The accurate automated brain tumor segmentation is important for disease analysis and control and increases the likelihood of survival. However, it faces significant challenges due to the low contrast of the tissue bo...
详细信息
The accurate automated brain tumor segmentation is important for disease analysis and control and increases the likelihood of survival. However, it faces significant challenges due to the low contrast of the tissue boundary and the small size of the tumor. Convolutional Neural networks are a common automated image evaluation technique that has greatly improved current state-of-the-art precision in the task of segmenting brain tumors. This paper presents an advanced encoder-decoder algorithm with Depthwise Atrous Spatial Pyramid Pooling network (EDD-Net). Firstly, Dilated-ResNet block with Squeeze-and Excitation is introduced in the encoder and decoder module to derive image features adaptively and focuses on the relevant characteristics of the brain segmentation task with fewer parameters. Then, the Depthwise atrous Spatial Pyramid Pooling(DSPP) technique is used as the transition and output layers of the network, to achieve the multi-scale extraction of the feature image and preserves more spatial information. Furthermore, to speed up learning we propose a down-sampling module, while the up-sampling module can more efficiently aggregate low- and high-level feature information. The proposed method is evaluated on the Brats 19 dataset. Experiments demonstrate that the EDD-Net provides high accuracy and robustness in small tumor segmentation. On the online validation set, the suggested ensemble achieved dice scores of 0.813, 0.873, and 0.866 for tumor enhancement, whole tumor enhancement, and tumor core enhancement, respectively, performing favorably compared with existing state-of-the-art architectures.
Tongue and its movements can be used for several medical-related tasks, such as identifying a disease and tracking a rehabilitation. To be able to focus on a tongue region, the tongue segmentation is needed to compute...
详细信息
Tongue and its movements can be used for several medical-related tasks, such as identifying a disease and tracking a rehabilitation. To be able to focus on a tongue region, the tongue segmentation is needed to compute a region of interest for a further analysis. This paper proposes an encoder-decoder CNN-based architecture for segmenting a tongue in an image. The encoder module is mainly used for the tongue feature extraction, while the decoder module is used to reconstruct a segmented tongue from the extracted features based on training images. In addition, the residual multi-kernel pooling (RMP) is also applied into the proposed network to help in encoding multiple scales of the features. The proposed method is evaluated on two publicly available datasets under a scenario of front view and one tongue posture. It is then tested on a newly collected dataset of five tongue postures. The reported performances show that the proposed method outperforms existing methods in the literature. In addition, the re-training process could improve applying the trained model on unseen dataset, which would be a necessary step of applying the trained model on the real-world scenario.
Pitch bearing is one of the critical components of the electric pitch system in wind turbines (WTs), and its early and reliable fault warning is of great significance to ensure the operational reliability and safety o...
详细信息
Pitch bearing is one of the critical components of the electric pitch system in wind turbines (WTs), and its early and reliable fault warning is of great significance to ensure the operational reliability and safety of the entire system. In this article, we develop a condition monitoring system for pitch bearing with available supervisory control data acquisition (SCADA) data. Motivated by the three-blade symmetrical structure of WTs, we propose a new symmetry-aware pitch feature encoder-decoder network named PitchNet. A group feature encoding strategy is first designed to extract intragroup variable-wise feature information. Then, a feature attention mechanism is applied to discover important intergroup features by dynamically assigning different weights to learned pitch representations. Finally, a feature decoder is used to reconstruct the pitch input features and derive the reconstruction residuals for early fault detection. The proposed PitchNet is evaluated using real cases of pitch-bearing cracks. Results have demonstrated that the proposed PitchNet presents better early fault warning ability and lower false alarm rates (FARs), which can provide promising and reliable monitoring results for wind farm operators.
Directed Energy Deposition (DED) is a metal additive manufacturing technology that is gaining popularity for its ability to rapidly manufacture virtually any metal components no matter how complex the shapes and prope...
详细信息
Directed Energy Deposition (DED) is a metal additive manufacturing technology that is gaining popularity for its ability to rapidly manufacture virtually any metal components no matter how complex the shapes and properties are. However, the current lack of real-time geometry monitoring and control is hindering the wider dissemination of DED in industries. This study developed and validated a geometry monitoring methodology which can achieve real-time inspection of the melt pool and newly solidified layer, and layer-wise inspection of the deposited layer during DED process. An encoder-decoder network was developed and applied to the profile images from the laser line scanner to obtain track profiles. A point cloud generation method was proposed to convert the obtained track profiles into 3D point cloud data using intrinsic/extrinsic calibration and printing position. Experiments have been successfully conducted to validate the proposed methodology by depositing multi-layer X-shape objects.
In recent years, the deployment of structural health monitoring (SHM) systems has become paramount for safeguarding critical infrastructures. Notwithstanding, the development of an unsupervised deep learning framework...
详细信息
In recent years, the deployment of structural health monitoring (SHM) systems has become paramount for safeguarding critical infrastructures. Notwithstanding, the development of an unsupervised deep learning framework capable of learning from long-term sensor data remains a critical challenge, particularly in accurately assessing the exact damage location. This study addresses this gap by proposing a novel approach for rapid bridge damage assessment. The proposed method employed a deep overcomplete encoder-decoder network (DOEDN) to reconstruct the acceleration data acquired from each sensor node on the bridge. The reconstruction losses generated by the DOEDN framework are then used as damage-sensitive features. Additionally, a damage indicator based on Gaussian processes is introduced to assess the damage location and evaluate its severity. The performance and sensitivity of the proposed DOEDN framework are evaluated through long-term monitoring acceleration data from a numerical highway bridge model and the well-known full-scale Z24 bridge. Furthermore, comparative assessments against the regular deep undercomplete encoder-decoder network are conducted using metrics including mean absolute error, coefficient of determination (R2), and mean intersection over union. The results show that the proposed DOEDN framework can reasonably assess the damage location and evaluate its severity across various structural scenarios in the bridge, even in the presence of temperature variations, thus providing a practical and effective solution for bridge health monitoring.
As the application of online partial discharge (PD) measurement increases the importance of denoising becomes increasingly obvious. Besides denoising a PD signal to detect and calculate discharge amplitude, pulse reco...
详细信息
As the application of online partial discharge (PD) measurement increases the importance of denoising becomes increasingly obvious. Besides denoising a PD signal to detect and calculate discharge amplitude, pulse reconstructing is required to compute rise time, fall time and other features which play a crucial role in determining discharge severity and identifying defect types. In this paper, a deep learning method based on adversarial deep network by using encoder-decoder network as a generator is developed to perform a denoising and signal reconstruction task. The issue is to provide a database that contains both noisy and denoised data because every denoising method has its limitation, and it is not possible to train a network with real data and their perfect denoised pulses. Therefore, a synthetic database is developed, and signal deformation and noise added synthetically. The trained network's performance is evaluated under actual conditions in two distinct laboratories on various days, with differing noise levels and waveforms. The proposed method outperforms the wavelet method in denoising synthetic data and shows an improvement on real data, while successfully reconstructing the PD pulses. Enhancing the network's performance further, it underwent fine-tuning with actual noise, which led to a marked enhancement in its denoising ability and overall capabilities.
Marine oil spills are currently among the most challenging problems in marine environments. Synthetic Aperture Radar (SAR), with all-day and all-weather observation, has demonstrated great potential in oil disaster mo...
详细信息
Marine oil spills are currently among the most challenging problems in marine environments. Synthetic Aperture Radar (SAR), with all-day and all-weather observation, has demonstrated great potential in oil disaster monitoring. A simple encoder-decoder network model was proposed by using SAR images for effective oil spill detection. The proposed model incorporates an optimized U-Net architecture that reduces computational requirements while maintaining detection performance. This is achieved through shortened encoder and decoder stages, depthwise separable convolutions, group normalization, and bilinear interpolation-based upsampling. To improve the model's generalization, the auto-learning rate and focal loss function are also included. Two public SAR datasets acquired by different sensors have been used to illustrate the efficiency of the proposed model. In addition, the polarimetric information has also been assessed for the detection of oil spill monitoring. The results show that the model can achieve high efficiency at a small model size. The sub-dataset with polarimetric features also achieves a high F1-score of 91.65% and an IoU of 84.59%.
The visual quality of images captured in rainy weather is degraded by rain streaks, which may significantly reduce the accuracy of computer vision systems. In this paper, an internal and external transmission encoder-...
详细信息
The visual quality of images captured in rainy weather is degraded by rain streaks, which may significantly reduce the accuracy of computer vision systems. In this paper, an internal and external transmission encoder-decoder network is proposed for eliminating rain streaks in a single image. Since rain streaks and background structures tend to be spatially long, we apply an internal connected aggregation unit to leverage global information and aggregate features at different scales to help remove rain streaks while restoring scene details. To fully exploit effective features of the encodernetwork and remove redundant information, we utilize an external connected enhancement unit to obtain effective feature maps for predicting clear outputs. By utilizing the efficient information inside and outside the unit, the proposed method can remove the rain streaks of different sizes and directions in the image and restore clear background structure details. Extensive experimental results on synthetic datasets and natural samples show that our method can achieve better rain removal performance than other advanced methods.
Measuring and extracting abrasive grains on the entire surface of monolayer brazing grinding wheels to analyze the distribution of abrasive grains are of great significance to grinding research and grinding wheel manu...
详细信息
Measuring and extracting abrasive grains on the entire surface of monolayer brazing grinding wheels to analyze the distribution of abrasive grains are of great significance to grinding research and grinding wheel manufacture. It is not easy to carry out the work with traditional methods. In this paper, a linear CCD is used to acquire the entire grinding wheel surface image, and an improved encoder-decoder network based on Efficientnet, ASPP, and Skip Connections is designed to promote the accuracy and speed of abrasive grains prediction. Based on dataset creation, transfer learning, and hyper-parameter testing, 91.21% abrasive grain semantic segmentation accuracy was finally obtained. If we focus on whether the abrasive grains are recognized without considering semantic errors, an accuracy rate of 99.6% is obtained. The method can provide basic data for grinding mechanism research and abrasive tool manufacturing.
Height estimation from single remote sensing image is a challenging inherently ambiguous and technically ill-posed problem that we address in this study by resorting to deep learning approach. A spatial enhanced and m...
详细信息
Height estimation from single remote sensing image is a challenging inherently ambiguous and technically ill-posed problem that we address in this study by resorting to deep learning approach. A spatial enhanced and multi-scale aggregated encoder-decoder network, SM-EDNet, is proposed, which takes a single image as input and produces an estimated height map as output. First, residual network (ResNet) is applied to extract low-level and deep features to cope with the heterogeneous characteristics of remote sensing scenes. Then, the multi-scale context information is aggregated through DenseASPP (Dense Atrous Spatial Pyramid Pooling) by extracting features from multiple dilated convolution layers. The skip connection is constructed by using the structure preserving model, DULR, to aggregate ResNet low-level features and multi-scale high-level features. The deformable convolution module is constructed to enhance the sensitivity to differences in geometric shapes of ground objects. For model training, three-layer deep supervision mechanism is designed to counteract the adverse effects of unstable gradients changes. Experimental results on three benchmark datasets, including ISPRS Vaihingen, ISPRS Potsdam, and DFC2018, show that the proposed method achieves the most outstanding performance compared with the state-of-the-art networks. The source codes are available at: https://***/xjh0929/2HEIGHT.
暂无评论