The difference in the statistical distribution between the training images and the test images reduces the performance of the steganalysis model, and this sample mismatch phenomenon makes it difficult to improve the accuracy of the detection model. To solve this problem, this paper proposed a deep-learning image steganalysis method based on generalized Gaussian distribution feature clustering. Firstly, a generalized Gaussian distribution is used to fit the coefficients of the image and extract statistical distribution features; Then, the extracted features are clustered to obtain corresponding category labels to achieve sample pre-classification; Finally, different subclasses of samples are used to train the network to achieve more reliable steganalysis. Experimental results show that this scheme can effectively divide images with similar statistical distributions into the same subcategory, reduce the impact of inconsistent statistical characteristics between the training image and the detection image on the steganalysis network, and improve the accuracy of steganalysis.
The operation, maintenance, and decommissioning of nuclear facilities carry high risks and are exposed to high levels of radiation. It is crucial to ensure the safety of personnel involved in nuclear facility decommissioning. When workers work in the radiation area, it is necessary to carry out radiation protection through path planning. Considering the complexity and variability of the environment in the operation process, we designed a new dose minimum path planning algorithm based on the multi-layer grid. Based on this algorithm, the computational complexity of path planning in complex scenarios is greatly reduced, ensuring the optimal dose while reducing the computational time of the algorithm. The simulation results show that the dynamic minimum dose path planning method can effectively ensure the safety of staff during the decommissioning of nuclear facilities.
The electrocardiogram (ECG) signal is vulnerable to being interfered with some unknown noises in the acquisition process due to their low frequency and amplitude, which leads to the loss of significant information in the signals. Recent deep learning models have achieved encouraging results in denoising, however, the generalization ability of the model is not robust to various noises, and the gradient difference between the denoised signal and the original signal is always ignored. In this paper, we propose a deep learning denoising method based on half instance normalization (HIN) block and gradient difference max (GDM) loss function, which includes two stages. In the first stage, we input the noisy ECG signal to obtain the denoised signal. In the second stage, we reconstruct the denoised signal by fusing the preliminary results of the first stage and correct the waveform distortion caused by the first stage denoising, to reduce the loss of information. A novel loss function is also proposed, which can consider the difference between the slope of the denoised ECG signal and the clean ECG signal. The experimental results on MIT-BIH databases show that our method reaches optimal performance both in signal-to-noise ratio (SNR) and root-mean-square error (RMSE).
For the networked linear switching system, a minimal conservative Kalman filter design method is proposed in the network transmission environment with multiple packet losses of both measurement and switching signal packets. Both the measurement signal and the switching signal transmitted in the same packet may be lost through the network. The random multiple packet dropout phenomenon is assumed to be subject to the Bernoulli random binary distribution. A novel Kalman filter is designed to estimate the state values by minimizing the filtering error covariance. Then, the desired Kalman gain matrices are derived, which is a novel switched Kalman gain matrix containing a random change of switching rate. In addition, the stability of the Kalman filter is analyzed in this paper. Finally, simulation results show the effectiveness of the proposed filter.
Traffic light detection is a part of modern intelligent driving, which affects the driving safety of smart cars. However, the object of traffic lights is small and the actual road conditions are complex. The classic object detection algorithm cannot achieve a good detection effect on small signal lights. For this situation, This paper proposes an improved yolov5 model. Because the object of the traffic light is small, the feature extracted by the P2 layer with a small receptive field is introduced. Inspired by the MOAT model, a C3_MOAT structure is proposed. At the same time, a structural re-parameterized RepConv module is added to the Head part of YOLOV5 to improve the feature extraction of the model. ability. Through the above improvements, the ability of the model to detect signal light objects is improved. Experiments show that the improved algorithm improves the map@.5 and map@.5:.95 indicators in the Bosch dataset by 2.6% and 2.7% respectively, and in the S2TLD dataset map@.5 and map@.5:.95 Indicators were increased by 2.2% and 1.6%. Compared with the YOLOV5 network, it has a good improvement.
The existing array non-contact ECG sensing technology mostly uses rigid electrode or conductive fabric electrode, in which the rigid electrode is too hard to make the electrode difficult to close to the human body well, and the conductive fabric electrode is too soft to cause the electrodes to be wrinkled, both of which will lead to poor ECG signal quality. In order to overcome the problem that it is difficult to obtain high-quality ECG signals due to electrode materials, this study combined the manufacturing process of Flexible Printed Circuit (FPC) with the surface treatment process of Electroless Nickel/Immersion (ENIG) and adopted polyimide (PI) and rolled copper as electrode materials, which makes the array electrode electrodes extremely flexible, tough, and flat. At the same time, in terms of electrode design, the test and verification of the electrode unit also ensured the rationality of the array electrode. Subsequently, the electrode was verified by electrode test and human experiment, and the results showed that the electrode could receive ECG signals well, and compared with the ECG signals obtained by the traditional Ag/AgCl electrodes, the signal-to-noise ratio (SNR) of the signals obtained by the two types of electrodes reached above 38 dB and the waveform characteristics were highly consistent.
It is necessary to detect overlapping communities in complex network research tasks, since the actual social network is complex, and many nodes do not fully belong to a certain organization. Therefore, we propose a new algorithm to detect overlapping communities by incorporating the fuzzy logic theory. By calculating and transforming the node similarity, the most closely connected nodes are divided into the corresponding communities by controlling the parameters, and this process generates a preliminary nonoverlapping community structure. The remaining boundary nodes are redistributed according to the degree of membership among the nodes. The nodes that meet certain conditions are identified as overlapping nodes, while the rest of the nodes are divided into the corresponding communities according to their inclinations. In this manner, the communities can be divided and the overlapping nodes can be identified. The proposed algorithm is validated by applying it to three real datasets. The experimental results indicate that the proposed algorithm can successfully find the overlapping nodes by detecting the community structure, which provides guidance for further investigation.
Due to the poor stability of aerial detection equipment, the small proportion of ground target pixels and the occlusion of high-rise objects, air-to-ground target detection has brought great challenges. In order to detect ground targets accurately, we propose a modified YOLOV5 model based on visual attention.The ConvMixer architecture based on the hybrid idea is added to the backbone network of YOLOv5 to obtain more context information. In addition, the GhostNetV2 module is used to improve the neck network of YoloV5 to reduce the number of model parameters and calculation. K-means++ clustering algorithm is used to select the initial anchor box which is as close as possible to the global optimal solution. Finally, Wise-IoU evaluation index is used to improve the original IoU measurement. The experimental results show that the AP value of the proposed method is 0. 797, which is 6. 7% higher than that of YOLOv5. On the basis that the number of parameters and amount of computation are less than the original model, the method effectively inhibits the ground complex background interference and improves the accuracy of air-to-ground target detection.
Classrooms with double-sided windows in southern China lack good measures to improve the daylight factor and reduce glare due to large windows, so the lighting indoor environment is usually uncomfortable due to the high glare and uneven distribution of daylight on sunny days. Based on the digital simulation of the light environment, this paper studies the effect of the properties of 5 types of glass material in improving the factor of daylight and daylight glare. First, a classroom was chosen as a research object and the evaluation parameters (daylight factor and glare) were measured. Secondly, a simulation model was created and the measured and simulated data were compared to verify the validity of the data. Five types of glass were selected and their effect was studied to improve the daylight factor and reduce glare. simulation Results show that: The degree of transmittance of glass technologies has a significant impact on the improvement of classroom lighting and the daylight factor. Controlling the degree of transmittance of the glass results in a greater degree of glare control and thus improved visual comfort. The computer numerical simulations used in this research assisted to a high extent in providing a prediction model to study the effect of glazing techniques in improving the daylight factor and reducing the glare.
Using the method of inserting a fast-response buffer stage between the error amplifier and the power transistor, combined with a single-stage folded cascode amplifier circuit and a PMOS transistor power device, an optimized linear regulator circuit structure for digital chips is designed. Based on the 0.13um process library of SMIC, the output voltage is 1.204V, the power supply voltage regulation rate is 1mv/V, the response time is less than 1.15us, the load regulation rate is 19.8uV/V, the maximum load current is 20mA, the load capacitance is 2~100pF, and the quiescent current is 6.97uA. The simulation results show that the circuit has superior performance parameters, simple structure, easy implementation, no off-chip capacitance, fast instantaneous response, small static current. It can provide a stable low voltage power supply for digital chips, and has high application value.
With the continuous improvement of the digitalization level in the field of energy and electric power, the combination of blockchain technology and business in the field of energy and electric power is getting deeper and deeper. This paper designs a blockchain consensus algorithm combined with artificial intelligence, aiming at the efficiency, security and other problems existing in the existing blockchain consensus algorithm. First, on the basis of ensuring the reliability of the blockchain system, combined with the integrity performance of the consensus node of the blockchain, combined with the recurrent neural network(RNN) to predict the credibility value of the node, screened out the reliable consensus node to participate in the consensus of the energy blockchain, and punished the malicious node. This algorithm can effectively enhance the consensus efficiency of the energy blockchain and improve the level of data privacy protection. We will further support the construction of new power systems.
Traditional recommendation systems often rely on explicit feedback, such as user ratings and favorites. However, this approach suffers from data sparsity and inaccuracy issues. In contrast, implicit feedback data, such as user browsing history and click records, is much richer. Therefore, this paper uses implicit feedback data as the basis for recommendations, models user behavior to uncover their potential interests, and transforms them into vector representations.This paper also introduces interactive interest modeling. This approach models user interest representations based on implicit feedback user behavior sequences. For sequence recommendation problems, we use a Transformer model to extract deeper features from interactive interest representations and use an attention mechanism to capture users' real-time interests. Experimental results on Kaggle datasets demonstrate that this method outperforms traditional implicit feedback recommendation methods in terms of accuracy and personalization.
In Electroencephalogram signal analysis, feature extraction and signal recognition accuracy can be limited by the loss of information resulting from single-domain analysis methods. To address this, we propose a preprocessing method that integrates temporal and frequency characteristics using a modified generalized Stockwell transform, developed from wavelet and short-time Fourier transforms. This method effectively generates time-frequency spectral maps for left and right hand motion imagery Electroencephalogram signals, which are then used to extract features using a multi-scale convolutional neural network with frequency domain spectral interception, space domain convolution, and time domain variance layers. Finally, the extracted features are input to a fully connected layer for softmax mapping classification. We evaluate the performance of our method on the IV2a dataset of the BCI Competition, achieving an accuracy of 78.98% and a kappa value of 0.719. These results demonstrate that our proposed method has better generalization ability and recognition performance compared to other related methods that combine time-frequency analysis and traditional neural networks.
Dispersion effects are important in rendering scenes that contain translucent objects. Existing ray-tracing algorithms must rely on high-end GPU hardware to achieve real-time rendering when rendering dispersion. Based on the image-space caustic map technique, a concise and efficient real-time dispersion rendering method is proposed in the paper, in which the rendering efficiency is improved by reducing the number of monochromatic lights that need to be emitted based on the idea of single charge coupled device image sensor. A variable-size photon drawing strategy is proposed for filling the gaps formed by the insufficient number of monochromatic light samples. The experimental results show that the proposed method can achieve real-time rendering on PC. Our proposed method can simulate the whole continuous spectrum with 7 monochromatic lights of discrete sampled spectrum, reducing the computation and storage of rendering, and it improves the problem of noise formed by insufficient number of emitted rays.
The research on the influence maximization problem in signed networks has a wide range of applications, because this way of influence propagation is more similar to real social networks. However, it is still a very challenging problem to accurately select a set of the most influential set of seed nodes. Most of the existing heuristics focus on the topology of the network and thus there are omissions in the selection of seed nodes. To solve the above problem, we propose an information entropy-based node attribute influence maximization (IENAIM) algorithm. The algorithm combines the social relationships of nodes with the closeness of nodes to select the most influential nodes. Comparative experiments in four real social network datasets validate the superiority of our algorithm in terms of influence propagation results.
Single-cell RNA-seq (scRNA-seq) provides RNA expression profiles for each cell independently, enabling accurate classification of single cells and identification of potential cell heterogeneity. Due to the high dimension and noise of scRNA-seq data, direct use of dimensionality reduction methods ignore potential cell features and corrupt the structure of data. To address these problems, we propose a Hyperbolic feature Fusion for scRNA-seq Classification (HFSC). Specifically, we map the scRNA-seq data to a low-dimensional hyperbolic space. Since the hyperbolic space does not satisfy the requirement of permutation invariance, we further perform hyperbolic transformations on the above hyperbolic features. After that, the fusion of hyperbolic features and original Euclidean features not only preserves the low dimensional hyperbolic features of the data, but also integrates the original features of the data, which takes into account the representation of cells in different spaces. Finally, we perform the node classification task on six datasets. Experimental results show that HFSC has good performance, and higher classification accuracy compared with the other baselines.
The volume morphology of human brain hippocampi have important medical significance for the diagnosis of neuropsychiatric diseases. This paper proposes a visualization system based on V-Net for MRI human brain hippocampi segmentation. We improved the V-Net 3D image segmentation network by bottleneck architecture, and we designed a dice loss function in the network to obtain better hippocampus segmentation results. Then we segment bilateral hippocampi by our improved network model, and the segmentation results are displayed on the GUI interface, allowing doctors to slide layer by layer to observe segmented effect. In the network model of this article included 132 samples. 100 of samples were selected randomly used to train, and the remaining 32 were used to test the performance of our model. During training, we adopted fivefold cross-validation method. The experimental result shows that compared with the V-Net prototype, the use of the network model in this paper for segmentation maintains average dice score of bilateral hippocampi segmentation at 0.88, while the weight model is compressed to more than 15 times, and segmentation time is reduced by 0.34s. This proves that our improved network model can segment human brain bilateral hippocampi accurately, and it also have advantages of high segmentation efficiency and small size network model. And our developed visualization system can better assist doctors in human brain disease diagnosis.
Accurate prediction of the Exhaust Gas Temperature (EGT) of the Aircraft Auxiliary Power Unit (APU) can effectively monitor the future operating status of the APU and prevent safety accidents from occurring. An APU performance parameter prediction method that incorporates Atrous Convolutional Neural Network (ACNN), Bi-directional Long Short-Term Memory (Bi-LSTM) and Attention mechanism is proposed. By combining the characteristics of different networks, the purpose of improving the accuracy of EGT prediction is achieved. ACNN is introduced to extract features from the input parameters, and then the Bi-LSTM network is used to learn and train the extracted feature parameters and combine with the Attention mechanism to give different weights to different feature parameters to achieve parameter prediction. The experimental results show that for single-step and multi-step prediction of EGT, the prediction accuracy of the proposed model is significantly improved compared with other prediction models, providing a certain reference for APU performance change trend prediction.
Most of the dynamic convolution algorithms adopted at this stage use the SE attention mechanism, but the attention mechanism, as a key part of dynamic convolution, has not attracted enough attention, and the relevant research is insufficient. In this paper, an exquisite ODConv which is called Channel-Spatial dynamic convolution is proposed. CSConv introduces the spatial attention module and the channel attention module into the ODConv in parallel, so that the convolution kernel pays more attention to the basic characteristics of the input and effectively improves the accuracy of the model and the efficiency of the convolution kernel. The experimental results show that CSConv achieves good results in the four datasets of ImageNet, COCO, HRRSD and DIOR.
In this paper, a multi-vision system is used to study camera calibration, target feature extraction and feature matching algorithms. First, the multi-vision system used is calibrated, and the basic composition is binocular cameras and monocular cameras, and the number can be expanded according to needs. Then, the camera is calibrated, and the internal and external parameters of each camera and the position relationship between the image coordinate system and the world coordinate system are obtained. In terms of target recognition, the feature point extraction and matching method based on SIFT algorithm is used to identify the target. In order to eliminate false matching and make the recognition result more accurate, the RANSAC algorithm is also used to eliminate the initial mismatched feature points of the SIFT algorithm, retain the accurate matching feature points, and realize the accurate identification of the target.
