Trained Network Model Research Articles

Identification of the Original Images

In social networks, people usually communicate online by posting and passing images, people will find that there are many similar or even visually identical images. When the images determine whether the published information is true or false, original images became crucial, and at the same time, deep learning excelled in image identification. Combining the original image identification and deep learning methods, this paper proposes the original image identification method based on the deep network models, aiming to use the deep network models to identify the original images between the original images and a bunch of variant images of the original images, and obtain satisfactory results. In this way, through the trained network models, the original images can be identified and can help the viewers to determine the correctness of the information. Experimental results verify the feasibility, effectiveness and security of the proposed method.

Research on Algorithm for Authenticating the Authenticity of Calligraphy Works Based on Improved EfficientNet Network

Calligraphy works have high artistic value, but there is the rampant problem of forgery. Indeed, the authentication of traditional calligraphy heavily relies on calligraphers’ subjective judgment. Therefore, spurred by the recent development of neural networks, this paper proposes a method for authenticating calligraphy works based on an improved EfficientNet network. Specifically, the developed method utilizes the character box algorithm to efficiently extract individual calligraphy characters, which are then augmented and used as the training set for the model. The training process employs CBAM and Self-Attention modules to enhance the attention mechanism of the EfficientNet network. The trained network model is used to judge the calligraphy works’ similarity; tested on authentic works, imitated works, and works from other calligraphers; and compared with other networks. The experimental results demonstrate that the proposed method effectively achieves the authentication of calligraphy works, and the improved CBAM-EfficientNet network and SA-EfficientNet network achieve better authentication performance.

An Improved Mask R-CNN Micro-Crack Detection Model for the Surface of Metal Structural Parts.

Micro-crack detection is an essential task in critical equipment health monitoring. Accurate and timely detection of micro-cracks can ensure the healthy and stable service of equipment. Aiming at improving the low accuracy of the conventional target detection model during the task of detecting micro-cracks on the surface of metal structural parts, this paper built a micro-cracks dataset and explored a detection performance optimization method based on Mask R-CNN. Firstly, we improved the original FPN structure, adding a bottom-up feature fusion path to enhance the information utilization rate of the underlying feature layer. Secondly, we added the methods of deformable convolution kernel and attention mechanism to ResNet, which can improve the efficiency of feature extraction. Lastly, we modified the original loss function to optimize the network training effect and model convergence rate. The ablation comparison experiments shows that all the improvement schemes proposed in this paper have improved the performance of the original Mask R-CNN. The integration of all the improvement schemes can produce the most significant performance improvement effects in recognition, classification, and positioning simultaneously, thus proving the rationality and feasibility of the improved scheme in this paper.

Study on deterioration identification method of rubber bearings for bridges based on YOLOv4 deep learning algorithm

How to quickly and accurately identify the bridge rubber bearing deterioration plays an important role in ensuring the bridge structure and road safety. This paper selects the common rubber bearings of domestic bridges as the research object, and proposes an improved YOLOv4-based bridge rubber bearing deterioration detection algorithm to address the reasons for the difficulty in detecting bridge rubber bearing deterioration due to large scale variations and small sample data sets. An image dataset (named HRBD) with annotations is constructed from real inspection scenarios, and the data is expanded by image processing means such as rotation, translation and brightness transformation, so that this dataset has sufficient data complexity and solves the problem of overfitting due to insufficient samples for network training. The anchor applicable to this dataset was regained by the K-means++ clustering algorithm, and then the CA module was inserted into the YOLOv4 backbone network for more accurate anchor localization. The improved YOLOv4 network was used for migration learning to train the dataset, and finally the trained network model was used for detection on the test set. The experimental results show that the improved YOLOv4 bridge rubber bearing deterioration detection and identification network can effectively identify and locate bridge rubber bearings and their deterioration types (crack damage, shear deformation, bearing void).

Measures for the Integration of Digital Media Art and Film and Animation Creation Based on Virtual Reality Technology

Abstract In this paper, we propose a scene animation network (FSGAN) method that uses spectral features to perceive the deep style of an image, which uses three loss functions for perceiving the distribution of image style features in the same style domain to generate anime style-consistent images. The recurrent consistency adversarial network is supplemented with a local reinforcement module based on self-attention, adapted to the animalization of portrait scenes based on their characteristics and needs. This module guides the network to pay attention to the salient regions in the image and combines layer normalization and instance normalization methods through the weighting of the attention weight coefficients to construct a portrait animation network based on local reinforcement perception. Then, based on the digital media technology to complete the virtual reality scene construction, the constructed space for three-dimensional division, and the use of generative adversarial network training model, the proposed key events and random sampling technology combined with the control of virtual animation scene generation method, to achieve the controllability and diversity of the virtual animation scene generation. The first-stage design efficiency, the second-stage design efficiency and the overall design efficiency of the animation creation method in this paper are all basically stabilized between 0.8-1. Compared with the traditional method, they are higher by 0.3932, 0.3596, and 0.5635, respectively. The extreme pixel percentage, standard deviation and contrast of the animation images created by this paper’s method are 0.0349, 0.0037 and 0.3382, respectively, and the overall color performance is better. Thus, the method presented in this paper has the best overall effect on processing anime images using digital media art effects.

Reading recognition of pointer meters based on an improved UNet++ network

Pointer meters are widely used in modern industries, such as petrochemical applications, substations, and nuclear power plants. To overcome the reading errors and inaccurate measurements due to uneven or fluctuating illumination in practical applications, this paper proposes an improved UNet++ network for recognizing pointer meter readings. First, the scale invariant feature transform feature-matching algorithm is used to adjust the captured tilted meter images to a symmetrical and upright shape. Then, the UNet++ network is used to segment the scale and pointer regions in the dashboard to eliminate background interference. Furthermore, part of the convolution in the UNet++ network is replaced with dilated convolution with different expansion rates to expand the perceptual field during network training. In the UNet++ network jump connection, the attention mechanism module is also introduced in the path to enhance the region’s features to be segmented and suppress the parts of the non-segmented area. A hybrid loss function is used for the network model training to prevent the imbalance of the segmented region share. Finally, the distance method is used to read the gauge representation. Experiments were conducted to compare the performance of the proposed method with that of the original UNet++ network in terms of feasibility and precision. The experimental results showed that the recognition reading accuracy was significantly improved by the enhanced network, with the accuracy, sensitivity, and specificity reaching 98.65%, 84.33%, and 99.38%, respectively. Furthermore, when using the improved UNet++ network for numerical reading, the average relative error was only 0.122%, indicating its robustness in a natural environment.

ADGNN: Towards Scalable GNN Training with Aggregation-Difference Aware Sampling

Distributed computing is promising to enable large-scale graph neural network (GNN) model training. However, care is needed to avoid excessive computational and communication overheads. Sampling is promising in terms of enabling scalability, and sampling techniques have been proposed to reduce training costs. However, online sampling introduces large overheads, and while offline sampling that is done only once can eliminate such overheads, it instead introduces information loss and accuracy degradation. Thus, existing sampling techniques are unable to improve simultaneously both efficiency and accuracy, particularly at low sampling rates. We develop a distributed system, ADGNN, for full-batch based GNN training that adopts a hybrid sampling architecture to enable a trade-off between efficiency and accuracy. Specifically, ADGNN employs sampling result reuse techniques to reduce the cost associated with sampling and thus improve training efficiency. To alleviate accuracy degradation, we introduce a new metric,Aggregation Difference (AD), that quantifies the gap between sampled and full neighbor set aggregation. We present so-called AD-Sampling that aims to minimize the Aggregation Difference with an adaptive sampling frequency tuner. Finally, ADGNN employs anAD -importance-based sampling technique for remote neighbors to further reduce communication costs. Experiments on five real datasets show that ADGNN is able to outperform the state-of-the-art by up to nearly 9 times in terms of efficiency, while achieving comparable accuracy to the non-sampling methods.

Ghost imaging object recognition based on self-attention mechanism network

A network model based on the self-attention mechanism is proposed to address the difficulties in extracting features from ghost imaging targets, low recognition efficiency, and potential errors. First, a ghost imaging detection system is constructed using a laser, spatial light modulator, bucket detector, etc. The object is illuminated with speckles generated by the spatial light modulator. The detected data are then input into the self-attention mechanism network model for training. Experimental results show that for the handwritten digits in the experimental dataset, the highest accuracy and average accuracy of the self-attention mechanism network are 99.13% and 96.41%, respectively. This experiment demonstrates the potential of using the self-attention mechanism network for target recognition in ghost imaging, improving the speed of target recognition and significantly enhancing the accuracy of recognition.

SVFGNN: A privacy-preserving vertical federated graph neural network model training framework based on split learning

SVFGNN: A privacy-preserving vertical federated graph neural network model training framework based on split learning

RETRACTED ARTICLE: Dual Siamese Anchor Points Adaptive Tracker with Transformer for RGBT Tracking

Due to environmental conditions, such as rainy days, foggy days, and dim lighting, objects in visible light images are not prominently displayed, leading to an easy loss of targets during tracking. In recent years, many RGB visible light trackers have achieved significant success in addressing visual tracking challenges. However, these trackers perform poorly when tracking targets under special conditions, such as occlusions and low-light scenarios. In contrast, objects in thermal infrared images are more distinct in poor lighting conditions. Given this characteristic, researchers have shown increased interest in the development of trackers that combine thermal infrared and visible light imagery. However, some mainstream RGBT (red–green–blue and thermal) algorithms, such as MANET and ADNET, are based on the anchor-based theory, requiring consideration of anchor box sizes and introducing a substantial number of hyperparameters. This can lead to suboptimal performance when tracking dynamically changing targets. Moreover, these models rely on convolutional neural networks for feature extraction, which have limitations in capturing global features. In this paper, we introduce a novel training network model called DAPAT, which combines the anchor-free concept with Transformer theory. DAPAT differs from previous models in several ways. Specifically, we have designed a straightforward model to extract precise global features from template and search images. We have also incorporated two enhancement modules into the model to improve template and search images of different sizes while suppressing the impact of non-target images. We employ a dual-stream feature fusion network to reduce the loss of image feature information due to feature correlation operations. Finally, we compare the performance of the tracking model proposed in this paper with some advanced RGBT trackers on three data sets (RGBT234, RGBT210, and GTOT). The test results demonstrate that our tracker exhibits improvements in robustness and success rate, among other performance aspects.

Back-propagating errors through artificial neural networks for variable selection

Variable selection is one of the most important contents of machine learning research and a large number of variable selection methods have been proposed at present. From deep learning perspective, we propose a new variable selection method based on multi-layer artificial neural networks in this paper. The method propagates errors gained form a trained network model through the network from output layer to input layer and the values gained by input layer give information about the relative contribution. The variables with big contribution value are selected, which we call neural inverse propagation (NIP). Specifically, after training a neural network, the gained errors between the actual network outputs and the desired outputs go through the network from top to down and arrive at the input layer. The absolute values gained by input units provide information about the relative importance (contribution) of input units, which is that the larger the value is, the more important the corresponding input variable is. We remove k variables corresponding to the first k minimum absolute value gained by input units to the predictive variables, to perform variable selection. The efficiency of the proposed method is illustrated through its application to classification and regression tasks on a number of synthetic datasets and a real-world image classification dataset.

Beyond cocaine and heroin use: a stacking ensemble-based framework for predicting the likelihood of subsequent substance use disorder using demographics and personality traits

Heroin and cocaine addiction is a global health issue that can result in death. These substances have the ability to change cognitive functions and impulsive behavior control. This study explores the relationship between heroin and cocaine use and the development of additional psychoactive substance addictions, such as cannabis and nicotine. It also investigates whether lighter drug use, like cannabis, leads to heavier drug use. Using subsets of heroin, cocaine, and cannabis data for neural network model training, stacking ensemble learning is employed to uncover these connections. These models predict the risk of subsequent substance abuse based on the history of heroin, cocaine, or cannabis abuse, incorporating demographic factors and personality traits. Results reveal significant impacts: Heroin abuse substantially increases the risk of cannabis and nicotine use (F-scores of 0.95 and 0.94, respectively). Cocaine abuse shows an even stronger association (Accuracy: 0.88) and can lead to heroin and cannabis use. Additionally, cannabis use is linked to subsequent cocaine use. These results have important implications for precision medicine, emphasizing the importance of personalized medications in preventing subsequent addiction development.

A Lightweight Low-dose PET Image Super-resolution Reconstruction Method based on Convolutional Neural Network.

PET imaging is one of the most widely used neurological disease screening and diagnosis techniques. Since PET involves the radiation and tolerance of different people, the improvement that has always been focused on is to cut down radiation, in the meantime, ensuring that the generated images with low-dose tracer and generated images with standard-dose tracer have the same details of images. We propose a lightweight low-dose PET super-resolution network (SRPET-Net) based on a convolutional neural network. In this research, We propose a method for accurately recovering highresolution (HR) PET images from low-resolution (LR) PET images. The network learns the details and structure of the image between low-dose PET images and standard-dose PET images and, afterward, reconstructs the PET image by the trained network model. The experiments indicate that the SRPET-Net can achieve a superior peak signal-to-noise ratio (PSNR) and structural similarity index measurement (SSIM) values. Moreover, our method has less memory consumption and lower computational cost. In our follow-up work, the technology can be applied to medical imaging in many different directions.

Multi-scale fusion for few-shot remote sensing image classification

ABSTRACT Few-shot image classification is currently valuable research in the field of remote sensing image applications. The difficulty in obtaining training data for a large number of labelled remote sensing images leads to difficulties in applying traditional deep learning-based remote sensing image classification methods. Existing few-shot remote sensing image classification methods usually use a large number of base datasets to train network models in the pre-training phase and perform few-shot classification tasks after model fine-tuning on new datasets in the meta-testing phase. However, an insufficient number of remote sensing images lead to inaccurate features being extracted by the model, and the large amount of new class data for fine-tuning is also not easily accessible. In addition, using a pre-trained model to extract features from the new dataset will create a”negative transfer” problem. In this paper, we aim to address the above challenges in two ways. First, we use a subset of ImageNet, a readily available few-shot natural image dataset, for model pre-training, and eliminate the fine-tuning operation to simulate a real-world application scenario where remote sensing data is extremely scarce. Second, due to the significant differences in scale and style between the ImageNet dataset and the remote sensing dataset, which lead to a serious negative transfer problem in the model, we design a multi-scale feature fusion module to comprehensively decide the labels of query samples considering all scales to compensate for the scale differences and alleviate the”negative transfer” problem. We conducted experiments on four benchmark remote sensing datasets and achieved satisfactory performance.

A numerical control machining tool path step error prediction method based on BP neural network

Step error calculation of numerical control (NC) machining tool path is a premise for generating high-quality tool path and promoting its application. At present, iterative methods are generally used to calculate step error, and the computation time increases when accuracy improves. Neural networks can be calculated on GPUs and cloud platforms, which is conducive to reducing computation time and improving accuracy through continuous learning. This article innovatively introduces a BP neural network model to predict step error values. Firstly, the core parameters required for step error calculation are taken as the data samples to construct the neural network model, and map to the same scale through Z-score normalization to eliminate the adverse effects of singular parameters on the calculation results. Then, considering only a small number of parameters determine theoretical values of step error, the Dropout technique can drop hidden layer neurons with a certain probability, which is helpful to avoid overfitting and used in the neural network model design. In the neural network model training, this paper adds the Stochastic Gradient Descent with Momentum (SGDM) optimizer to the back propagation of network training in order to improves the network’ stability and accuracy. The proposed neural network predicts step error of samples from three surface models, the results show that the prediction error decreases as sample training increases. After trained by 15% of the surface samples, the neural network predicts the step errors of the remaining samples. Compared with theoretical values, more than 99% of the predicted values have an absolute error less than 1 μm. Moreover, the cost time is only one-third of the geometric method, which verifies the effectiveness and efficiency of our method.

Accurate Visual Simultaneous Localization and Mapping (SLAM) against Around View Monitor (AVM) Distortion Error Using Weighted Generalized Iterative Closest Point (GICP).

Accurately estimating the pose of a vehicle is important for autonomous parking. The study of around view monitor (AVM)-based visual Simultaneous Localization and Mapping (SLAM) has gained attention due to its affordability, commercial availability, and suitability for parking scenarios characterized by rapid rotations and back-and-forth movements of the vehicle. In real-world environments, however, the performance of AVM-based visual SLAM is degraded by AVM distortion errors resulting from an inaccurate camera calibration. Therefore, this paper presents an AVM-based visual SLAM for autonomous parking which is robust against AVM distortion errors. A deep learning network is employed to assign weights to parking line features based on the extent of the AVM distortion error. To obtain training data while minimizing human effort, three-dimensional (3D) Light Detection and Ranging (LiDAR) data and official parking lot guidelines are utilized. The output of the trained network model is incorporated into weighted Generalized Iterative Closest Point (GICP) for vehicle localization under distortion error conditions. The experimental results demonstrate that the proposed method reduces localization errors by an average of 39% compared with previous AVM-based visual SLAM approaches.

Intelligent navigation method for multiple marine autonomous surface ships based on improved PPO algorithm

To achieve the autonomous navigation of multiple Marine Autonomous Surface Ship (MASS), an intelligent planning and decision-making method for MASS based on Rapidly-exploring Random Trees star (RRT-star) and improved Proximal Policy Optimization (PPO) algorithm is proposed. The novelty of the study is: (1) enhancing the traditional PPO algorithm by the Generalized Advantage Estimation (GAE) and the Long Short-Term Memory (LSTM) network, which facilitate accelerated convergence of average reward, predict the state space, and enhance accuracy in estimating the advantage function. (2) proposing a complete reward function, which can not only guide MASS to navigate towards the waypoint, but also ensure that MASS complies with the COLREGs during collision avoidance. It is worth highlighting that the trained network model can be generalized to different scenarios. We attached the trained neural network model to different MASS and simulated it in different open and narrow waters. Even in emergency situations, the method can still make it deviate from COLREGs and make flexible collision avoidance decisions. The results show that this method can handle the multi-MASS encounter situation well and make MASS reach the waypoint safely.

Forecast of pain degree of lumbar disc herniation based on back propagation neural network.

To further explore the pathogenic mechanism of lumbar disc herniation (LDH) pain, this study screens important imaging features that are significantly correlated with the pain score of LDH. The features with significant correlation imaging were included into a back propagation (BP) neural network model for training, including Pfirrmann classification, Michigan State University (MSU) regional localization (MSU protrusion size classification and MSU protrusion location classification), sagittal diameter index, sagittal diameter/transverse diameter index, transverse diameter index, and AN angle (angle between nerve root and protrusion). The BP neural network training model results showed that the specificity was 95 ± 2%, sensitivity was 91 ± 2%, and accuracy was 91 ± 2% of the model. The results show that the degree of intraspinal occupation of the intervertebral disc herniation and the degree of intervertebral disc degeneration are related to LDH pain. The innovation of this study is that the BP neural network model constructed in this study shows good performance in the accuracy experiment and receiver operating characteristic experiment, which completes the prediction task of lumbar Magnetic Resonance Imaging features for the pain degree of LDH for the first time, and provides a basis for subsequent clinical diagnosis.

Applications of Machine Learning to Predict the Flexural Bearing Capacity of Hollow Core Slabs After Fire Exposure

Conventional evaluation of the overall mechanical properties and ultimate flexural capacity of prestressed hollow core slabs after a fire exposure depends heavily on the inversion of fire scene temperature. To avoid this drawback, this paper presents a new methodology which combines a generalized regression neural network (GRNN) with conventional non-destructive testing technology. Thereby, a neural network model for predicting the material performance parameters after fire exposure is obtained based on conventional testing indices. A hollow core slab bridge is used as an example, and the applicability of the trained network model is confirmed using numerical simulation and a field failure test. Results show that the overall relative error of GRNN in predicting the key performance parameters of the bridge after fire exposure is less than 10%. Further, because of the good thermal inertia of the concrete, the relative error in predicting the material performance parameters of steel after a fire is less than 5%. Moreover, the ultimate flexural capacity of the prestressed hollow core slab after a fire can be accurately evaluated by feeding the material performance parameters predicted by GRNN neural network into the finite element (FE) model.

Object Detection Network Based on Module Stack and Attention Mechanism

Currently, visual computer applications based on convolutional neural networks are rapidly developing. However, several problems remain: (1) high-quality graphics processing equipment is needed, and (2) the trained network model has several unnecessary convolution operations. These problems result in a single-stage target detection network that often requires unnecessary computing power and is difficult to apply to equipment with insufficient computing resources. To solve these problems, based on YOLOv5, a YOLOv5-L (YOLOv5 Lightweight) network structure is proposed. This network is improved using YOLOv5. First, to enhance the inference speed of the detector on the CPU, the PP-LCNet (PaddlePaddle-Lightweight CPU Net) is employed as the backbone network. Second, the focus module is removed, and the end convolution module in the head network is replaced by a deep separable convolution module, which eliminates redundant operations and reduces the amount of computation. The experimental results show that YOLOv5-L enables a 48% reduction in model parameters and computation compared to YOLOv5, a 35% increase in operation speed, and a less than 2% reduction in accuracy, which is significant in the environment of low-performance computing equipment.