The speed of packet forwarding is crucial for the overall transmission performance of network devices. A large number of studies have proposed an acceleration scheme for network traffic processing, which requires not only efficient routing lookup algorithms, but also high-performance architecture and forwarding control architecture. This paper proposes a fast packet forwarding framework based on the existing kernel to achieve the processing of network traffic acceleration. When the router receives the first packet of the same stream, it saves the content to be forwarded to the new fast forwarding table. In this process, it builds a database. The generated fast forwarding table can be used to guide packet forwarding. Once hit, it can directly extract the next hop information in the table, To guide the forwarding of messages. Experiments have shown that for frequently arriving traffic, it can be directly forwarded through a fast forwarding path. The performance testing experimental results show that building a fast forwarding framework and selecting appropriate routing lookup algorithms can directly forward packets that hit the table entry content when facing a large number of flows, greatly improving the forwarding efficiency.
The possibilistic fuzzy c-means clustering (PFCM) algorithm is a hybrid clustering algorithm based on the fuzzy c-means clustering (FCM) algorithm. It can improve the anti-noise ability of FCM and the stability of PCM. However, there are great challenges when it is applied in image segmentation for images with strong noise injection. Therefore, we propose an improved Possibilistic Fuzzy C-Means Clustering with adaptive region-level information (RIPFCM). The first is that both the region-level information and the original information are integrated into the objective function of PFCM. The second is that we introduce a method to adaptively set the weighted factors which are used to balance the influence of the region-level information and the original information. Finally, experiments of color image segmentation show that the proposed RIPFCM method achieves better results compared with several clustering algorithms.
With the continuous development of GNSS-R technology, its application scope is also expanding. At present, quite fruitful results have been achieved in moving target detection, remote sensing of Marine and land surface parameters. Power monitoring is very important for GNSS-R, and the temperature change of receiver antenna is an important reason affecting its power. Aiming at the problem that the antenna temperature of the receiver cannot be accurately controlled, this paper proposes a fuzzy neural network PID control algorithm through research, and compares the algorithm with the conventional PID algorithm by simulation. The results show that the step response curve of the system under the control of this algorithm can quickly reach the steady state value. And the system test, can be the constant temperature system temperature control within +/- 0.2 degrees C, to meet the high precision temperature control requirements, and the output noise power is stable.
Affected by haze, images often face color distortion, resolution reduction and other image quality degradation problems. The existing dehazing methods based on convolutional neural network(CNN) often perform well on large-scale synthetic datasets, but lack robustness in the processing of real haze images. This is because haze images in reality are often non-homogeneous. Due to the fact that the haze texture faced by the real haze processing is more complex, it is easier to destroy the texture details. Meanwhile, the paired training images are difficult to collect, and the small-scale data set is easy to lead to overfitting. To address these challenges, we propose a dehazing approach based on ensemble learning, DB-GAN, which uses Res2Net pre-trained by ImageNet as the encoder in the knowledge adaptation branch to improve the generalization ability of the network and avoid overfitting. In the data fitting branch, deep bilateral learning is used to learn the structure of the features from the full-resolution and low-resolution inputs, respectively, to better learn the color features and boundary features. We then map the different features by a fusion tail. Finally, we demonstrate the effectiveness of our approach through extensive experimental results.
To address the problems of difficult extraction of tomato disease leaf features, scattered disease locations, and certain similarities that lead to model misclassification, this study proposes an improved ResNet50 model for tomato leaf disease identification with ten types of tomato leaves as research objects. ECA-ResNet is constructed to enhance the convolutional neural network feature map differentiation process and classification accuracy. After that, the inception module is modified and integrated to the backbone network to address the problem of misclassification caused on by similar diseases. The proposed parallel multi-branch structure of the inception module fuses the different scale features of three branches to enhance the recognition of subtle diseases such as small spots, thus improving the feature expression capability of the network. The results show that the classification with 99.24% accuracy, and 99.23% precision. The model is compared with several deep convolutional models, including AlexNet, VGG16, GoogLeNet, MobileNetV2, and other improved models. The results show that the method has better classification performance, offering new approaches for the detection of leaf diseases in tomato as well as the advancement of intelligent agriculture.
The co-occurrence matrix is a way to describe the spatially relevant information of an image. The image thresholding method based on the co-occurrence matrix has its advantages over the thresholding method based on 2D grey-scale histogram. The traditional 2D grey-scale histogram requires two thresholds for binary segmentation, whereas the grey-scale co-occurrence matrix requires only one threshold for binary segmentation. When performing multi-threshold segmentation of an image, using co-occurrence matrix reduces the number of thresholds by half compared to using two-dimensional grey-scale histogram. In this paper, exponential entropy is used to implement co-occurrence matrix threshold selection, and swarm intelligent optimization algorithms is used to reduce the computational effort of multi-threshold selection. This method is compared with the 2D histogram-based method and the traditional log-entropy-based threshold segmentation method on BSD500 images, and the metrics of PSNR, FSIM, and SSIM are analysed in conjunction with the three intelligent optimization algorithms, and the methods in this paper are all superior to other comparative methods.
In the process of navigation and positioning, one of the factors affecting the positioning accuracy is the spatial geometric distribution of the satellite. It is very important to select the satellite with a good spatial distribution. For the traditional satellite selection algorithm based on azimuth Angle and pitch Angle, Because in the satellite selection process, the first lowest satellite is forced to add a fixed angle as a criterion for satellite selection,the satellite selection in the area with dense satellite distribution is more concentrated, which is not conducive to positioning. On the basis of calculating the azimuth and elevation, In this paper, a satellite selection algorithm based on hierarchical clustering is proposed.
The fuzzy C-means (FCM) algorithm is one of the most widely used algorithms in unsupervised pattern recognition learning. However, real-world data is more complex and there may be some irrelevant features in the data that affect the final clustering results of FCM. The weighted clustering algorithm increases the importance of relevant features in the data by assigning different weights to features of different dimensions, and at the same time weakens the influence of irrelevant features on the clustering results. However, both the weighted clustering algorithm and the FCM algorithm will have classification errors as the observation noise increases. As a distance measure between two distributions, relative entropy is added to the objective function as a regularization function, which can minimize the distance within the cluster and maximize the difference between clusters. Therefore, this paper proposes a new feature-weighted relative entropy clustering algorithm (REFCM_EW). The REFCM_EW algorithm combines feature weight and relative entropy, which not only enhances the importance of relevant features in the data but also has better noise detection capability. Experimental results show that REFCM_EW has a good effect on the strip data.
The newly proposed sparse subset selection (DS3) algorithm can effectively perform the clustering of data and the selection of representatives for each cluster (i.e. a subset of the entire data set) simultaneously. It can be formulated as a row-sparsity of membership matrix and minimizing coding loss problem. However, the regularization parameter in DS3 has a great impact on the performance of the algorithm. Different regularization parameter values will lead the algorithm to select a different number of representatives or clusters. However, for a given data set, which can have a definite number of classes, the DS3 algorithm can not automatically select the optimal regularization parameter. In view of this problem that DS3 has, noting that clustering validity criteria can be used to determine the number of clusters and the close relationship between the number of clusters and the value of the regularization parameter, we propose to use the clustering validity criteria to determine the optimal regularization parameter and thus propose an unsupervised DS3 (UDS3) algorithm that can automatically determine the regularization parameter. Compared with the original DS3 algorithm, the proposed UDS3 algorithm will give more constructive clustering results when the number of classes of the data set to be clustered is not given. Video summarization experiments on several Youtube data sets show the effectiveness of the proposed UDS3 algorithm.
The cutset-type possibilistic c-means (C-PCM) clustering algorithm is an efficient method to overcome the coincident clustering problem of the possibilistic c-means (PCM) clustering algorithm. However, when the C-PCM is applied to image segmentation, it tends to generate poor results for noisy images due to the lack of spatial information. Moreover, most of the traditional methods improved clustering segmentation methods by adding spatial information, which will increase the computational complexity and running time of the algorithm. To solve these problems, this paper proposes a fast C-PCM algorithm with spatial information named LLink-CPCM. First, the bilateral filter is used to obtain local spatial information and color information. Second, the information is introduced into the objective function of the C-PCM. In addition, a membership link is introduced into the objective function to decrease the computational complexity and improve the iteration speed. Experiments on color images show that the proposed LLink-CPCMcan improve the segmentation accuracies and reduce the iteration numbers compared with several clustering methods.
The relative entropy fuzzy c-means (REFCM) clustering algorithm is a classical soft clustering approach. REFCM incorporates relative entropy as a regularization term within the objective function of the fuzzy c-means (FCM) clustering algorithm, effectively enhancing its robustness to noise. However, its effectiveness is mainly limited to the clustering of spherical shapes. The kernel trick technique is applied to the REFCM clustering algorithm, wherein the data is mapped to a suitable feature space using a nonlinear mapping to solve the problem. Combining or selecting kernels is very important for efficient kernel clustering. Therefore, we propose a multiple kernel relative entropy fuzzy c-means clustering algorithm (MKREFCM). This algorithm extends single-kernel learning to multi-kernel learning. By integrating several kernels and automatically adjusting weights, the algorithm's sensitivity to kernel selection in the clustering process is minimized. Finally, experiments on non-spherical datasets demonstrate the effectiveness of the proposed algorithm.
With the development of semiconductor technology, the performance of computer processors has been greatly improved, resulting in increased requirements for memory performance.[1] Due to the unbalanced development between the processor and the memory, the storage rate, bandwidth, and operating frequency of the memory are difficult to meet the work requirements of the processor, which limits the overall work efficiency of the computer. This paper conducts in-depth research on DDR3 SDRAM(double-data-rate 3 synchronous dynamic RAM), analyzes various state transitions of DDR3 operation, and divides DDR3 state transitions into initialization module, register configuration module, read and write training module, self-refresh module, and state control module. Use Verilog code for RTL-level implementation, use vivado simulation software for simulation testing, and observe the simulation waveform and timing.[2] The simulation results show that the DDR3 SDRAM controller initialization, refresh, register configuration and read and write modules and other functions are normally implemented.
In order to improve the problems of traditional one-dimensional chaotic mapping, such as small chaotic interval, discontinuous chaotic interval, simple chaotic behavior and low initial value sensitivity, a two-dimensional chaotic system based on 1D-LSE (Logistic self-embedding) chaotic mapping is proposed and further extended to a three-dimensional. Taking the improved two-dimensional chaotic system as an example, various performance indicators, such as the Lyapunov exponent, bifurcation diagram, sensitivity to initial values, bit change rate, and sequence randomness, are analyzed. The experimental simulation results show that the proposed two-dimensional chaotic system has a continuous chaotic interval, a larger chaotic interval, better performance in generating random sequences, higher sensitivity to initial values, and more complex chaotic characteristics.
Quantum satellite communication is a communication method with strong security and large coverage area, and it has become a hot research topic at home and abroad. The successful launch of the quantum satellite has laid an important foundation for future integrated quantum communication. When further studying the communication performance between quantum satellite-ground links, it is found that atmospheric particulate matter has a significant impact on quantum satellite communication performance. In this paper, by analyzing the extinction characteristics of atmospheric particles, a mathematical model of particle concentration, transmission distance and satellite-earth quantum energy is established, and the influence of atmospheric particles on communication parameters, bit error rate and entanglement degree is analyzed. The simulation results show that with the increase of particle concentration, the performance indicators of quantum satellite communication are affected to a certain extent. As the concentration of particulate matter in the atmosphere continues to decrease, the energy of the optical quantum signal continues to increase, and the quantum bit error rate increases with the increase in the concentration of atmospheric particulate matter and the transmission distance. The degree of quantum entanglement decreases with the increase of the particle size of atmospheric particles, and increases with the increase of visibility. The luminosity of light quantum will decrease with the increase of transmission height and the increase of surface particle concentration. By studying the optimization and adjustment strategy of quantum signal based on fidelity quantum relay satellite, simulation ensures the feasibility of this strategy. It is observed that the research in this paper can provide the necessary theoretical reference for the development of quantum space-ground integrated communication network.
Scene text image super-resolution is to design a method to reconstruct super-resolution text images from low-resolution scene text images, which aim to improve the recognition rate of low-resolution text images. Data driven based deep neural network methods are attractive in this field due to superior performance. However, most existing super-resolution methods for scene text images focus more on the quality of the images not text, which leaves a space to improve performance. In the paper, we design a multi-head attention submodule to exploit the masks from four aspects in order to pay more attention to the text information. Specifically, scene text image super-resolution based on multi-head attention is proposed, abbreviated as STISR-MHA, which includes a text image correction, the text feature enhancement, and the super-resolution image reconstruction submodules. Where, the text image correction submodule can correct the distortion of the image; the text feature enhancement submodule can enhance its text image semantic information; the super-resolution image reconstruction submodule can restore a super-resolution image from low-resolution scene text image. Experiments on the benchmark TextZoom dataset show that the proposed model achieved good results in terms of recognition accuracy and image quality assessment metrics.
Automatic Target Detection (ATD) leverages machine learning to efficiently process datasets that are too large for humans to evaluate quickly enough for practical applications. Technological and natural factors such as the type of sensor, collection conditions, and environment can affect image interpretability. Synthetic Aperture Radar (SAR) sensors are sensitive to different issues than optical sensors. While SAR imagery can be collected at any time of day and in almost any weather conditions, some conditions are uniquely challenging. Properties of targets and the environment can affect the radar signatures. In this experiment, we simulated these effects in quantifiable increments to measure how strongly they impact the performance of a machine learning model when detecting targets. The experiments demonstrate the differences in image interpretability for machine learning versus human perception.
Speedy Measurement of Arabidopsis Traits (SMART) is a high-throughput phenotyping pipeline that processes images to observe leaf-specific traits. The current pipeline utilizes a k-means clustering segmentation method. This method has limitations in some cases where leaves display dynamic change in color and shape. Segment-Anything, a new foundational model in segmentation, offers a new method of high-resolution segmentation of complex geometries present in plant systems. Here we present a new method for segmentation of plant using Segment-Anything and Grounded-Dino. A method of obtaining individual leaves was implemented to describe the advantages and limitations of the prompt-based method of deep learning models. An analysis of the segmentation results from Segment-Anything demonstrates that this is a powerful method for providing statistically valuable data for biological insights into novel plant traits under nutrient stress. 1
Massive Open Online Courses (MOOCs) are a contemporary approach to education, providing a large number of open courses to facilitate students' access to knowledge. However, the absence of personalized recommendations based on specific knowledge concepts has decreased students' enthusiasm because of the many courses with different emphases. To address this issue, this paper proposes a catechism resource recommendation model, HGNNRec, which is based on a heterogeneous graph neural network combined with an attention mechanism. The model captures learners' interest preferences by constructing a heterogeneous information network, extracting rich semantic information using meta-paths, performing node feature extraction and weight assignment via graph convolutional networks and attention mechanisms, and finally incorporating matrix decomposition methods for recommendation. The experimental results demonstrate that the proposed method outperforms various baseline and existing methods in predicting and recommending concepts of interest to users. The model can effectively address the problem of personalized recommendations for learners and improve the overall learning experience in MOOCs.
Dynamic graph representation learning aims at learning the low-dimensional vector representation of nodes in dynamic networks, and the learned network representation can be used for various dynamic network analysis tasks.In dynamic networks, it is very important to accurately model time information, but most embedding methods completely ignore it, or use a series of static snapshots to approximate dynamic networks, which will lose a lot of important time information. Besides, community structure is an important feature of complex networks, and capturing network community structure during dynamic network representation learning has a positive impact on learning outcomes. In this paper, we propose a dynamic graph neural network representation learning method with Community-enhanced temporal features, the TCGCN. TCGCN introduces the community structure information of the dynamic network into the extracted temporal features, and then integrates the temporal features into the method based on graph convolution network to learn the representation of the nodes, so that the learning results can keep time dependence, and can maintain the community structure and global network structure features, comprehensive learning to the change of the dynamic network rule. The experimental results show that TCGCN can greatly improve the quality of dynamic network representation learning.
