Disaster early warning method and system based on key target detection and image segmentation

Abstract

The invention belongs to the field of disaster early warning, and provides a disaster early warning method and system based on key target detection and image segmentation, and the method comprises the steps: obtaining the real-time image data of a site, and carrying out the preprocessing of an image; processing the preprocessed image through a YOLOv5 target detection algorithm, detecting a key target in a disaster site, and outputting a category label and a bounding box of the target; performing pixel-by-pixel segmentation processing on a bounding box region output by target detection by using a U-Net image segmentation algorithm to generate an accurate segmentation mask of the target; and inputting the image segmentation mask, historical disaster data, meteorological data and geographic information system data into an LSTM prediction model, carrying out multi-source data fusion and time sequence analysis, and outputting dynamic prediction information and early warning levels of disasters. According to the invention, real-time monitoring and early warning of disasters are realized, the response speed of the early warning system is obviously improved, more comprehensive disaster early warning information is provided, high precision of disaster prediction is ensured, and the phenomena of false alarm and missing alarm are reduced.

Description

Technical Field

[0001] The present invention belongs to the field of disaster warning, and in particular relates to a disaster warning method and system based on key target detection and image segmentation. Background Art

[0002] With the acceleration of global climate change and urbanization, the frequent occurrence of natural disasters such as fires, floods, and earthquakes poses a major threat to the safety of human society. Traditional disaster monitoring and early warning systems usually rely on sensors, manual observations, or simple image analysis methods. Their response speed and accuracy are subject to many limitations, making it difficult to cope with complex and changing disaster scenarios. Especially in the early stages of a disaster, quickly and accurately predicting the spread of a disaster and assessing its potential threat range are the keys to effective early warning and timely response.

[0003] In recent years, with the rapid development of artificial intelligence and deep learning technologies, disaster detection and prediction methods based on computer vision have gradually attracted widespread attention. Target detection algorithms (such as YOLO, SSD, etc.) can effectively identify key targets in images, while image segmentation algorithms (such as U-Net, FCN, etc.) can accurately extract boundary information of disaster areas. However, existing single image processing algorithms often lack the ability to comprehensively analyze complex disaster scenes, and it is difficult to combine multi-source data (such as meteorological, geographic information, historical disaster data, etc.) for comprehensive risk assessment and trend prediction. In addition, disasters have significant temporal evolution characteristics, and existing methods lack the ability to predict the dynamic development of disasters, resulting in delayed or inaccurate warning information. Summary of the invention

[0004] In order to solve the problems in the prior art, the present invention provides a disaster warning method based on key target detection and image segmentation, the method comprising the following steps: Acquire real-time image data on site and perform pre-processing operations on the images; The preprocessed images are processed using the YOLOv5 target detection algorithm to detect key targets at the disaster site and output the target’s category label and its bounding box. Use the U-Net image segmentation algorithm to perform pixel-by-pixel segmentation on the bounding box area output by the target detection to generate an accurate segmentation mask of the target; The image segmentation mask, historical disaster data, meteorological data, and geographic information system data are input into the LSTM prediction model to perform multi-source data fusion and time series analysis, and output dynamic prediction information and warning levels of disasters.

[0005] Furthermore, in the basic network of YOLOv5, deep convolution and residual connection are introduced to improve the feature extraction capability, and a feature pyramid network structure is added.

[0006] Furthermore, an adaptive Anchor box is introduced into the YOLOv5, and the target size distribution in the disaster scene is analyzed by the K-Means clustering algorithm, and the adaptive Anchor box is regenerated to better match the key target size in the disaster scene, avoiding the false detection and missed detection problems caused by the Anchor box being too large or too small.

[0007] Furthermore, in the U-Net, the bounding box area output by YOLOv5 is adaptively adjusted, and only the region of interest detected by the target is used as the input of the U-Net, thereby reducing the interference of irrelevant information caused by the full image processing.

[0008] Furthermore, the U-Net introduces deep residual connections in the encoder part and introduces dilated convolutions in the convolutional layer of the U-Net.

[0009] Another aspect of the present invention is to provide a disaster warning system based on key target detection and image segmentation, the system comprising the following modules: An acquisition module is used to acquire real-time image data on site and perform preprocessing operations on the image; The detection module is used to process the preprocessed images using the YOLOv5 target detection algorithm, detect key targets at the disaster site, and output the target's category label and its bounding box; The segmentation module is used to perform pixel-by-pixel segmentation processing on the bounding box area output by target detection using the U-Net image segmentation algorithm to generate an accurate segmentation mask of the target; The prediction module is used to input the image segmentation mask and historical disaster data, meteorological data, and geographic information system data into the LSTM prediction model, perform multi-source data fusion and time series analysis, and output dynamic prediction information and warning levels of disasters.

[0010] Furthermore, in the basic network of YOLOv5, deep convolution and residual connection are introduced to improve the feature extraction capability, and a feature pyramid network structure is added.

[0011] Furthermore, an adaptive Anchor box is introduced into the YOLOv5, and the target size distribution in the disaster scene is analyzed by the K-Means clustering algorithm, and the adaptive Anchor box is regenerated to better match the key target size in the disaster scene, avoiding the false detection and missed detection problems caused by the Anchor box being too large or too small.

[0012] Furthermore, in the U-Net, the bounding box area output by YOLOv5 is adaptively adjusted, and only the region of interest detected by the target is used as the input of the U-Net, thereby reducing the interference of irrelevant information caused by the full image processing.

[0013] Furthermore, the U-Net introduces deep residual connections in the encoder part and introduces dilated convolutions in the convolutional layer of the U-Net.

[0014] The present invention has the following beneficial effects: The improved YOLOv5 and U-Net algorithms can quickly detect and accurately segment key targets in disaster scenarios, and combined with the LSTM model for dynamic prediction, real-time monitoring and early warning of disasters can be achieved, significantly improving the response speed of the early warning system.

[0015] The present invention uses image segmentation technology to generate pixel-by-pixel precise segmentation masks, accurately extracts the boundaries and shapes of key targets, and combines multi-source data (such as meteorology, geographic information systems, and historical disaster data) for in-depth analysis to ensure high accuracy of disaster prediction and reduce false alarms and missed reports.

[0016] By integrating image segmentation masks with historical disaster data, meteorological data, and geographic information system (GIS) data, and comprehensively analyzing the multi-dimensional characteristics of disasters, the model can capture the dynamic changes of disasters in complex scenarios and provide more comprehensive disaster warning information.

[0017] The LSTM model has powerful time series processing capabilities, can memorize and analyze the patterns of disaster changes over time, and predict the expansion trend of disasters. By updating real-time data, the system can continuously and adaptively adjust the prediction results to improve the accuracy and effectiveness of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0020] Below, the invention is preferably described in conjunction with the accompanying drawings and specific implementation methods.

[0021] This embodiment solves the above problem through the following steps: In one embodiment, reference Figure 1, the present invention proposes a disaster warning method based on key target detection and image segmentation, the purpose of which is to accurately identify key targets through real-time monitoring of disaster scenes using deep learning algorithms, and accurately define the scope of the disaster area in combination with advanced image segmentation technology. By combining target detection and segmentation technology, this method can not only identify key objects at the disaster site, such as flames, flood water levels, collapsed buildings, etc., but also accurately obtain the boundaries and shapes of these targets, and further integrate and analyze with multi-source data to generate disaster development trend predictions and warning information. Specifically, the method of the present invention includes the following steps: Acquire real-time image data on site and perform preprocessing operations on the images.

[0022] Real-time image information of the disaster site is collected by cameras, drones, satellites or other image acquisition devices installed at the disaster site. The image data may include color images or infrared images of multiple dimensions; the image acquisition device can capture changes in the scene in real time and transmit the image data to the data processing system. In the image preprocessing stage, the acquired image data is subjected to operations such as noise removal, brightness adjustment, contrast enhancement and image scaling to improve the quality and clarity of the image and ensure the accuracy of target detection and image segmentation in subsequent image processing steps. The preprocessing also includes selecting an appropriate image enhancement algorithm according to the specific application scenario, removing redundant interference information in the image, such as background noise and irrelevant objects, and maintaining the integrity of key target areas, so as to provide optimized input for subsequent algorithm processing. The preprocessed images are processed using the YOLOv5 target detection algorithm to detect key targets at the disaster site and output the target’s category label and its bounding box.

[0023] This step processes the preprocessed image data through the improved YOLOv5 target detection algorithm. The YOLOv5 algorithm is optimized in many aspects for disaster scenes. The specific steps are as follows: In order to adapt to different types of disaster scenarios, we first build and expand a disaster scenario dataset. The dataset includes image data of typical disasters such as fire, flood, and building damage after earthquake, covering image samples under different weather and lighting conditions. At the same time, we manually annotate key targets in the dataset (such as flames, smoke, floods, building cracks, etc.) to generate high-quality label files to ensure that the model can fully learn the detailed features of the disaster scene during training.

[0024] Through targeted data set expansion and annotation, the model can better understand the complex environment of the disaster site and improve the detection accuracy of different disaster types and key targets.

[0025] In the basic network of YOLOv5, deep convolution and residual connections are introduced to improve feature extraction capabilities. By adding the Feature Pyramid Network (FPN) structure, the model can extract image features from multiple scales, especially for small-scale targets (such as sparks at the beginning of a fire, tiny floating objects in a flood, cracks in a building, etc.), significantly improving the detection performance of the model.

[0026] In fire detection, the initial stage of a flame is often a small-scale fire point with unclear features. The FPN structure enhances the ability to extract multi-scale features, allowing the model to effectively capture these small targets. In addition, for flood scenes, the flood water level may appear as a narrow linear target in the image. The introduction of FPN can enhance the perception of these linear areas.

[0027] Traditional YOLOv5 uses predefined anchor boxes to locate the boundaries of target objects. In view of the diverse shapes and sizes of targets in disaster scenes, the improved algorithm adopts an adaptive anchor box mechanism. The K-Means clustering algorithm is used to analyze the target size distribution in disaster scenes and regenerate adaptive anchor boxes to better match the key target sizes in disaster scenes and avoid false detection and missed detection problems caused by too large or too small anchor boxes.

[0028] The introduction of the adaptive anchor box enables the model to more accurately align the boundaries of target objects when facing complex and diverse disaster scenarios. For example, in the scene of building collapse after an earthquake, the adaptive anchor box can accurately capture irregular shapes such as cracks and collapsed walls.

[0029] The weights of detection categories are adjusted according to the importance of different key targets in disaster scenarios. Key targets such as flames, flood water levels, and building cracks have higher detection priorities, while the weights of irrelevant objects in the background (such as road surfaces, trees, etc.) are relatively lower. By introducing a category weight mechanism, the algorithm pays more attention to key targets during training, thereby improving its detection confidence and positioning accuracy.

[0030] For example, in a fire detection scenario, the weights of flames and smoke are increased, allowing the model to more accurately and preferentially detect the source of the fire when encountering such high-risk targets. At the same time, in a flood scenario, increasing the weight of the flood water level helps to accurately identify the expansion boundary of the flood.

[0031] In view of the overlapping of multiple targets that may occur in disaster scenarios (such as the simultaneous existence of multiple fire sources, and the overlap of flood areas and obstacle areas), the non-maximum suppression (NMS) algorithm in YOLOv5 is improved. Based on the original NMS, a category sensitivity adjustment strategy is introduced to allow targets with different risk levels to exist at the same time without over-suppression. This improvement can prevent the model from missing high-risk targets due to over-suppression.

[0032] In flood scenarios, when the flooded area overlaps with nearby buildings, the optimized NMS algorithm can distinguish different categories, such as buildings and water bodies, while retaining the bounding boxes of these targets to ensure the integrity of the detection results.

[0033] Through the above improvements, the improved YOLOv5 algorithm not only improves the detection accuracy and speed of key targets in disaster scenes, but also ensures the recognition ability of small targets in complex scenes, meeting the real-time and accuracy requirements of the disaster warning system.

[0034] The U-Net image segmentation algorithm is used to perform pixel-by-pixel segmentation on the bounding box area output by the target detection to generate an accurate segmentation mask of the target.

[0035] In this step, the improved U-Net image segmentation algorithm is used to perform pixel-by-pixel precise segmentation processing on the bounding box area output by the target detection to generate the segmentation mask of the key targets in the disaster scene.

[0036] When processing full-size images, the original U-Net model may reduce the accuracy of segmentation of key targets due to interference from background information in the image. To improve the accuracy of segmentation, the improved U-Net adaptively adjusts the bounding box area output by YOLOv5, and only uses the region of interest (ROI) detected by the target as the input of U-Net, reducing the interference of irrelevant information caused by full image processing. This adaptive adjustment allows U-Net to focus on the fine segmentation of key targets by cropping the local area of ​​the target object in the image.

[0037] For example, in a fire scene, after YOLOv5 detects the bounding box area of ​​the flame, U-Net no longer processes the entire image, but only performs accurate pixel-by-pixel segmentation on the flame area to generate a segmentation mask for the flame. This can avoid the influence of background information (such as the sky and buildings) on the segmentation results and improve segmentation accuracy.

[0038] In order to enhance the ability to segment complex targets in disaster scenes, the improved U-Net introduces multi-scale feature fusion and attention mechanisms. In the encoder and decoder of U-Net, multi-level feature maps are cascaded to fuse image information of different scales. In particular, the attention mechanism is introduced during layer-by-layer upsampling in the decoding stage. This mechanism enables the model to focus on the key features of the target area in the disaster scene while ignoring the interference of irrelevant areas.

[0039] In flood scenes, the shape of the flood is complex and easily confused with the background ground. By introducing the attention mechanism, U-Net can more accurately capture the boundaries and changing shapes of the flood area, focus on the edge of the water flow, and ensure the accuracy and completeness of the segmentation results.

[0040] In order to solve the segmentation problem of complex structures in disaster scenes (such as collapsed buildings, cracks, etc.), the improved U-Net model introduces deep residual connections in the encoder part to avoid the gradient vanishing problem that may occur in deep convolutional networks. In addition, by introducing dilated convolution in the convolution layer of U-Net, the model can effectively increase the receptive field, while maintaining the resolution, capturing the global features of a large range of targets.

[0041] In post-earthquake building scenes, collapsed buildings usually have complex edges and structures. Through the combination of residual connections and dilated convolutions, U-Net can effectively capture the subtle structure of the building during segmentation while maintaining an overall understanding of the large-scale collapsed area and generating accurate segmentation masks for the collapsed area of ​​the building.

[0042] Targets in disaster scenes usually have irregular shapes and boundaries, and the contrast between the target and the background is not obvious, which can easily cause segmentation errors. To solve this problem, the improved U-Net adopts an optimization strategy based on a hybrid loss function, combining the cross-entropy loss and the Dice coefficient loss. The cross-entropy loss is used for overall pixel classification, while the Dice coefficient loss focuses on improving the segmentation accuracy of the target boundary, which is particularly suitable for the segmentation of small targets in unbalanced data.

[0043] In fire scenes, the flames have irregular shapes and blurred edges. Through the optimized loss function, the model can better distinguish the flames from the surrounding background, significantly improve the segmentation accuracy of the flame boundaries, and generate a fine flame segmentation mask.

[0044] Through the above-mentioned improvements, the U-Net model significantly improves the segmentation accuracy and processing efficiency of targets in disaster scenarios, ensuring that the disaster warning system can quickly generate accurate segmentation masks and provide precise data support for disaster monitoring, assessment and emergency response.

[0045] The image segmentation mask, historical disaster data, meteorological data, and geographic information system data are input into the LSTM prediction model to perform multi-source data fusion and time series analysis, and output dynamic prediction information and warning levels of disasters.

[0046] In this step, the image segmentation mask and historical disaster data, meteorological data, and geographic information system (GIS) data are input into the long short-term memory network (LSTM) prediction model to perform multi-source data fusion and time series analysis, ultimately achieving dynamic disaster prediction and warning level assessment. To ensure the accuracy and real-time nature of the analysis, the present invention uses an improved LSTM architecture to dynamically predict the development trend and potential threats of disasters by combining multi-source information of disaster scenes. The specific implementation steps are as follows: Image segmentation masks, historical disaster data, meteorological data, and geographic information system (GIS) data are standardized and time-aligned. Various data sources have different structures and sampling frequencies, and they need to be preprocessed first to ensure effective fusion.

[0047] Image segmentation mask: Extract the spatial location and shape information of the disaster target from the segmentation mask output by U-Net. Each mask represents the boundary of a key target (such as a flame, a water level, or a collapsed building). The mask data is discretized into a feature vector to extract spatial information such as area, boundary length, and shape change.

[0048] Historical disaster data: includes historical records of similar disasters, such as the occurrence, extension, and evolution of previous fires and floods. Historical data is arranged in time series and standardized to align with other data sources.

[0049] Meteorological data: Meteorological data mainly includes real-time wind speed, wind direction, temperature, humidity, rainfall, etc. This type of data has a time series nature and affects the development trend of disasters, and needs to be associated with the current disaster status.

[0050] Geographic Information System (GIS) data: GIS data provides geographical information of disaster areas, such as terrain, height, river distribution, building distribution, etc. GIS data is usually two-dimensional or three-dimensional spatial data and needs to be converted into a vector form that is compatible with LSTM model input. Features such as terrain slope and building density are extracted to assist disaster prediction.

[0051] Normalize all types of input data to ensure that all data feature values ​​fall into the same range (such as [0, 1]) to avoid data of different magnitudes affecting model training.

[0052] Use time windows to slice the data, split historical meteorological data, disaster records, etc. into the same time segments, and align them with the time segments corresponding to the image segmentation mask.

[0053] The image segmentation mask is fused with other data sources through feature concatenation to generate a time series feature matrix. The input of each time step includes features from multiple different sources, which are used together to represent the state of the disaster at a certain point in time.

[0054] The features extracted from each data source are combined into a complete input vector using feature concatenation. For each time step, the image segmentation mask, meteorological data, historical disaster data, and GIS data are mapped to features of the same dimension and merged through matrix concatenation.

[0055] In this way, the model is able to consider information from multiple sources simultaneously and make predictions in the context of the time series.

[0056] The LSTM network receives the input feature vector at each time step and processes it through multiple layers of LSTM units. At each time step, the LSTM unit updates the current hidden state based on the current input and the previous hidden state.

[0057] LSTM can remember the long-term dynamic changes of disasters, such as the gradual rise of flood levels and the spread of fires. The model predicts the future evolution trend of disasters by gradually processing the input feature sequence.

[0058] The model is stacked with multiple layers of LSTM units to enhance its learning ability for complex time series. The first layer of LSTM receives the input feature vector and outputs a hidden state, which is passed to the next layer of LSTM and processed in turn.

[0059] The input data is trained by sliding the time window, and the model is trained using a supervised learning method. The loss function can use the mean square error (MSE) to measure the difference between the future disaster state predicted by the model and the actual state.

[0060] The output of the LSTM model represents the dynamic development trend of the disaster. The model outputs a vector at each time step, indicating the changing trend of the disaster in the next few time steps (such as the speed of fire spread, flood level changes, etc.), and combines historical disaster data and real-time input to generate a warning level.

[0061] Disaster expansion prediction: The output vector of LSTM is used to predict the expansion range of future disasters. The output vector can represent the spread speed of fire, the rising rate of water level, or the damage process of buildings.

[0062] Warning level assessment: Combine the predicted disaster development trend with the risk model to determine the current warning level. The warning level division is based on a series of thresholds. For example, when the fire spreads to a certain area or the flood level rises above a certain height, different levels of warnings (such as yellow warning, red warning, etc.) will be triggered.

[0063] The disaster change trend vector output by the LSTM model is used to evaluate the warning level according to the predetermined risk threshold. For example, if the LSTM predicts that the flood level will rise by more than 0.5 meters in the next 30 minutes, a red warning is triggered and a warning message is sent to the relevant departments.

[0064] The output of the model not only includes the development trend of the disaster, but also provides a dynamic prediction map of the disaster expansion in combination with the time dimension to assist decision-making.

[0065] For example, in a fire scene, the flame segmentation mask generated by U-Net is processed by LSTM and combined with current meteorological data such as wind speed and temperature. LSTM predicts that the flame will spread 100 meters to the east in the next 15 minutes, and based on historical fire records and terrain data, a red warning level is given, indicating that it may spread to nearby residential areas.

[0066] On the other hand, the present invention also provides a disaster warning system based on key target detection and image segmentation, comprising: An acquisition module is used to acquire real-time image data on site and perform preprocessing operations on the image; The detection module is used to process the preprocessed images using the YOLOv5 target detection algorithm, detect key targets at the disaster site, and output the target's category label and its bounding box; The segmentation module is used to perform pixel-by-pixel segmentation processing on the bounding box area output by target detection using the U-Net image segmentation algorithm to generate an accurate segmentation mask of the target; The prediction module is used to input the image segmentation mask and historical disaster data, meteorological data, and geographic information system data into the LSTM prediction model, perform multi-source data fusion and time series analysis, and output dynamic prediction information and warning levels of disasters.

[0067] The prior art mentioned in the above background technology section and specific embodiment section of the present invention can be used as part of the present invention to understand the meaning of some technical features or parameters.

Claims

1. A disaster warning method based on key target detection and image segmentation, characterized in that: The method comprises the following steps: Acquire real-time image data on site and perform pre-processing operations on the images; The preprocessed images are processed using the YOLOv5 target detection algorithm to detect key targets at the disaster site and output the target’s category label and its bounding box. Use the U-Net image segmentation algorithm to perform pixel-by-pixel segmentation on the bounding box area output by the target detection to generate an accurate segmentation mask of the target; The image segmentation mask, historical disaster data, meteorological data, and geographic information system data are input into the LSTM prediction model to perform multi-source data fusion and time series analysis, and output dynamic prediction information and warning levels of disasters.

2. The disaster warning method based on key target detection and image segmentation according to claim 1 is characterized in that: In the basic network of YOLOv5, deep convolution and residual connection are introduced to improve the feature extraction capability, and a feature pyramid network structure is added.

3. The disaster warning method based on key target detection and image segmentation according to claim 1 is characterized in that: An adaptive Anchor frame is introduced into the YOLOv5, and the target size distribution in the disaster scene is analyzed by the K-Means clustering algorithm, and the adaptive Anchor frame is regenerated to better match the key target size in the disaster scene, avoiding false detection and missed detection problems caused by the Anchor frame being too large or too small.

4. The disaster warning method based on key target detection and image segmentation according to claim 1 is characterized in that: In the U-Net, the bounding box area output by YOLOv5 is adaptively adjusted, and only the region of interest detected by the target is used as the input of the U-Net to reduce the interference of irrelevant information caused by the full image processing.

5. The disaster warning method based on key target detection and image segmentation according to claim 1 is characterized in that: The U-Net introduces deep residual connections in the encoder part and introduces hole convolution in the convolution layer of the U-Net.

6. A disaster warning system based on key target detection and image segmentation, characterized in that: The system includes the following modules: An acquisition module is used to acquire real-time image data on site and perform preprocessing operations on the image; The detection module is used to process the preprocessed images using the YOLOv5 target detection algorithm, detect key targets at the disaster site, and output the target's category label and its bounding box; The segmentation module is used to perform pixel-by-pixel segmentation processing on the bounding box area output by target detection using the U-Net image segmentation algorithm to generate an accurate segmentation mask of the target; The prediction module is used to input the image segmentation mask and historical disaster data, meteorological data, and geographic information system data into the LSTM prediction model, perform multi-source data fusion and time series analysis, and output dynamic prediction information and warning levels of disasters.

7. The disaster warning system based on key target detection and image segmentation according to claim 6, characterized in that: In the basic network of YOLOv5, deep convolution and residual connection are introduced to improve the feature extraction capability, and a feature pyramid network structure is added.

8. The disaster warning system based on key target detection and image segmentation according to claim 6, characterized in that: An adaptive Anchor frame is introduced into the YOLOv5, and the target size distribution in the disaster scene is analyzed by the K-Means clustering algorithm, and the adaptive Anchor frame is regenerated to better match the key target size in the disaster scene, avoiding false detection and missed detection problems caused by the Anchor frame being too large or too small.

9. The disaster warning system based on key target detection and image segmentation according to claim 6, characterized in that: In the U-Net, the bounding box area output by YOLOv5 is adaptively adjusted, and only the region of interest detected by the target is used as the input of the U-Net to reduce the interference of irrelevant information caused by the full image processing.

10. The disaster warning system based on key target detection and image segmentation according to claim 6, characterized in that: The U-Net introduces deep residual connections in the encoder part and introduces hole convolution in the convolution layer of the U-Net.

Images