Explore the vast range of titles available.

Diabetic retinopathy (DR) is a common complication of long-term diabetes, affecting the human eye and potentially leading to permanent blindness. The early detection of DR is crucial for effective treatment, as symptoms often manifest in later stages. The manual grading of retinal images is time-consuming, prone to errors, and lacks patient-friendliness. In this study, we propose two deep learning (DL) architectures, a hybrid network combining VGG16 and XGBoost Classifier, and the DenseNet 121 network, for DR detection and classification. To evaluate the two DL models, we preprocessed a collection of retinal images obtained from the APTOS 2019 Blindness Detection Kaggle Dataset. This dataset exhibits an imbalanced image class distribution, which we addressed through appropriate balancing techniques. The performance of the considered models was assessed in terms of accuracy. The results showed that the hybrid network achieved an accuracy of 79.50%, while the DenseNet 121 model achieved an accuracy of 97.30%. Furthermore, a comparative analysis with existing methods utilizing the same dataset revealed the superior performance of the DenseNet 121 network. The findings of this study demonstrate the potential of DL architectures for the early detection and classification of DR. The superior performance of the DenseNet 121 model highlights its effectiveness in this domain. The implementation of such automated methods can significantly improve the efficiency and accuracy of DR diagnosis, benefiting both healthcare providers and patients.

Since the impoundment of the Three Gorges Reservoir area in 2003, the potential risks of geological disasters in the reservoir area have increased significantly, among which the hidden dangers of landslides are particularly prominent. To reduce casualties and damage, efficient and precise landslide susceptibility evaluation methods are important. Multiple ensemble models have been used to evaluate the susceptibility of the upper part of Badong County to landslides. In this study, EasyEnsemble technology was used to solve the imbalance between landslide and nonlandslide sample data. The extracted evaluation factors were input into three bagging, boosting, and stacking ensemble models for training, and landslide susceptibility mapping (LSM) was drawn. According to the importance analysis, the important factors affecting the occurrence of landslides are altitude, terrain surface texture (TST), distance to residences, distance to rivers and land use. The influences of different grid sizes on the susceptibility results were compared, and a larger grid was found to lead to the overfitting of the prediction results. Therefore, a 30 m grid was selected as the evaluation unit. The accuracy, area under the curve (AUC), recall rate, test set precision, and kappa coefficient of a multi-grained cascade forest (gcForest) model with the stacking method were 0.958, 0.991, 0.965, 0.946, and 0.91, respectively, which a significantly better than the values produced by the other models.

The precision detection of dense small targets in orchards is critical for the visual perception of agricultural picking robots. At present, the visual detection algorithms for plums still have a poor recognition effect due to the characteristics of small plum shapes and dense growth. Thus, this paper proposed a lightweight model based on the improved You Only Look Once version 4 (YOLOv4) to detect dense plums in orchards. First, we employed a data augmentation method based on category balance to alleviate the imbalance in the number of plums of different maturity levels and insufficient data quantity. Second, we abandoned Center and Scale Prediction Darknet53 (CSPDarknet53) and chose a lighter MobilenetV3 on selecting backbone feature extraction networks. In the feature fusion stage, we used depthwise separable convolution (DSC) instead of standard convolution to achieve the purpose of reducing model parameters. To solve the insufficient feature extraction problem of dense targets, this model achieved fine-grained detection by introducing a 152 × 152 feature layer. The Focal loss and complete intersection over union (CIOU) loss were joined to balance the contribution of hard-to-classify and easy-to-classify samples to the total loss. Then, the improved model was trained through transfer learning at different stages. Finally, several groups of detection experiments were designed to evaluate the performance of the improved model. The results showed that the improved YOLOv4 model had the best mean average precision (mAP) performance than YOLOv4, YOLOv4-tiny, and MobileNet-Single Shot Multibox Detector (MobileNet-SSD). Compared with some results from the YOLOv4 model, the model size of the improved model is compressed by 77.85%, the parameters are only 17.92% of the original model parameters, and the detection speed is accelerated by 112%. In addition, the influence of the automatic data balance algorithm on the accuracy of the model and the detection effect of the improved model under different illumination angles, different intensity levels, and different types of occlusions were discussed in this paper. It is indicated that the improved detection model has strong robustness and high accuracy under the real natural environment, which can provide data reference for the subsequent orchard yield estimation and engineering applications of robot picking work.

Diabetes is a chronic disease characterized by the inability of the pancreas to produce enough insulin or the body’s inability to use insulin efficiently. This disease is becoming increasingly prevalent worldwide and can result in severe complications such as blindness, kidney failure, and stroke. Early detection of diabetes can potentially save millions of lives globally, making it a crucial focus of research. In this study, we propose a machine learning model to aid in predicting diabetes. The model comprises several machine learning methods: Linear Regression (LnR), Logistic Regression (LR), k-nearest neighbor (KNN), Naive Bayes (NB), Random Forest (RF), Support Vector Machine (SVM), and Decision Tree (DT). Prior to feeding the pre-processed data into the machine learning model for evaluation, we conducted several pre-processing steps, such as removing null values, standardizing data using normalization, and labeling data using the label encoding process. Imbalanced datasets can adversely affect the accuracy of machine learning algorithms, and we address this problem by balancing the datasets using the Synthetic Minority Oversampling Technique (SMOTE) method. We assessed the model’s performance on two datasets and found that the random forest algorithm produced optimal results, with 97% accuracy on the diabetes dataset 2019 and 80% accuracy on the Pima Indian dataset. However, using a balanced dataset, we can significantly reduce the number of false-negative detections.

In this research, 200 corrosion images of steel and 500 crack images of rubber bearing are collected and manually labeled to build the data set. Then the two data sets are respectively adopted to train VGG-Unet models in two methods, aiming to conduct Damage Segmentation by inputting different size of data set. One method is Squashing Segmentation to input squashed images from high resolution directly into VGG-Unet model while Cropping Segmentation uses cropped image with size 224 × 224 as input images. Because the proportion of damage pixels in the data set is different, the results produced by the two data sets are quite different. For large size damage (such as corrosion) segmentation, Cropping Segmentation has a better result while for minor damage (such as crack) segmentation, the result is opposite. The main reason is the gap in the concentration of valid data from the data set. To improve the capability of crack segmentation based on Cropping Segmentation, Background Data Drop Rate (BDDR) is adopted to reduce the quantity of background images to control the proportion of damage pixels from the data set in pixel-level. The ratio of damage pixels from the data set can be decided by different value of BDDR. By testing, the accuracy of Cropping Segmentation becomes relatively higher under BDDR being 0.8.

Peanut leaf diseases that occur throughout the growth process of peanuts seriously affect the peanut yield and quality. The timely and accurate identification and diagnosis of disease with appropriate early treatment measures can effectively avoid the risk of yield and quality losses caused by leaf lesions. Due to the low professional knowledge level of plant growers, the traditional manual diagnosis exhibits low accuracy and causes manpower wastage. Therefore, the present study proposed an online recognition method for peanut leaf diseases based on the data balance algorithm and deep transfer learning. The data balance algorithm was used to solve the problem of data distribution tilt. Furthermore, transfer learning was used to construct a peanut leaf disease recognition model to enhance the generalisation ability based on the lightweight convolutional neural network by removing the original network output layer, re-adding the normalisation and pooling layers, modifying the fully connected layer, and introducing the regularisation constraint strategies. Finally, the deployment and analysis of the model were completed in low-cost embedded devices. This allowed the rapid on-site identification of the healthy state, black spot disease, brown spot disease, net spot disease and mosaic disease for single leaves. Using the self-built peanut leaf disease dataset, three lightweight convolutional neural networks, namely MobileNet V2, Xception and NasNetMobile, were trained and deployed. Comparative experiments showed that the average macro accuracy for peanut leaf disease recognition reached 0.978, 0.990 and 0.974. The average on-site diagnostic accuracy of peanut leaf diseases exceeded 85%. Thus, the present study provides new methods for the development of a portable peanut leaf disease diagnostic device.

Large amounts of co-seismic landslides provide abundant debris sources following a strong earthquake and are prone to initiate and generate debris flow under heavy rainfall. Assessing the susceptibility they pose and what drives their movement in the years following the mainshock has not yet been attempted, primarily because multitemporal debris flow inventories are lacking. This study conducted statistical analyses using the multitemporal debris flow inventory (2008–2021) following the 2008 Wenchuan earthquake and a set of conditional factors (debris source, terrain, and hydrology). The dynamic susceptibility evaluation model was created using logistic regression. The temporal evolution of these factors affecting debris flow runout was explained. Our findings suggest that topography variables grew rapidly after the earthquake, while debris source factors dominated for seven years after the disaster before declining progressively to zero. These findings may have significant consequences for traditional susceptibility assessment models in regions where co-seismic landslides are the dominant debris flow source.

The application of deep learning methods in the research on intelligent recognition of fault images in Electric Multiple Units (EMUs) represents a crucial approach to alleviating the heavy workload of the Train of EMU Failures Detection System (TEDS) inspections and safeguarding train safety. In light of the current problems with TEDS image data, such as irregular formats, substantial quality disparities, and uneven sample distribution, which pose challenges for deep learning, this paper proposes a method for constructing an intelligent recognition dataset of TEDS EMU fault images. By integrating the characteristics of TEDS images with the prior knowledge of EMU operations, a unified data acquisition protocol and standardized dataset construction techniques are devised. Consequently, an intelligent recognition dataset of TEDS EMU fault images is established. This dataset accomplishes effective data integration and standardization and has been successfully implemented in the research on automatic recognition algorithms for TEDS EMU faults. It accelerates the progress of intelligent recognition technology for TEDS fault images and lays a solid technical foundation for ensuring the safety of EMU operations. A standardized dataset integrating TEDS image features and operational knowledge addresses quality and distribution issues, enabling efficient deep learning‐based fault recognition for safer EMU operations.

With the rapid growth and increasing sophistication of malicious software (malware), traditional detection methods face significant challenges in addressing emerging threats. Machine learning-based detection approaches rely on manual feature engineering, making it difficult for them to adapt to diverse attack patterns. In contrast, while deep learning methods can automatically extract features, they remain vulnerable to data imbalance and sample scarcity, which lead to poor detection performance for minority-class samples. To address these issues, this study proposes a semantic data augmentation approach based on a Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP), and designs a malware detection model that combines a Text Convolutional Neural Network (TextCNN) with a Transformer Encoder, termed Mal-CGP-TTN. The proposed model establishes a symmetry between local feature extraction and global semantic representation, where the convolutional and attention-based components complement each other to achieve balanced learning. First, the proposed method enriches the semantic diversity of the training data by generating high-quality synthetic samples. Then, it leverages multi-scale convolution and self-attention mechanisms to extract both local and global features of malicious behaviors, thereby enabling hierarchical semantic modeling and accurate classification of malicious activities. Experimental results on two public datasets demonstrate that the proposed method outperforms traditional machine learning and mainstream deep learning models in terms of accuracy, precision, and F1-score. Notably, it achieves substantial improvements in detecting minority-class samples.

The number of Unmanned Aerial Vehicles (UAVs) used in various industries has increased exponentially, and abnormal detection of UAVs is one of the primary technical means to ensure that UAVs can work normally. Currently, most anomaly detection models are trained using on-board logs from drones. However, in some cases, using these logs can be problematic due to data encryption, inconsistent descriptions of characteristics, and imbalanced positive and negative samples. Consequently, the on-board logs of UAVs may not be directly usable for training anomaly detection models. Given the above problems, this paper proposes a Time Line Modeling (TLM) method based on the UAV software-in-the-loop (SITL) simulation environment to obtain and process the on-board failure logs of drones. The Time Line Modeling method includes two stages: the Fault Time Point Anchoring Method and Fault Time Window Stretching Method. First, based on the SITL simulation environment, multiple flight missions were constructed. Failures of several common components of UAVs are designed. Secondly, the fault’s initial location and end location are determined by the method of Fault Time Point Anchoring, and the original collection of tagged UAV’s on-board data is realized. Then, in terms of data processing, the features that are not universal are removed, and the flight data of the UAV is optimized by using the data balance method of Time Window Stretching to achieve the balance of normal data and abnormal data. Finally, use of algorithms such as Sequential Minimal Optimization (SMO), Random Forest (RF), and Convolutional Neural Network (CNN) were used to experiment with the processed data. The experimental results showed that the data set obtained based on this method can be effectively applied to the training of machine learning-based anomaly detection models.