up to 14.1 MB). The classifiers' performance slightly depends on the signal quality, since a denoising step is performed during pre-processing. To this end, the signal-to-noise ratio (SNR) was acquired before and after the denoising, indicating an improvement between 15 and 30 dB. The trained and tested models occupy relatively little memory, enabling their implementation in resource-limited systems.

Detection research of insulating gloves wearing status based on improved YOLOv8s algorithm

The safety hazards may be caused by power grid operators not wearing insulating gloves according to regulations for live electrical working. Additionally, existing methods for detecting the wearing status of insulating gloves suffer from low recognition accuracy, slow detection speed, and large memory occupation by weight files. To address these issues, a Mixup-CA-Small-YOLOv8s (MCS-YOLOv8s) algorithm is proposed for detecting the wearing status of insulating gloves. Firstly, the mixup data augmentation technology using image mixing is introduced, increasing the data’s diversity and improving the model’s generalization ability. Secondly, the coordinate attention (CA) module is added to the original backbone network to strengthen the channel and positional information, suppressing the secondary feature information. Finally, a small target detection structure is designed by removing the last bottom feature detection layer in the original neck network and adding a shallow feature. The ability of small targets’ feature extraction is enhanced without increasing too much computation. The experimental results indicate that the mean average precision of the MCS-YOLOv8s algorithm on the test set is 0.912, the detection speed is 87 FPS, and the model’s weight memory occupies 15.7 MB. It is verified that the model has the advantages of high detection accuracy, fast speed, and small weight memory, which has great significance in ensuring the safe and stable operation of the power grid.

An online method for ship trajectory compression using AIS data

Abstract Vessel trajectories from the Automatic Identification System (AIS) play an important role in maritime traffic management, but a drawback is the huge amount of memory occupation which thus results in a low speed of data acquisition in maritime applications due to a large number of scattered data. This paper proposes a novel online vessel trajectory compression method based on the Improved Open Window (IOPW) algorithm. The proposed method compresses vessel trajectory instantly according to vessel coordinates along with a timestamp driven by the AIS data. In particular, we adopt the weighted Euclidean distance (WED), fusing the perpendicular Euclidean distance (PED) and synchronous Euclidean distance (SED) in IOPW to improve the robustness. The realistic AIS-based vessel trajectories are used to illustrate the proposed model by comparing it with five traditional trajectory compression methods. The experimental results reveal that the proposed method could effectively maintain the important trajectory features and significantly reduce the rate of distance loss during the online compression of vessel trajectories.

GA-GGD: Improving semantic discriminability in graph contrastive learning via Generative Adversarial Network

Graph contrastive learning has garnered considerable research interest due to its ability to effectively embed graph data without manual labels. Among them, methods based on Deep Graph Infomax (DGI) have been widely studied and favored in the industry because of their fast training speed, applicability to large-scale data. DGI-based methods usually obtain a noise graph through node shuffling. The proxy task of these methods encourages the encoder to distinguish whether the nodes come from the original graph or the noise graph, thereby maximizing the mutual information between the node representation and the graph it belongs to, while also maximizing the Jenson–Shannon divergence between the nodes of original graph and noise graph. However, we argue that these approaches only enable the encoder to differentiate between semantically meaningful graphs and noise graphs, but not to effectively identify different semantic graphs. This leads to the inability of the encoder to effectively embed information between different semantics, significantly reducing the robustness and affecting the performance of downstream tasks. In addition, this training mode makes the model more sensitive to attacks. To improve their semantic discriminability, we take advantage of the natural ability of generative adversarial networks to generate semantic data, proposing a method called Generative Adversarial Graph Group Discrimination (GA-GGD). Specifically, it consists of a graph group discriminator and a semantic attack generator. The discriminator aims to encode the graph and identify whether nodes originate from the original graph. The goal of the generator is to use random features and graph structure to find vulnerabilities of discriminator and generate node representations with similar but wrong semantic to confuse the discriminator. GA-GGD can improve the model’s semantic information embedding without significantly increasing computational overhead and memory occupancy. We test the effectiveness of the proposed model on commonly used data sets and large-scale datasets, as well as in various downstream tasks such as classification, clustering, and adversarial attacks defence. A wealth of experimental results confirm the efficacy of the proposed model.

PQSDC: a parallel lossless compressor for quality scores data via sequences partition and run-length prediction mapping.

The quality scores data (QSD) account for 70% in compressed FastQ files obtained from the short and long reads sequencing technologies. Designing effective compressors for QSD that counterbalance compression ratio, time cost, and memory consumption is essential in scenarios such as large-scale genomics data sharing and long-term data backup. This study presents a novel parallel lossless QSD-dedicated compression algorithm named PQSDC, which fulfills the above requirements well. PQSDC is based on two core components: a parallel sequences-partition model designed to reduce peak memory consumption and time cost during compression and decompression processes, as well as a parallel four-level run-length prediction mapping model to enhance compression ratio. Besides, the PQSDC algorithm is also designed to be highly concurrent using multicore CPU clusters. We evaluate PQSDC and four state-of-the-art compression algorithms on 27 real-world datasets, including 61.857 billion QSD characters and 632.908 million QSD sequences. (1) For short reads, compared to baselines, the maximum improvement of PQSDC reaches 7.06% in average compression ratio, and 8.01% in weighted average compression ratio. During compression and decompression, the maximum total time savings of PQSDC are 79.96% and 84.56%, respectively; the maximum average memory savings are 68.34% and 77.63%, respectively. (2) For long reads, the maximum improvement of PQSDC reaches 12.51% and 13.42% in average and weighted average compression ratio, respectively. The maximum total time savings during compression and decompression are 53.51% and 72.53%, respectively; the maximum average memory savings are 19.44% and 17.42%, respectively. (3) Furthermore, PQSDC ranks second in compression robustness among the tested algorithms, indicating that it is less affected by the probability distribution of the QSD collections. Overall, our work provides a promising solution for QSD parallel compression, which balances storage cost, time consumption, and memory occupation primely. The proposed PQSDC compressor can be downloaded from https://github.com/fahaihi/PQSDC.

Detection of Gannan Navel Orange Ripeness in Natural Environment Based on YOLOv5-NMM

In order to achieve fast and accurate detection of Gannan navel orange fruits with different ripeness levels in a natural environment under all-weather scenarios and then to realise automated harvesting of Gannan navel oranges, this paper proposes a YOLOv5-NMM (YOLOv5 with Navel orange Measure Model) object detection model based on the improvement in the original YOLOv5 model. Based on the changes in the phenotypic characteristics of navel oranges and the Chinese national standard GB/T 21488-2008, the maturity of Gannan navel oranges is tested. And it addresses and improves the problems of occlusion, dense distribution, small target size, rainy days, and light changes in the detection of navel orange fruits. Firstly, a new detection head of 160 × 160 feature maps is constructed in the detection layer to improve the multi-scale target detection layer of YOLOv5 and to increase the detection accuracy of the different maturity levels of Gannan navel oranges of small sizes. Secondly, a convolutional block attention module is incorporated in its backbone layer to capture the correlations between features in different dimensions to improve the perceptual ability of the model. Then, the weighted bidirectional feature pyramid network structure is integrated into the Neck layer to improve the fusion efficiency of the network on the feature maps and reduce the amount of computation. Lastly, in order to reduce the loss of the target of the Gannan Navel Orange due to occlusion and overlapping, the detection frame is used to remove redundancy using the Soft-NMS algorithm to remove redundant candidate frames. The results show that the accuracy rate, recall rate, and average accuracy of the improved YOLOv5-NMM model are 93.2%, 89.6%, and 94.2%, respectively, and the number of parameters is only 7.2 M. Compared with the mainstream network models, such as Faster R-CNN, YOLOv3, the original model of YOLOv5, and YOLOv7-tiny, it is superior in terms of the accuracy rate, recall rate, and average accuracy mean, and also performs well in terms of the detection rate and memory occupation. This study shows that the YOLOv5-NMM model can effectively identify and detect the ripeness of Gannan navel oranges in natural environments, which provides an effective exploration of the automated harvesting of Gannan navel orange fruits.

FT_HTlist: A fault-tolerant frequent itemset mining algorithm based on the linear table

This paper proposes a fault-tolerant frequent itemset mining algorithm (FT_HTlist) based on the linear table when the fault-tolerance is 1. The algorithm uses the method of concatenating 1 in the highest bit of the binary number of the known fault-tolerant frequent patterns to generate the candidate fault_tolerant patterns, called FT_Candidate. The algorithm is based on the data structure of the linear table for fault-tolerant frequent itemset mining. This method does not need recursion, so it reduces the consumption of mining space. At the same time, the paper proposed a deduplication algorithm to remove the support for repeat calculations. So the algorithm has a strong advantage in spatial performance. In addition, the algorithm only needs to mine two horizontal chains of the FT_Candidate, thus reducing the consumption of mining time. Finally, the paper shows the time performance and space performance of the proposed algorithm under sparse datasets and dense datasets. The results show that our algorithm has better mining time than other algorithms, and the horizontal chain reduces the memory occupation of the algorithm.

Crack Detection of Concrete Based on Improved CenterNet Model

Cracks on concrete surfaces are vital factors affecting construction safety. Accurate and efficient crack detection can prevent safety-related accidents. Using drones to photograph cracks on a concrete surface and detect them through computer vision technology has the advantages of accurate target recognition, simple practical operation, and low cost. To solve this problem, an improved CenterNet concrete crack-detection model is proposed. Firstly, a channel-space attention mechanism is added to the original model to enhance the ability of the convolution neural network to pay attention to the image. Secondly, a feature selection module is introduced to scale the feature map in the downsampling stage to a uniform size and combine it in the channel dimension. In the upsampling stage, the feature selection module adaptively selects the combined features and fuses them with the output features of the upsampling. Finally, the target size loss is optimized from a Smooth L1 Loss to IoU Loss to lessen its inability to adapt to targets of different sizes. The experimental results show that the improved CenterNet model reduces the FPS by 123.7 Hz, increases the GPU memory by 62 MB, increases the FLOPs by 3.81 times per second, and increases the AP by 15.4% compared with the original model. The GPU memory occupancy remained stable during the training process and exhibited good real-time performance and robustness.

Research on Handwriting Text Generation Algorithm Based on Generative Adversarial Network

The main task of this paper is to study the handwriting text generation method based on deep learning. Through understanding the development status of the research, It can be found that the current research on the generation of different handwriting styles still has some obvious defects, such as the need for manual intervention in character segmentation, failure to capture the global handwriting style, style collapse, failure to generate arbitrary length characters, and text. Finally, this paper proposes a handwritten text generation algorithm combining advantages of convolutional network and Transformer. Specifically, this paper first constructs a lightweight backbone network and uses lightweight MobileNetv3 network as the backbone network to realize feature extraction of input images. The Efficient Channel Attention module is introduced to replace the SE attention module of MobileNetv3, which makes the network pay more attention to the global and local style features of handwritten images. In the feature extraction part, the network reduces the number of parameters, calculation amount and video memory occupation. It can also extract rich feature information. The FID and KID indexes of this algorithm obtained 20.28 and 9.07×10-3 respectively, and the generation effect of handwritten pictures was excellent, which could effectively imitate the writing style of writers.

Efficient and Accurate Analog Voltage Measurement Using a Direct Sensor-to-Digital Port Interface for Microcontrollers and Field-Programmable Gate Arrays

Portable sensor systems are usually based on microcontrollers and/or Field-Programmable Gate Arrays (FPGAs) that are interfaced with sensors by means of an Analog-to-Digital converter (ADC), either integrated in the computing device or external. An alternative solution is based on the direct connection of the sensors to the digital input port of the microcontroller or FPGA. This solution is particularly interesting in the case of devices not integrating an internal ADC or featuring a small number of ADC channels. In this paper, a technique is presented to directly interface sensors with analog voltage output to the digital input port of a microcontroller or FPGA. The proposed method requires only a few passive components and is based on the measurements of the duty cycle of a digital square-wave signal. This technique was investigated by means of circuit simulations using LTSpice and was implemented in a commercial low-cost FPGA device (Gowin GW1NR-9). The duty cycle of the square-wave signal features a good linear correlation with the analog voltage to be measured. Thus, a look-up table to map the analog voltage values to the measured duty cycle is not required with benefits in terms of memory occupation. The experimental results on the FPGA device have shown that the analog voltage can be measured with a maximum accuracy of 1.09 mV and a sampling rate of 9.75 Hz. The sampling rate can be increased to 31.35 Hz and 128.31 Hz with an accuracy of 1.61 mV and 2.68 mV, respectively.

LiWi-HAR: Lightweight WiFi-Based Human Activity Recognition Using Distributed AIoT

Human activity recognition (HAR) based on WiFi channel state information (CSI) has received a lot of attentions recently due to its non-intrusive nature. Most CSI-based HAR systems use a WiFi router and a computing terminal for centralized processing, which makes it difficult to achieve real-time wide-range recognition. Recently, lightweight artificial intelligence internet of things (AIoT) devices are widely deployed. The equipped WiFi chips within such devices can collect and process CSI data in a distributed way. Thus, the AIoT devices extend the detection range of collecting CSI and enrich the applications. However, the memories of the AIoT devices are constrained and lack of appropriate lightweight CSI processing strategies. To address these challenges, we propose the LiWi-HAR system which employs a comprehensive lightweight CSI processing strategy in WiFi based AIoT devices. The proposed lightweight CSI processing strategy extracts the main related features while compressing the data size. Then, a double hidden layer BP neural network based on particle swarm optimization (PSO-BPNN) algorithm is developed for HAR. In this case, the computing memory occupation of the device is effectively reduced, and the real-time high accurate recognition is achieved. Extensive experimental results present that the efficiency of our system significantly outperforms other centralized deep learning based systems and the recognition accuracy achieves 91.7%.

Hand Gesture Authentication by Discovering Fine-Grained Spatiotemporal Identity Characteristics

Dynamic hand gesture is an emerging and promising biometric trait containing both physiological and behavioral characteristics. Possessing the two kinds of characteristics makes dynamic hand gesture have more identity information enabling more accurate and secure authentication theoretically, but also poses a challenge of efficient fine-grained spatiotemporal feature extraction. This challenge involves a seemingly paradoxical problem that high-frame-rate videos are required for behavioral characteristic analysis, but they can also introduce high computational costs. To mitigate this issue, we propose a Frequency Spatiotemporal Attention Network (FSTA-Net) with a focus on satisfying the high-performance and low-computation requirements of authentication systems. The FSTA-Net is established with a two-stage identity characteristic analysis paradigm for short- and long-term modeling. Specifically, considering that models prefer to analyze physiological characteristics which are relatively straightforward to understand, we first design a Behavior Enhanced (BE) module to emphasize hand motions and reduce redundant information to facilitate local identity feature distillation in the first stage. We then present a Frequency Spatiotemporal Attention (FSTA) module to summarize global identity features with decent FLOPs and GPU memory occupation in the second stage. Incorporating the BE and FSTA modules enables them to complement each other’s strengths, resulting in a clear-cut improvement in equal error rate and running speed. Extensive experiments on the SCUT-DHGA dataset demonstrate the superiority of the FSTA-Net. The code is available at https://github.com/SCUT-BIP-Lab/FSTA-Net.

Research on Project-based Teaching Reform of Nautical English in Higher Vocational Colleges under the Background of Informatization

Abstract The quality of higher vocational nautical English teaching is not only related to the student’s vocational ability and job competence but even related to the safety of the ship and personal safety during the voyage. In this paper, we design a reform plan for higher vocational nautical English teaching based on the Hadoop-MapReduce distributed system framework, design the reading and writing algorithms of teaching resources, use a genetic algorithm to improve Pick-KX dynamic load balancing algorithm and construct a nautical English project-based teaching resource base. The effectiveness of the improved Pick-KX dynamic load balancing algorithm in this paper is tested through experiments with 200 to 1000 concurrent users. Finally, two senior nautical English classes are selected to conduct a semester of controlled variable experiments on project-based teaching reform and analyze the changes in performance. It was found that when the number of online users increased from 200 to 1,000, the IO and memory occupancy increased by 60% and 458M, respectively, and the average time taken increased from 0.8 to 5.8 seconds, which was in a reasonable range, and the repository was able to run smoothly. After one semester of experiment, the average score of the experimental class increased by 10.88 points, and the distribution of scores changed towards the high end of the range, while the scores of the control class did not change significantly. This paper’s nautical English project-based teaching resource library is effective, and the project-based teaching reform is a significant improvement in students’ nautical English scores. This study provides unique ideas and feasible methods for the reform of nautical English teaching, as well as the application and innovation of project-based teaching.