With the continuous progress and development of society, people pay more and more attention to the safety of themselves and their families. As a daily function of each family, the access control system has attracted much attention for its practicability, security and convenience. In this paper, an intelligent access control system based on STM32 main control chip and its functional modules is designed. The system includes four functions: key unlock, fingerprint recognition unlock, face recognition unlock, and mobile phone APP unlock. At the same time, a solution for cloud control intelligent access control equipment is proposed. The scheme uses MQTT Internet of Things protocol to communicate with Baidu Tiangong cloud platform. Through the cloud platform to transmit instructions, it realizes the establishment of a secure two-way connection between intelligent access control and Baidu Tiangong cloud platform, and quickly realizes cloud control. This intelligent access control system effectively solves the diversity of people's access control unlocking functions, the security of security functions, and the actual needs of monitoring convenience through mobile phones.
Facial paralysis is a relatively frequent disease characterized by facial muscle movement disorders. Clinically, the patient's facial paralysis is usually assessed by the clinician according to assessment scales. The most popular scale for assessing facial paralysis in recent years is the Sunnybrook Facial Grading System (SFGS). However, the scale's assessment results are highly subjective and inefficient. Using artificial intelligence technology for the auxiliary diagnosis of facial paralysis assessment can improve the accuracy and objectivity of diagnosis. But the existing method or model based on computer vision for facial paralysis auxiliary diagnosis assessment mainly suffers from insufficient datasets and inadequate extraction of useful facial feature information. Therefore, this paper proposes an objective method for facial paralysis assessment based on transfer learning and an improved ResNet34 network model. This assessment method is based on SFGS, which assesses the degree of facial paralysis separately for the patient's eyes, nasolabial folds, and mouth when the face rests. Firstly, the keyframe images are extracted from the patients' captured facial videos, and the data's validity is increased by preprocessing with geometric normalization and grayscale normalization. Then, for the strengthened feature extraction problem, an improved resnet34 network (named ResNet34-SP) is designed with SoftPool as the pooling layer of the network, which retains more feature information. Finally, the Extended Cohn-Kanade (CK+) database was used for transfer learning pre-training to improve the problem of insufficient datasets and realize the automatic SFGS static score assessment of facial paralysis. The accuracy rates for the eyes, nasolabial folds, and mouth were 97%, 87%, and 88%, respectively, confirming the method's effectiveness.
Convolutional Neural Networks (CNNs) are currently considered as the state-of-the-art approach for solving many sophisticated computer vision problems. However, the performance of CNNs heavily depends on the quality of the annotated training dataset. In this paper, we present a simple method to enhance the robustness of CNNs to noisy datasets by proposing hinge loss with L2 norm regularization term as the loss function. The proposed method is compared with two commonly used loss functions, binary cross-entropy and mean square error, in binary classification tasks from Fashion-MNIST dataset. Noisy labels were added for evaluating the robustness and generalization performance. Experimental results demonstrate that the proposed loss function outperforms the other two loss functions.
In order to improve the accuracy of breast tumor recognition, this paper proposes a breast tumor aided diagnosis analysis model. The method combines support vector machine (SVM) and genetic algorithms (GA). Firstly, GA is used to reduce the dimension of feature data obtained from breast tumor and the grid search method based on k fold cross validation error is used to select the model parameters of SVM. Following, the classification and recognition of breast tumors are implemented by the proposed method. In comparison with the results of LVQ, PSO-SVM and BP neural network breast tumor diagnosis methods, recognition accuracy of GA-SVM based method is more better than the traditional classification methods. As a result, it can provide decision support and assist doctors to minimize and avoid missing and faulty diagnostic cases when fine needle aspiration cytopathology is used, which is the traditional examination method, to diagnose breast tumors.
It is of no doubt that artificial intelligence (AI) is a cutting-edge technological science, in which manifold learning methods have become a hot research topic in recent years. t-distribution stochastic neighbor embedding (t-SNE) is a representative nonlinear dimensionality reduction method in manifold learning. The 2 main advantages of t-SNE are its ability to maintain structural characteristics and solve congestion problems. However, t-SNE is not parametric, so the 2 initializations of the algorithm will result in 2 different embeddings, thus, a guiding index is added to t-SNE, and thus index t-SNE algorithm has been formed. Nevertheless, like traditional t-SNE, the optimization method of index t-SNE is still prone to fall into the trap of local minimum, resulting in no global optimal solution. And the closer to the global optimal solution, the more oscillatory the parameter optimization path will be, which means that the optimization efficiency of index t-SNE is not high. The root of causing the above problems is that the optimization method of index t-SNE relies on gradient or derivative information. In this paper, a derivative-free swarm intelligence method - Sobol sequence initialized Harris hawk optimization (HHO) has been utilized to overcome the above shortcomings, and the advantages of the proposed method have been proved via numerical experiments.
The neuron model is a widely-used approach for classification problems that imitates the behavior of neurons in the brain. However, most existing neuron models do not account for the non-linear characteristics of dendrites and synapses. To address this limitation, we propose a novel deep learning model, called GoogLeDNet, which combines the GoogLeNet architecture with the dendritic neuron model (DNM). By incorporating the non-linear characteristics of dendrites and synapses, our model closely resembles the structure of biological neurons and demonstrates superior performance on the CelebA gender classification dataset. Compared to other baseline models, including AlexNet, MobileNet, ShuffleNet, and GoogLeNet, our model achieves an accuracy of 93.3% and an F1 score of 93.2%.
ECG signal analysis plays a crucial role in the diagnosis of cardiovascular diseases. Aiming at the problem of ECG signal classification and diagnosis, we propose a classification method based on model fusion of MobileNetV3 and BiLSTM_Attention. Firstly, considering the high solving efficiency of lightweight neural network, MobileNetV3 is used to extract local features in the signal. Secondly, on the basis of extracting local features, the Bi-directional long and short term memory network (BiLSTM) is used to further extract the time sequence relationship of ECG signals. Finally, the attention mechanism is introduced to extract different weights for different features to enhance the classification ability of the model. Example analysis shows that, the classification accuracy of this algorithm is 98.46%, and each epoch takes 21.07 seconds on average, which has excellent comprehensive performance and potential clinical application value.
Traditional tracking often have an online updating process, which often leads to heavy computing burden. We propose an off-line tracking re-detection framework, which is composed of a Siamese convolution network and a correlation filtering tracker. The Siamese convolution network is trained on a large-scale video data to obtain high object discrimination ability, to be a re-detection network. We use this Siamese network to re-detect the tracking results of correlation filtering tracker. In the tracking process, we do not have to update detection network online due to its high object discrimination ability. This mechanism reduces the computational burden of the algorithm and achieves high tracking speed. We also propose a hybrid judgment mechanism to make the re-detection network get better tracking decisions, as well as to re-detect tracked objects after tracking challenges occurring. Experiment results on large-scale data sets demonstrate that the proposed tracking framework outperforms the compared methods and high tracking speed.
As a regression algorithm, logistic regression can be used for not only regression prediction, but also classification. By employing Sigmod function of nonlinear transformation to obtain a value within (0, 1), it can be given the characteristic of probability in logistic regression. According to the rule "value is 1 when $p > 0.5$ , and value is 0 when $p < 0.5$ ", logistic regression can be recognized as a 2-class classification problem and is highly explanatory. However, in the process of solving the parameters of the loss function of traditional logistic regression, gradient descent (GD) method is often used, which means that there is a requirement for the convexity of the function, and it is necessary to reasonably control the size of the hyperparameter in GD, otherwise it is easy to fall into the local optimum value and cannot obtain the global optimal solution. As new methods need to be found to solve the problem of loss function optimization, in this paper, a Faure sequence initialized fruit fly optimization algorithm (FSFFOA) method has been proposed to replace GD method to complete the optimization process. As a swarm intelligence algorithm, FSFFOA has many parameters that can be used to adjust the algorithm accuracy and speed, and has good performance in solving nonlinear optimization problems. Via numerical experiment, the effectiveness of FSFFA has been proved.
This paper studies the rainwater drainage system of underground garage in urban community, and designs an intelligent system based on ARM chip for underground garage water blocking. This system including water level monitoring system, sound and light alarm system and multi-function dam, this paper make the prototype and complete the test of the prototype, testing data proves that water resistance effect is remarkable.
Detecting fake news can be a challenging task. In this paper, we propose a transformer based method to identify and classify texts of fake news. The model is trained and tested on a COVID-19 dataset. The prediction of the proposed method achieved an accuracy of over 95%. The overall performance is superior to three other state-of-the-art machine learning methods, namely Naïve Bays (NB), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP).
Users' viewing behavior could affect their perception and evaluation of design works. Taking into account users' visual attention as a subjective cognition cue, we used eye-tracking evidence to identify users' focus areas for further analysis. We conducted experiments to extract the image features of design images and the reviewers' eye-tracking data, aiming to predict the product design ranking in the competition through fusion data analysis. In particular, we collected 1,504 product design images from a design competition. Four deep convolutional neural networks were selected to explore the best aesthetics computation model. The experimental results show that using design images and eye-tracking data fusion can improve the model prediction performance. Finally, MobileNet-V3 achieves the highest classification accuracy of 74.75%. This suggests the proposed method can provide useful insights into personalized aesthetics evaluation and user-centered design perception.
The rapid development of the computer industry has promoted the development of the entire field of computer vision. The research on target feature extraction is a prerequisite for the application of computer vision technology. The article focuses on the target feature extraction technology of monocular cameras. It proposes to design a feature extraction algorithm based on the OpenCV open-source computer vision library on the basis of obtaining workpiece images from BASLER industrial cameras, which can improve the accuracy and speed of industrial robot grasping and meet the dynamic and real time requirements in actual production.
Dendritic neurons play a crucial role in information processing in neural circuits. Inspired by these neurons, researchers have developed dendritic neural models (DNM) that integrate their properties into conventional deep learning models, yielding outstanding results in various tasks. In this study, DDenseNet model is proposed. DDenseNet combines the advantages of DenseNet and DNM to better simulate brain neuron characteristics and improve deep learning model performance. Layer Normalization (LayerNorm) is added to our model to stabilize data feature distributions and increase convergence speed. Experimental results show that DDenseNet outperforms traditional DenseNet and even some established classic deep learning models in classification tasks. The study suggests that using DNM as a classifier has the potential to create more efficient deep learning models for classification tasks.
LVI-SAM is a new lidar-vision-inertial SLAM fusion framework, which realizes the tight coupling between the visual inertial-system and the lidar inertial system through smoothing and mapping, and it can achieve high-precision and robust real-time state estimation and map building. However, in dynamic scenes, due to the interference of dynamic targets, the extracted lidar feature points may produce incorrect constraints when used for pose optimization, which will lead to the degradation of model accuracy and robustness. To solve the performance degradation problem caused by the interference of dynamic targets in dynamic scenes, an improved LVI-SAM algorithm based on the removal of dynamic target point clouds is proposed, that is, in the system's front end, the region growth algorithm is used to effectively segment the front and rear spots in the point cloud data to remove dynamic targets. Then the processed feature points are used for pose estimation of the lidar inertial navigation subsystem, which further improves the constraint of the lidar odometer in the factor map, and improves the accuracy of depth information obtained by the visual-inertial navigation subsystem, so as to improve the overall positioning accuracy of the system. The simulation experiment on M2DGR dataset shows that the improved algorithm improves the accuracy and robustness of pose estimation in dynamic scenes compared with LVI-SAM.
The construction industry plays a crucial role in contributing to the economy. However, poor real-time monitoring of construction progress often leads to ineffective project management and budget overruns. To address this issue, this paper proposes a computer vision-based system for monitoring construction progress and presents its corresponding architecture. The study explores various technologies including portable data acquisition devices, cloud-edge collaborative computing architecture, progress calculation based on time evolution and multimodal fusion, and construction progress data structure based on lightweight format. These findings serve as the essential groundwork for subsequent development and validation of the system.
Sustainable energy systems, which encompass both renewable energy and energy efficiency, aim to reduce reliance on finite resources such as fossil fuels while minimizing negative impacts on the environment and promoting economic and social development. Wind energy, in particular, is an increasingly important renewable source of energy that can help combat climate change and is cost-competitive with traditional sources of electricity. To maximize energy production, optimizing wind turbine layout is crucial and involves determining the optimal placement and configuration of turbines within a given space. This article proposes a novel approach called fractional-order particle swarm optimization (FOPSO) to address this problem and demonstrates that it outperforms other state-of-the-art optimization algorithms in terms of both solution quality and power generation efficiency.
Artificial intelligence(AI) enterprises, along with the technologies they propose, have had a tremendous impact on the world today. There exists co-occurrence relationships among these enterprises and technologies. However, existing research often represents such co-occurrence relationships using common graphs or simpler data structures, which are inadequate for revealing the underlying correlation. We crawled and mined Chinese patent data, categorizing AI technologies into 57 domains, and combining technology data with enterprise information to construct a technology-enterprise hypergraph. This hypergraph, along with its dual form, is then transformed into a co-occurrence graph through feature calculations. We further applied fast unfolding algorithm to partition the graph, resulting in 6 technology clusters and 9 enterprise clusters that possess strong interpretability.
Credit card fraud detection is a challenging problem due to imbalanced data and adversarial attacks. We propose an adversarial learning with sum of top-K loss (AST) framework for this task. Our approach integrates adversarial learning into logistic regression to mitigate the blind spots of traditional machine learning methods, such that fraudsters manipulate features to evade detection. To further improve the robustness of classification under imbalanced data, we introduce a sum of top-K loss, replacing the plain empirical loss. We have designed a gradient descent approach to optimize the adversarial logistic regression model with the sum of top-K loss. Experiments validate the effectiveness of our AST framework over the prevalent machine learning techniques and existed adversarial learning algorithm.
