Search

Bone age assessment (BAA) is a typical clinical technique for diagnosing endocrine and metabolic diseases in children's development. Existing deep learning-based automatic BAA models are trained on the Radiological Society of North America dataset (RSNA) from Western populations. However, due to the difference in developmental process and BAA standards between Eastern and Western children, these models cannot be applied to bone age prediction in Eastern populations. To address this issue, this paper collects a bone age dataset based on the East Asian populations for model training. Nevertheless, it is laborious and difficult to obtain enough X-ray images with accurate labels. In this paper, we employ ambiguous labels from radiology reports and transform them into Gaussian distribution labels of different amplitudes. Furthermore, we propose multi-branch attention learning with ambiguous labels network (MAAL-Net). MAAL-Net consists of a hand object location module and an attention part extraction module to discover the informative regions of interest (ROIs) based only on image-level labels. Extensive experiments on both the RSNA dataset and the China Bone Age (CNBA) dataset demonstrate that our method achieves competitive results with the state-of-the-arts, and performs on par with experienced physicians in children's BAA tasks. [ABSTRACT FROM AUTHOR]

Land desertification is one of the serious ecological and environmental problems facing mankind today, which threatens the survival and development of human society. China is one of the countries with the most serious land desertification problems in the world. Therefore, it is of great theoretical value and practical significance to carry out accurate identification and monitoring of land desertification and its influencing factors in ecologically fragile areas of China. This is conducive to curbing land desertification and ensuring regional ecological security. Minqin County, Gansu Province, located in northwestern China, is one of the most serious areas of land desertification, which is also one of the four sandstorm sources in China. Based on ENVINet5, this paper constructs a high-precision land desertification identification method with an accuracy of 93.71%, which analyzes the trend and reasons of land desertification in this area, provides suggestions for disaster prevention in Minqin County. and provides a reference for other similar areas to make corresponding desertification control policies. [ABSTRACT FROM AUTHOR]

Dragon fruit is one of the most popular fruits in China and Southeast Asia. It, however, is mainly picked manually, imposing high labor intensity on farmers. The hard branches and complex postures of dragon fruit make it difficult to achieve automated picking. For picking dragon fruits with diverse postures, this paper proposes a new dragon fruit detection method, not only to identify and locate the dragon fruit, but also to detect the endpoints that are at the head and root of the dragon fruit, which can provide more visual information for the dragon fruit picking robot. First, YOLOv7 is used to locate and classify the dragon fruit. Then, we propose a PSP-Ellipse method to further detect the endpoints of the dragon fruit, including dragon fruit segmentation via PSPNet, endpoints positioning via an ellipse fitting algorithm and endpoints classification via ResNet. To test the proposed method, some experiments are conducted. In dragon fruit detection, the precision, recall and average precision of YOLOv7 are 0.844, 0.924 and 0.932, respectively. YOLOv7 also performs better compared with some other models. In dragon fruit segmentation, the segmentation performance of PSPNet on dragon fruit is better than some other commonly used semantic segmentation models, with the segmentation precision, recall and mean intersection over union being 0.959, 0.943 and 0.906, respectively. In endpoints detection, the distance error and angle error of endpoints positioning based on ellipse fitting are 39.8 pixels and 4.3°, and the classification accuracy of endpoints based on ResNet is 0.92. The proposed PSP-Ellipse method makes a great improvement compared with two kinds of keypoint regression method based on ResNet and UNet. Orchard picking experiments verified that the method proposed in this paper is effective. The detection method proposed in this paper not only promotes the progress of the automatic picking of dragon fruit, but it also provides a reference for other fruit detection. [ABSTRACT FROM AUTHOR]

Extracting high-accuracy landslide areas using deep learning methods from high spatial resolution remote sensing images is a hot topic in current research. However, the existing deep learning algorithms are affected by background noise and landslide scale effects during the extraction process, leading to poor feature extraction effects. To address this issue, this paper proposes an improved mask regions-based convolutional neural network (Mask R-CNN) model to identify the landslide distribution in unmanned aerial vehicles (UAV) images. The improvement of the model mainly includes three aspects: (1) an attention mechanism of the convolutional block attention module (CBAM) is added to the backbone residual neural network (ResNet). (2) A bottom-up channel is added to the feature pyramidal network (FPN) module. (3) The region proposal network (RPN) is replaced by guided anchoring (GA-RPN). Sanming City, China was selected as the study area for the experiments. The experimental results show that the improved model has a recall of 91.4% and an accuracy of 92.6%, which is 12.9% and 10.9% higher than the original Mask R-CNN model, respectively, indicating that the improved model is more effective in landslide extraction. [ABSTRACT FROM AUTHOR]

Frequent outbreaks of cyanobacterial blooms have become one of the most challenging water ecosystem issues and a critical concern in environmental protection. To overcome the poor stability of traditional detection algorithms, this paper proposes a method for detecting cyanobacterial blooms based on a deep-learning algorithm. An improved vegetation-index method based on a multispectral image taken by an Unmanned Aerial Vehicle (UAV) was adopted to extract inconspicuous spectral features of cyanobacterial blooms. To enhance the recognition accuracy of cyanobacterial blooms in complex scenes with noise such as reflections and shadows, an improved transformer model based on a feature-enhancement module and pixel-correction fusion was employed. The algorithm proposed in this paper was implemented in several rivers in China, achieving a detection accuracy of cyanobacterial blooms of more than 85%. The estimate of the proportion of the algae bloom contamination area and the severity of pollution were basically accurate. This paper can lay a foundation for ecological and environmental departments for the effective prevention and control of cyanobacterial blooms. [ABSTRACT FROM AUTHOR]

Weather radars are commonly used to track the development of convective storms due to their high resolution and accuracy. However, the coverage of existing weather radar is very limited, especially in mountainous and ocean areas. Geostationary meteorological satellites can provide near global coverage and near real-time observations, which can compensate for the lack of radar observations. In this paper, a deep learning method was used to estimate the radar composite reflectivity from observations of China's new-generation geostationary meteorological satellite FY-4A and topographic data. The derived radar reflectivity products from satellite observations can be used over regions without radar coverage. In general, the deep learning model can reproduce the overall position, shape, and intensity of the radar echoes. In addition, evaluation of the reconstruction radar observations indicates that a modified model based on the attention mechanism (Attention U-Net model) has better performance than the traditional U-Net model in terms of all statistics such as the probability of detection (POD), critical success index (CSI), and root-mean-square error (RMSE), and the modified model has stronger capability on reconstructing details and strong echoes. [ABSTRACT FROM AUTHOR]

The accurate prediction of photovoltaic (PV) power is essential for planning power systems and constructing intelligent grids. However, this has become difficult due to the intermittency and instability of PV power data. This paper introduces a deep learning framework based on 7.5 min-ahead and 15 min-ahead approaches to predict short-term PV power. Specifically, we propose a hybrid model based on singular spectrum analysis (SSA) and bidirectional long short-term memory (BiLSTM) networks with the Bayesian optimization (BO) algorithm. To begin, the SSA decomposes the PV power series into several sub-signals. Then, the BO algorithm automatically adjusts hyperparameters for the deep neural network architecture. Following that, parallel BiLSTM networks predict the value of each component. Finally, the prediction of the sub-signals is summed to generate the final prediction results. The performance of the proposed model is investigated using two datasets collected from real-world rooftop stations in eastern China. The 7.5 min-ahead predictions generated by the proposed model can reduce up to 380.51% error, and the 15 min-ahead predictions decrease by up to 296.01% error. The experimental results demonstrate the superiority of the proposed model in comparison to other forecasting methods. [ABSTRACT FROM AUTHOR]

It is an objective reality that deaf-mute people have difficulty seeking medical treatment. Due to the lack of sign language interpreters, most hospitals in China currently do not have the ability to interpret sign language. Normal medical treatment is a luxury for deaf people. In this paper, we propose a sign language recognition system: Heart-Speaker. Heart-Speaker is applied to a deaf-mute consultation scenario. The system provides a low-cost solution for the difficult problem of treating deaf-mute patients. The doctor only needs to point the Heart-Speaker at the deaf patient and the system automatically captures the sign language movements and translates the sign language semantics. When a doctor issues a diagnosis or asks a patient a question, the system displays the corresponding sign language video and subtitles to meet the needs of two-way communication between doctors and patients. The system uses the MobileNet-YOLOv3 model to recognize sign language. It meets the needs of running on embedded terminals and provides favorable recognition accuracy. We performed experiments to verify the accuracy of the measurements. The experimental results show that the accuracy rate of Heart-Speaker in recognizing sign language can reach 90.77%. [ABSTRACT FROM AUTHOR]

In recent years, garbage classification has become a hot topic in China, and legislation on garbage classification has been proposed. Proper garbage classification and improving the recycling rate of garbage can protect the environment and save resources. In order to effectively achieve garbage classification, a lightweight garbage object detection model based on deep learning techniques was designed and developed in this study, which can locate and classify garbage objects in real-time using embedded devices. Focusing on the problems of low accuracy and poor real-time performances in garbage classification, we proposed a lightweight garbage object detection model, YOLOG (YOLO for garbage detection), which is based on accurate local receptive field dilation and can run on embedded devices at high speed and with high performance. YOLOG improves on YOLOv4 in three key ways, including the design of DCSPResNet with accurate local receptive field expansion based on dilated–deformable convolution, network structure simplification, and the use of new activation functions. We collected the domestic garbage image dataset, then trained and tested the model on it. Finally, in order to compare the performance difference between YOLOG and existing state-of-the-art algorithms, we conducted comparison experiments using a uniform data set training model. The experimental results showed that YOLOG achieved A P 0.5 of 94.58% and computation of 6.05 Gflops, thus outperformed YOLOv3, YOLOv4, YOLOv4-Tiny, and YOLOv5s in terms of comprehensive performance indicators. The network proposed in this paper can detect domestic garbage accurately and rapidly, provide a foundation for future academic research and engineering applications. [ABSTRACT FROM AUTHOR]

The multi-functional buoy is an important facility for assisting the navigation of inland waterway ships. Therefore, real-time tracking of its position is an essential process to ensure the safety of ship navigation. Aiming at the problem of the low accuracy of multi-functional buoy drift prediction, an integrated deep learning model incorporating the attention mechanism and ResNet-GRU (RGA) to predict short-term drift values of buoys is proposed. The model has the strong feature expression capability of ResNet and the temporal memory capability of GRU, and the attention mechanism can capture important information adaptively, which can solve the nonlinear time series drift prediction problem well. In this paper, the data collected from multi-functional buoy #4 at Nantong anchorage No. 2 in the Yangtze River waters in China were studied as an example, and first linear interpolation was used for filling in missing values; then, input variables were selected based on Pearson correlation analysis, and finally, the model structure was designed for training and testing. The experimental results show that the mean square error, mean absolute error, root mean square error and mean percentage error of the RGA model on the test set are 5.113036, 1.609969, 2.261202 and 15.575886, respectively, which are significantly better than other models. This study provides a new idea for predicting the short-term drift of multi-functional buoys, which is helpful for their tracking and management. [ABSTRACT FROM AUTHOR]

In this paper, we consider water surface object detection in natural scenes. Generally, background subtraction and image segmentation are the classical object detection methods. The former is highly susceptible to variable scenes, so its accuracy will be greatly reduced when detecting water surface objects due to the changing of the sunlight and waves. The latter is more sensitive to the selection of object features, which will lead to poor generalization as a result, so it cannot be applied widely. Consequently, methods based on deep learning have recently been proposed. The River Chief System has been implemented in China recently, and one of the important requirements is to detect and deal with the water surface floats in a timely fashion. In response to this case, we propose a real-time water surface object detection method in this paper which is based on the Faster R-CNN. The proposed network model includes two modules and integrates low-level features with high-level features to improve detection accuracy. Moreover, we propose to set the different scales and aspect ratios of anchors by analyzing the distribution of object scales in our dataset, so our method has good robustness and high detection accuracy for multi-scale objects in complex natural scenes. We utilized the proposed method to detect the floats on the water surface via a three-day video surveillance stream of the North Canal in Beijing, and validated its performance. The experiments show that the mean average precision (MAP) of the proposed method was 83.7%, and the detection speed was 13 frames per second. Therefore, our method can be applied in complex natural scenes and mostly meets the requirements of accuracy and speed of water surface object detection online. [ABSTRACT FROM AUTHOR]

In recent years, machine learning for trading has been widely studied. The direction and size of position should be determined in trading decisions based on market conditions. However, there is no research so far that considers variable position sizes in models developed for trading purposes. In this paper, we propose a deep reinforcement learning model named LSTM-DDPG to make trading decisions with variable positions. Specifically, we consider the trading process as a Partially Observable Markov Decision Process, in which the long short-term memory (LSTM) network is used to extract market state features and the deep deterministic policy gradient (DDPG) framework is used to make trading decisions concerning the direction and variable size of position. We test the LSTM-DDPG model on IF300 (index futures of China stock market) data and the results show that LSTM-DDPG with variable positions performs better in terms of return and risk than models with fixed or few-level positions. In addition, the investment potential of the model can be better tapped by the reward function of the differential Sharpe ratio than that of profit reward function. [ABSTRACT FROM AUTHOR]

Precipitation has an important impact on people's daily life and disaster prevention and mitigation. However, it is difficult to provide more accurate results for rainfall nowcasting due to spin-up problems in numerical weather prediction models. Furthermore, existing rainfall nowcasting methods based on machine learning and deep learning cannot provide large-area rainfall nowcasting with high spatiotemporal resolution. This paper proposes a dual-input dual-encoder recurrent neural network, namely Rainfall Nowcasting Network (RN-Net), to solve this problem. It takes the past grid rainfall data interpolated by automatic weather stations and doppler radar mosaic data as input data, and then forecasts the grid rainfall data for the next 2 h. We conduct experiments on the Southeastern China dataset. With a threshold of 0.25 mm, the RN-Net's rainfall nowcasting threat scores have reached 0.523, 0.503, and 0.435 within 0.5 h, 1 h, and 2 h. Compared with the Weather Research and Forecasting model rainfall nowcasting, the threat scores have been increased by nearly four times, three times, and three times, respectively. [ABSTRACT FROM AUTHOR]

The number of patients with cardiovascular diseases is rapidly increasing in the world. The workload of existing clinicians is consequently increasing. However, the number of cardiovascular clinicians is declining. In this paper, we aim to design a mobile and automatic system to improve the abilities of patients' cardiovascular health management while also reducing clinicians' workload. Our system includes both hardware and cloud software devices based on recent advances in Internet of Things (IoT) and Artificial Intelligence (AI) technologies. A small hardware device was designed to collect high-quality Electrocardiogram (ECG) data from the human body. A novel deep-learning-based cloud service was developed and deployed to achieve automatic and accurate cardiovascular disease detection. Twenty types of diagnostic items including sinus rhythm, tachyarrhythmia, and bradyarrhythmia are supported. Experimental results show the effectiveness of our system. Our hardware device can guarantee high-quality ECG data by removing high-/low-frequency distortion and reverse lead detection with 0.9011 Area Under the Receiver Operating Characteristic Curve (ROC–AUC) score. Our deep-learning-based cloud service supports 20 types of diagnostic items, 17 of them have more than 0.98 ROC–AUC score. For a real world application, the system has been used by around 20,000 users in twenty provinces throughout China. As a consequence, using this service, we could achieve both active and passive health management through a lightweight mobile application on the WeChat Mini Program platform. We believe that it can have a broader impact on cardiovascular health management in the world. [ABSTRACT FROM AUTHOR]

Modern satellite and aerial imagery outcomes exhibit increasingly complex types of ground objects with continuous developments and changes in land resources. Single remote-sensing modality is not sufficient for the accurate and satisfactory extraction and classification of ground objects. Hyperspectral imaging has been widely used in the classification of ground objects because of its high resolution, multiple bands, and abundant spatial and spectral information. Moreover, the airborne light detection and ranging (LiDAR) point-cloud data contains unique high-precision three-dimensional (3D) spatial information, which can enrich ground object classifiers with height features that hyperspectral images do not have. Therefore, the fusion of hyperspectral image data with airborne LiDAR point-cloud data is an effective approach for ground object classification. In this paper, the effectiveness of such a fusion scheme is investigated and confirmed on an observation area in the middle parts of the Heihe River in China. By combining the characteristics of hyperspectral compact airborne spectrographic imager (CASI) data and airborne LiDAR data, we extracted a variety of features for data fusion and ground object classification. Firstly, we used the minimum noise fraction transform to reduce the dimensionality of hyperspectral CASI images. Then, spatio-spectral and textural features of these images were extracted based on the normalized vegetation index and the gray-level co-occurrence matrices. Further, canopy height features were extracted from airborne LiDAR data. Finally, a hierarchical fusion scheme was applied to the hyperspectral CASI and airborne LiDAR features, and the fused features were used to train a residual network for high-accuracy ground object classification. The experimental results showed that the overall classification accuracy was based on the proposed hierarchical-fusion multiscale dilated residual network (M-DRN), which reached an accuracy of 97.89%. This result was found to be 10.13% and 5.68% higher than those of the convolutional neural network (CNN) and the dilated residual network (DRN), respectively. Spatio-spectral and textural features of hyperspectral CASI images can complement the canopy height features of airborne LiDAR data. These complementary features can provide richer and more accurate information than individual features for ground object classification and can thus outperform features based on a single remote-sensing modality. [ABSTRACT FROM AUTHOR]

Humanoid robots are equipped with humanoid arms to make them more acceptable to the general public. Humanoid robots are a great challenge in robotics. The concept of digital twin technology complies with the guiding ideology of not only Industry 4.0, but also Made in China 2025. This paper proposes a scheme that combines deep reinforcement learning (DRL) with digital twin technology for controlling humanoid robot arms. For rapid and stable motion planning for humanoid robots, multitasking-oriented training using the twin synchro-control (TSC) scheme with DRL is proposed. For switching between tasks, the robot arm training must be quick and diverse. In this work, an approach for obtaining a priori knowledge as input to DRL is developed and verified using simulations. Two simple examples are developed in a simulation environment. We developed a data acquisition system to generate angle data efficiently and automatically. These data are used to improve the reward function of the deep deterministic policy gradient (DDPG) and quickly train the robot for a task. The approach is applied to a model of the humanoid robot BHR-6, a humanoid robot with multiple-motion mode and a sophisticated mechanical structure. Using the policies trained in the simulations, the humanoid robot can perform tasks that are not possible to train with existing methods. The training is fast and allows the robot to perform multiple tasks. Our approach utilizes human joint angle data collected by the data acquisition system to solve the problem of a sparse reward in DRL for two simple tasks. A comparison with simulation results for controllers trained using the vanilla DDPG show that the designed controller developed using the DDPG with the TSC scheme have great advantages in terms of learning stability and convergence speed. [ABSTRACT FROM AUTHOR]

Weather radar echo is the data detected by the weather radar sensor and reflects the intensity of meteorological targets. Using the technique of radar echo extrapolation, which is the prediction of future echoes based on historical echo observations, the approaching short-term weather conditions can be forecasted, and warnings can be raised with regard to disastrous weather. Recently, deep learning based extrapolation methods have been proposed and show significant application potential. However, there are two limitations of existing extrapolation methods which should be considered. First, few methods have investigated the impact of the evolutionary process of weather systems on extrapolation accuracy. Second, current deep learning methods usually encounter the problem of blurry echo prediction as extrapolation goes deeper. In this paper, we aim to address the two problems by proposing a Multi-Level Correlation Long Short-Term Memory (MLC-LSTM) and integrate the adversarial training into our approach. The MLC-LSTM can exploit the spatiotemporal correlation between multi-level radar echoes and model their evolution, while the adversarial training can help the model extrapolate realistic and sharp echoes. To train and test our model, we build a real-life multi-level weather radar echoes dataset based on raw CINRAD/SA radar observations provided by the National Meteorological Information Center, China. Extrapolation experiments show that our model can accurately forecast the motion and evolution of an echo while keeping the predicted echo looking realistic and fine-grained. For quantitative evaluation on probability of detection (POD), false alarm rate (FAR), critical success index (CSI), and Heidke skill score (HSS) metrics, our model can reach average scores of 0.6538 POD, 0.2818 FAR, 0.5348 CSI, and 0.6298 HSS, respectively when extrapolating 15 echoes into the future, which outperforms the current state-of-the-art extrapolation methods. Both the qualitative and quantitative experimental results demonstrate the effectiveness of our model, suggesting that it can be effectively applied to operational weather forecasting practice. [ABSTRACT FROM AUTHOR]

This paper presents a novel approach for semantic segmentation of building roofs in dense urban environments with a Deep Convolution Neural Network (DCNN) using Chinese Very High Resolution (VHR) satellite (i.e., GF2) imagery. To provide an operational end-to-end approach for accurately mapping build roofs with feature extraction and image segmentation, a fully convolutional DCNN with both convolutional and deconvolutional layers is designed to perform building roof segmentation. We selected typical cities with dense and diverse urban environments in different metropolitan regions of China as study areas, and sample images were collected over cities. High performance GPU-mounted workstations are employed to perform the model training and optimization. With the building roof samples collected over different cities, the predictive model with convolution layers is developed for building roof segmentation. The validation shows that the overall accuracy (OA) and the mean Intersection Over Union (mIOU) of DCNN-based semantic segmentation results are 94.67% and 0.85, respectively, and the CRF-refined segmentation results achieved OA of 94.69% and mIOU of 0.83. The results suggest that the proposed approach is a promising solution for building roof mapping with VHR images over large areas in dense urban environments with different building patterns. With the operational acquisition of GF2 VHR imagery, it is expected to develop an automated pipeline of operational built-up area monitoring, and the timely update of building roof map could be applied in urban management and assessment of human settlement-related sustainable development goals over large areas. [ABSTRACT FROM AUTHOR]