Fully Convolutional Neural Network Research Articles

Simulation and prediction of daytime surface urban heat island intensity under multiple scenarios via fully connected neural network

The intensification of the Surface Urban Heat Island (SUHI), driven by urbanization, land use and land cover (LULC) changes, and population growth, presents significant environmental and public health risks in urban areas. Simulating and predicting SUHI, particularly through the identification of future high SUHI intensity (SUHII) zones, has been recognized as a critical step in mitigating these effects. This study employs a Fully Convolutional Neural Network (FCNN) model, trained on data from four research sites, to simulate the current daytime SUHII across six validation sites in Singapore, utilizing 15 key independent variables identified in previous studies. The model exhibits high validation accuracy, achieving 87.45%. Three projection scenarios, based on projected population growth and LULC changes, predict a decrease in High SUHII across all validation sites, ranging from 98.3% to 9%. This reduction is attributed to the LULC improvements proposed in the 2019 Master Plan. Spatial analysis of the predicted SUHII maps indicates that the majority of High SUHII locations across scenarios remain consistent with the current situation. This research also suggests that the model could be a valuable tool for urban planners, allowing them to assess whether new urban development plans will effectively reduce High SUHII to desired thresholds, thereby mitigating SUHII in urban environments.

Diabetic retinopathy data augmentation and vessel segmentation through deep learning based three fully convolution neural networks

ProblemThe eye fundus imaging is used for early diagnosis of most damaging concerns such as diabetic retinopathy, retinal detachments and vascular occlusions. However, the presence of noise, low contrast between background and vasculature during imaging, and vessel morphology lead to uncertain vessel segmentation. AimThis paper proposes a novel retinalblood vessel segmentation method for fundus imaging using a Difference of Gaussian (DoG) filter and an ensemble of three fully convolutional neural network (FCNN) models. MethodsA Gaussian filter with standard deviation σ1 is applied on the preprocessed grayscale fundus image and is subtracted from a similarly applied Gaussian filter with standard deviation σ2 on the same image. The resultant image is then fed into each of the three fully convolutional neural networks as the input. The FCNN models' output is then passed through a voting classifier, and a final segmented vessel structure is obtained.The Difference of Gaussian filter played an essential part in removing the high frequency details (noise) and thus finely extracted the blood vessels from the retinal fundus with underlying artifacts. ResultsThe total dataset consists of 3832 augmented images transformed from 479 fundus images. The result shows that the proposed method has performed extremely well by achieving an accuracy of 96.50%, 97.69%, and 95.78% on DRIVE, CHASE,and real-time clinical datasets respectively. ConclusionThe FCNN ensemble model has demonstrated efficacy in precisely detecting retinal vessels and in the presence of various pathologies and vasculatures.

YOLO-SegNet: A Method for Individual Street Tree Segmentation Based on the Improved YOLOv8 and the SegFormer Network

In urban forest management, individual street tree segmentation is a fundamental method to obtain tree phenotypes, which is especially critical. Most existing tree image segmentation models have been evaluated on smaller datasets and lack experimental verification on larger, publicly available datasets. Therefore, this paper, based on a large, publicly available urban street tree dataset, proposes YOLO-SegNet for individual street tree segmentation. In the first stage of the street tree object detection task, the BiFormer attention mechanism was introduced into the YOLOv8 network to increase the contextual information extraction and improve the ability of the network to detect multiscale and multishaped targets. In the second-stage street tree segmentation task, the SegFormer network was proposed to obtain street tree edge information more efficiently. The experimental results indicate that our proposed YOLO-SegNet method, which combines YOLOv8+BiFormer and SegFormer, achieved a 92.0% mean intersection over union (mIoU), 95.9% mean pixel accuracy (mPA), and 97.4% accuracy on a large, publicly available urban street tree dataset. Compared with those of the fully convolutional neural network (FCN), lite-reduced atrous spatial pyramid pooling (LR-ASPP), pyramid scene parsing network (PSPNet), UNet, DeepLabv3+, and HRNet, the mIoUs of our YOLO-SegNet increased by 10.5, 9.7, 5.0, 6.8, 4.5, and 2.7 percentage points, respectively. The proposed method can effectively support smart agroforestry development.

Within-Class Constraint Based Multi-task Autoencoder for One-Class Classification

Autoencoders (AEs) have attracted much attention in one-class classification (OCC) based unsupervised anomaly detection. The AEs aim to learn the unity features on targets without involving anomalies and thus the targets are expected to obtain smaller reconstruction errors than anomalies. However, AE-based OCC algorithms may suffer from the overgeneralization of AE and fail to detect anomalies that have similar distributions to target data. To address these issues, a novel within-class constraint based multi-task AE (WC-MTAE) is proposed in this paper. WC-MTAE consists of two different task: one for reconstruction and the other for the discrimination-based OCC task. In this way, the encoder is compelled by the OCC task to learn the more compact encoded feature distribution for targets when minimizing OCC loss. Meanwhile, the within-class scatter based penalty term is constructed to further regularize the encoded feature distribution. The aforementioned two improvements enable the unsupervised anomaly detection by the compact encoded features, thereby addressing the issue of the overgeneralization in AEs. Comparisons with several state-of-the-art (SOTA) algorithms on several non-image datasets and an image dataset CIFAR10 are provided where the WC-MTAE is conducted on 3 different network structures including the multilayer perception (MLP), LeNet-type convolution network and full convolution neural network. Extensive experiments demonstrate the superior performance of the proposed WC-MTAE. The source code would be available in future.

UAV Visual Tracking with Enhanced Feature Information

Unmanned aerial vehicles (UAVs) visual tracking is an important research direction. The tracking object is lost due to the problems of target occlusion, illumination variation, flight vibration and so on. Therefore, based on a Siamese network, this study proposes a UAVs visual tracker named SiamDFT++ to enhance the correlation of depth features. First, the network width of the three-layer convolution after the full convolution neural network is doubled, and the appearance information of the target is fully utilized to complete the feature extraction of the template frame and the detection frame. Then, the attention information fusion module and feature deep convolution module are proposed in the template branch and the detection branch, respectively. The feature correlation calculation methods of the two depths can effectively suppress the background information, enhance the correlation between pixel pairs, and efficiently complete the tasks of classification and regression. Furthermore, this study makes full use of shallow features to enhance the extraction of object features. Finally, this study uses the methods of deep cross-correlation operation and complete intersection over union to complete the matching and location tasks. The experimental results show that the tracker has strong robustness in UAVs short-term tracking scenes and long-term tracking scenes.

Automatic Optic Nerve Assessment From Transorbital Ultrasound Images: A Deep Learning-based Approach.

Transorbital Ultrasonography (TOS) is a promising imaging technology that can be used to characterize the structures of the optic nerve and the potential alterations that may occur in those structures as a result of an increase in intracranial pressure (ICP) or the presence of other disorders such as multiple sclerosis (MS) and hydrocephalus. In this paper, the primary objective is to develop a fully automated system that is capable of segmenting and calculating the diameters of structures that are associated with the optic nerve in TOS images. These structures include the optic nerve diameter sheath (ONSD) and the optic nerve diameter (OND). A fully convolutional neural network (FCN) model that has been pre-trained serves as the foundation for the segmentation method. The method that was developed was utilized to collect 464 different photographs from 110 different people, and it was accomplished with the assistance of four distinct pieces of apparatus. An examination was carried out to compare the outcomes of the automatic measurements with those of a manual operator. Both OND and ONSD have a typical inaccuracy of -0.12 0.32 mm and 0.14 0.58 mm, respectively, when compared to the operator. The Pearson correlation coefficient (PCC) for OND is 0.71, while the coefficient for ONSD is 0.64, showing that there is a positive link between the two measuring tools. A conclusion may be drawn that the technique that was developed is automatic, and the average error (AE) that was reached for the ONSD measurement is compatible with the ranges of inter-operator variability that have been discovered in the literature.

Exploring the Impact of Noise and Image Quality on Deep Learning Performance in DXA Images.

Segmentation of the femur in Dual-Energy X-ray (DXA) images poses challenges due to reduced contrast, noise, bone shape variations, and inconsistent X-ray beam penetration. In this study, we investigate the relationship between noise and certain deep learning (DL) techniques for semantic segmentation of the femur to enhance segmentation and bone mineral density (BMD) accuracy by incorporating noise reduction methods into DL models. Convolutional neural network (CNN)-based models were employed to segment femurs in DXA images and evaluate the effects of noise reduction filters on segmentation accuracy and their effect on BMD calculation. Various noise reduction techniques were integrated into DL-based models to enhance image quality before training. We assessed the performance of the fully convolutional neural network (FCNN) in comparison to noise reduction algorithms and manual segmentation methods. Our study demonstrated that the FCNN outperformed noise reduction algorithms in enhancing segmentation accuracy and enabling precise calculation of BMD. The FCNN-based segmentation approach achieved a segmentation accuracy of 98.84% and a correlation coefficient of 0.9928 for BMD measurements, indicating its effectiveness in the clinical diagnosis of osteoporosis. In conclusion, integrating noise reduction techniques into DL-based models significantly improves femur segmentation accuracy in DXA images. The FCNN model, in particular, shows promising results in enhancing BMD calculation and clinical diagnosis of osteoporosis. These findings highlight the potential of DL techniques in addressing segmentation challenges and improving diagnostic accuracy in medical imaging.

Time-domain signal extraction and image reconstruction through scattering medium based on fully convolutional neural network

In scattering environments, laser-active imaging systems are affected by medium absorption and multiple scattering, which reduces the system's ability to effectively detect objects hidden behind the scattering medium. Conventional penetrating scattering-medium imaging methods are less effective under dynamic and strong-scattering conditions. Current deep learning-based image reconstruction techniques only perform image reconstruction from the perspective of the image, without taking the characteristics of the difference between the distribution of the signal and noise in the initially acquired time-domain signals into consideration. In this study, we start from the underlying time-domain distribution characteristics between the signal and the noise and use the full convolutional neural network to learn the time-domain distribution discrepancy between the scattering noise and the echo signal. We directly establish the mapping between noise-containing and noise-free signals to complete the extraction of the target reflective echo signals and to realize target image reconstruction in the scattering medium. We experimentally prove that the proposed method can realize the reconstruction of the target intensity and distance images in a strong scattering environment. Finally, we experimentally demonstrate that the convolutional neural network (CNN) maintains its image reconstruction capability for scattered images outside the range of the training data distribution.

A lightweight deep convolutional network with inverted residuals for matching optical and SAR images

ABSTRACT Matching optical and synthetic aperture radar (SAR) images is often challenged by intricate geometric distortion and nonlinear radiation differences, leading to insufficient and unevenly distributed corresponding points. To tackle this issue, we propose a lightweight deep convolutional network with inverted residuals for optical and SAR image matching. Initially, a fully convolutional neural network (FCNN) is designed to extract high-level and semantic features, robustly capturing universal characteristics between optical and SAR images, adept at handling geometric distortion and nonlinear radiation changes. Notably, we integrate a lightweight architecture with inverted residuals into FCNN to adeptly extract local and global contextual information, facilitating feature reuse and minimizing the loss of crucial features. Additionally, a vector-refined module is deployed to refine dense features, filtering out redundant information. Subsequently, a coarse-to-fine strategy is employed to further eliminate gross errors or incorrect matches. Finally, we evaluate the performance of the proposed network in optical and SAR image matching against manually-designed methods and state-of-the-art deep learning techniques. Experimental results demonstrate that our network significantly surpasses existing methods in terms of the number of correct matches and matching accuracy. Specifically, our proposed network achieves at least a 2.8 times increase in correct matches and an 18% improvement in matching accuracy compared to existing methods.

Generating Multi-Depth 3D Holograms Using a Fully Convolutional Neural Network.

Efficiently generating 3D holograms is one of the most challenging research topics in the field of holography. This work introduces a method for generating multi-depth phase-only holograms using a fully convolutional neural network (FCN). The method primarily involves a forward-backward-diffraction framework to compute multi-depth diffraction fields, along with a layer-by-layer replacement method (L2RM) to handle occlusion relationships. The diffraction fields computed by the former are fed into the carefully designed FCN, which leverages its powerful non-linear fitting capability to generate multi-depth holograms of 3D scenes. The latter can smooth the boundaries of different layers in scene reconstruction by complementing information of occluded objects, thus enhancing the reconstruction quality of holograms. The proposed method can generate a multi-depth 3D hologram with a PSNR of 31.8dB in just 90ms for a resolution of 2160×3840 on the NVIDIA Tesla A100 40G tensor core GPU. Additionally, numerical and experimental results indicate that the generated holograms accurately reconstruct clear 3D scenes with correct occlusion relationships and provide excellent depth focusing.

Microseismic source location using deep learning: A coal mine case study in China

Microseismic source location is crucial for the early warning of rockburst risks. However, the conventional methods face challenges in terms of the microseismic wave velocity and arrival time accuracy. Intelligent techniques, such as the full convolutional neural network (FCNN), can capture spatial information but struggle with complex microseismic sequence. Combining the FCNN with the long short-term memory (LSTM) network enables better time-series signal classification by integrating multi-scale information and is therefore suitable for waveform location. The LSTM-FCNN model does not require extensive data preprocessing and it simplifies the microseismic source location through feature extraction. In this study, we utilized the LSTM-FCNN as a regression learning model to locate the seismic focus. Initially, the method of short-time-average/long-time-average (STA/LTA) arrival time picking was employed to augment spatiotemporal information. Subsequently, oversampling the on-site data was performed to address the issue of data imbalance, and finally, the performance of LSTM-FCNN was tested. Meanwhile, we compared the LSTM-FCNN model with previous deep-learning models. Our results demonstrated remarkable location capabilities with a mean absolute error (MAE) of only 7.16 m. The model can realize swift training and high accuracy, thereby significantly improving risk warning of rockbursts.

The applicability of automated marine clay gully delineation using deep learning in Norway

AbstractGullies and ravines are common landforms in raised marine fine‐grained deposits in Norway. Gullies in marine clay are significant landforms indicative of soil erosion and natural hazards and are of high conservation value. As a result of the substantial impact of human intervention over the past century, marine clay gullies are now red‐listed. To monitor the condition of these landforms, we need to improve our understanding of their spatial extent, complexity and morphology. We explore the applicability of automated approaches that use a methodology of combining deep learning (DL), fully convolutional neural networks (FCNNs) and an unmodified U‐Net model with ArcPy libraries and ground truth data to derive a high‐resolution map of gullies in raised marine fine‐grained deposits. Predictors used comprise solely terrain derivatives to broaden the usage of the pre‐trained model to other regions. Our best model achieved a precision score of 0.82 and a recall of 0.75. We find that our pre‐trained model can successfully predict gullies, also in blind‐test areas. The model performs better in regions with similar geological settings, scoring a length‐weighted overlap of >70% with reference datasets. The novelty of this study is that we demonstrate that the model's applicability for mapping routines increases when we post‐process the predictions by eliminating noise, especially by using the predictions derived from ensembled models. We, therefore, conclude that the pre‐trained models can effectively be used to supplement the geomorphological mapping of marine clay gullies in Norway. The outcome of this research contributes towards mapping the spatial extent and condition of red‐listed landforms in Norway, as well as the development of monitoring systems for future landscape change.

Dynamic monitoring of surface area and water volume of reservoirs using satellite imagery, computer vision and deep learning

The development of surface area and estimated volume measurements for water reservoirs using advanced computational techniques has become an increasingly important task for reservoir operation. Thus, the objective of this study is to determine the surface area and estimate the volume of water reservoirs through the segmentation of multispectral satellite images using a Fully Convolutional Neural Network (FCN). This paper introduces a universally applicable framework that utilizes deep convolutional neural networks for segmenting water bodies, particularly focusing on areas in Paraíba State, Northeast Brazil, susceptible to severe droughts. The study utilized Sentinel-2 satellite data obtained from Google Earth Engine for the period between January 2019 and September 2023. The extraction of water surface area from satellite images was achieved using the U-Net model, a deep learning framework based on FCN. The accuracy of water surface area extraction was analyzed using the Intersection-Over-Union (IoU) metric, also known as Jaccard Index. Subsequently, reservoir volumes were calculated using elevation-area-volume curves. The evaluation metrics for the results included root mean square error (RMSE), mean absolute percentage error (MAPE), coefficient of determination (R2), and Pearson correlation coefficient (r). The network demonstrated high accuracy, achieving a 95% IoU on the test set. This methodology was applied to four reservoirs, yielding promising results that closely aligned the monitored reservoir curves derived from satellite imagery with observed field data. The findings of this study are significant for research on semiarid region changes, extending their relevance to a broader understanding of global environmental dynamics. This work not only contributes to the field of remote sensing in hydrology but also offers practical insights for effective water resource management in drought-prone regions.

Object-Based Semi-Supervised Spatial Attention Residual UNet for Urban High-Resolution Remote Sensing Image Classification

Accurate urban land cover information is crucial for effective urban planning and management. While convolutional neural networks (CNNs) demonstrate superior feature learning and prediction capabilities using image-level annotations, the inherent mixed-category nature of input image patches leads to classification errors along object boundaries. Fully convolutional neural networks (FCNs) excel at pixel-wise fine segmentation, making them less susceptible to heterogeneous content, but they require fully annotated dense image patches, which may not be readily available in real-world scenarios. This paper proposes an object-based semi-supervised spatial attention residual UNet (OS-ARU) model. First, multiscale segmentation is performed to obtain segments from a remote sensing image, and segments containing sample points are assigned the categories of the corresponding points, which are used to train the model. Then, the trained model predicts class probabilities for all segments. Each unlabeled segment’s probability distribution is compared against those of labeled segments for similarity matching under a threshold constraint. Through label propagation, pseudo-labels are assigned to unlabeled segments exhibiting high similarity to labeled ones. Finally, the model is retrained using the augmented training set incorporating the pseudo-labeled segments. Comprehensive experiments on aerial image benchmarks for Vaihingen and Potsdam demonstrate that the proposed OS-ARU achieves higher classification accuracy than state-of-the-art models, including OCNN, 2OCNN, and standard OS-U, reaching an overall accuracy (OA) of 87.83% and 86.71%, respectively. The performance improvements over the baseline methods are statistically significant according to the Wilcoxon Signed-Rank Test. Despite using significantly fewer sparse annotations, this semi-supervised approach still achieves comparable accuracy to the same model under full supervision. The proposed method thus makes a step forward in substantially alleviating the heavy sampling burden of FCNs (densely sampled deep learning models) to effectively handle the complex issue of land cover information identification and classification.

Fault-tolerant Integration of Charge and Discharge Information Security for Multi-Scale Time Series Models

In this paper, the power mode of winding and converter topology are integrated. Considering the filter characteristics and electromagnetic torque control of the system during charging/discharging, the integrated structures of one-phase full bridge, three-phase full bridge and three-phase four-arm bridge are studied. The single-phase all-bridge integrated structure can completely suppress the electromagnetic torque generated by the motor during charging and discharging, and make the filtering effect of the system better. In this project, a multi-scale full convolutional deep neural network method based on time series is studied. The multi-scale full convolution time series model is used for training, and the future continuous signals can be monitored according to the trained model based on the power signal data collected from high-voltage power lines. Experiments show that the proposed method can effectively extract the required data.

Design of an Iterative Method for Enhanced Retinal Image Analysis Using Stacked Deep Learning Operations

The necessity for advanced diagnostic techniques in ophthalmology has become increasingly evident, particularly for conditions such as glaucoma and diabetic retinopathy, which necessitate precise retinal image analysis. Current methodologies, while effective to a degree, fall short in terms of accuracy, efficiency, and adaptability, especially in semantic segmentation and disease classification tasks. These limitations underscore the need for more sophisticated and reliable approaches to retinal image analysis. In response, this work introduces a comprehensive framework leveraging Fully Convolutional Neural Networks (FCNNs) for optic disc segmentation and Cup-to-Disc Ratio (CDR) estimation. The selection of FCNNs is predicated on their demonstrated proficiency in semantic segmentation, ability to discern spatial dependencies, and capacity to learn intricate structures within fundus images without relying on manually engineered features. This approach aims to surpass existing segmentation accuracy benchmarks, targeting over 95% Intersection over Union (IoU) while reducing the mean absolute error in CDR estimation to below 10%. To address the challenges of data variability and quality, the framework incorporates Contrast Limited Adaptive Histogram Equalization (CLAHE) for synthetic data generation, enhancing local contrast while preserving image integrity levels. This method is expected to improve contrast in augmented images by 30%, simultaneously enhancing image quality metrics. Furthermore, the introduction of an overlapping sliding window technique with adaptive patch sizes ensures meticulous coverage and analysis of fundus images, significantly elevating lesion detection sensitivity and reducing false positives. Lastly, the framework employs an innovative multi-disease classification strategy utilizing ensemble learning and stacking of CNN architectures. This synergistic approach amalgamates multiple base models to diminish overfitting risks and enhance generalization capabilities, aiming for a classification accuracy surpassing 95% in binary assessments and 85% for specific retinal diseases. The proposed model not only addresses the current gaps in retinal image analysis but also sets a new standard for precision, reliability, and efficiency in diagnosing and managing ocular diseases, marking a significant step forward in the application of artificial intelligence in ophthalmology.

Segmentation and deep learning to digitalize the kinematics of flow-type landslides

Flow-type landslides, including subaerial and submarine debris flows, have poor spatiotemporal predictability. Therefore, researchers rely heavily on experimental evidence in revealing complex flow mechanisms and evaluating theoretical models. To measure the velocity field of experimental flows, conventional image analysis tools for measuring soil deformation and hydraulics have been borrowed. However, these tools were not developed for capturing the kinematics of fast-moving soil–water mixtures over complex terrain under non-uniform lighting conditions. In this study, a new framework based on deep learning was used to automatically digitalize the kinematics of experimental flow-type landslides. Captured images were broken into sequences and binarized using a fully convolutional neural network (FCNN). The proposed framework was demonstrated to outperform classic image processing algorithms (e.g., particle image velocimetry, trainable Weka segmentation, and thresholding algorithms) over a wide range of experimental conditions. The FCNN model was even able to process images from consumer-grade cameras under complex shadow, light, and boundary conditions. This feature is most useful for field-scale experimentation. With fewer than 15 annotated training images, the FCNN digitalized experimental flows with an accuracy of 97% in semantic segmentation.

Cross-and-Diagonal Networks: An Indirect Self-Attention Mechanism for Image Classification.

In recent years, computer vision has witnessed remarkable advancements in image classification, specifically in the domains of fully convolutional neural networks (FCNs) and self-attention mechanisms. Nevertheless, both approaches exhibit certain limitations. FCNs tend to prioritize local information, potentially overlooking crucial global contexts, whereas self-attention mechanisms are computationally intensive despite their adaptability. In order to surmount these challenges, this paper proposes cross-and-diagonal networks (CDNet), innovative network architecture that adeptly captures global information in images while preserving local details in a more computationally efficient manner. CDNet achieves this by establishing long-range relationships between pixels within an image, enabling the indirect acquisition of contextual information. This inventive indirect self-attention mechanism significantly enhances the network's capacity. In CDNet, a new attention mechanism named "cross and diagonal attention" is proposed. This mechanism adopts an indirect approach by integrating two distinct components, cross attention and diagonal attention. By computing attention in different directions, specifically vertical and diagonal, CDNet effectively establishes remote dependencies among pixels, resulting in improved performance in image classification tasks. Experimental results highlight several advantages of CDNet. Firstly, it introduces an indirect self-attention mechanism that can be effortlessly integrated as a module into any convolutional neural network (CNN). Additionally, the computational cost of the self-attention mechanism has been effectively reduced, resulting in improved overall computational efficiency. Lastly, CDNet attains state-of-the-art performance on three benchmark datasets for similar types of image classification networks. In essence, CDNet addresses the constraints of conventional approaches and provides an efficient and effective solution for capturing global context in image classification tasks.

Adaptive mask-based brain extraction method for head CT images.

Brain extraction is an important prerequisite for the automated diagnosis of intracranial lesions and determines, to a certain extent, the accuracy of subsequent lesion identification, localization, and segmentation. To address the problem that the current traditional image segmentation methods are fast in extraction but poor in robustness, while the Full Convolutional Neural Network (FCN) is robust and accurate but relatively slow in extraction, this paper proposes an adaptive mask-based brain extraction method, namely AMBBEM, to achieve brain extraction better. The method first uses threshold segmentation, median filtering, and closed operations for segmentation, generates a mask for the first time, then combines the ResNet50 model, region growing algorithm, and image properties analysis to further segment the mask, and finally complete brain extraction by multiplying the original image and the mask. The algorithm was tested on 22 test sets containing different lesions, and the results showed MPA = 0.9963, MIoU = 0.9924, and MBF = 0.9914, which were equivalent to the extraction effect of the Deeplabv3+ model. However, the method can complete brain extraction of approximately 6.16 head CT images in 1 second, much faster than Deeplabv3+, U-net, and SegNet models. In summary, this method can achieve accurate brain extraction from head CT images more quickly, creating good conditions for subsequent brain volume measurement and feature extraction of intracranial lesions.

Pore characterization was achieved based on the improved U-net deep learning network model and scanning electron microscope images

The SEM image method is commonly used in the qualitative characterization of shale pores. The development of shale micro-reservoir pores can be visually observed through SEM images, but the efficiency of manual image processing is low and subjective. The introduction of deep learning greatly improves the efficiency of pore analysis. In this paper, the argon ion polishing SEM image of Longmaxi Formation shale in southern Sichuan is taken as an example. Intelligent identification and quantitative characterization of pores in shale SEM images are realized by Pore-net network model. Pore-net is based on the U-net network model. The way the model reads the data is changed so that the model does not focus on the region of interest. The number of convolutional layers of the model is increased. The Canny edge extraction algorithm is added. It not only reduces the workload of data set production, but also enhances the ability of network model to identify pores. The results show that the deep learning semantic image segmentation method is suitable for pore recognition of shale SEM images. The full convolutional neural network model is used to train the manually labeled shale SEM images, which can realize the intelligent recognition and quantitative characterization of shale SEM images. Compared with FCN and DeepLab V3+ network model, Pore-net performs better. Only 170 data sets are used to train the model. The Pore-net network model still has a good recognition effect on pores, which solves the problem of low accuracy of traditional pore recognition methods. The deviation between the porosity calculated by the Pore-net network model and the experimental data is between 12% and 19%. Compared with the porosity results calculated by the binarization method and other network models, the results calculated by Pore-net are closer to the real values, which proves that the porosity calculated by the Pore-net network model is reliable.