An unmanned aerial vehicle image boundary segmentation method, system and storage medium

Abstract

The embodiment of the present application relates to the technical field of computer vision and image processing, and discloses a kind of unmanned aerial vehicle image boundary segmentation method, system and storage medium, method includes: according to the local gradient amplitude and texture complexity of each pixel position in unmanned aerial vehicle disaster image, the structure element of corresponding pixel position is generated, the generated structure element is used to carry out open operation to image, and edge enhancement image is generated;Extract the morphological features of edge enhancement image;Extract the spectral features of unmanned aerial vehicle disaster image;Spectral features and morphological features are fused by gate attention mechanism, so that the disaster boundary segmentation map finally generated can accurately locate and segment disaster boundary.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and image processing, and particularly relates to a UAV image boundary segmentation method and system and a storage medium. BACKGROUND

[0002] Deep learning techniques, particularly convolutional neural networks based on encoder-decoder architecture such as U-Net, have been widely applied in image semantic segmentation tasks. In order to improve the segmentation accuracy, existing technologies often introduce attention mechanisms or multi-scale feature fusion modules to enhance the model's ability to capture key features. In addition, when dealing with images with clear boundaries, some methods combine mathematical morphology operations (such as erosion, dilation) as preprocessing or postprocessing steps to enhance or modify the edges of the image.

[0003] However, the existing technical solutions still have the following shortcomings: first, the mathematical morphology operations used usually use fixed-size and shape structural elements, which are difficult to adapt to local feature changes in different regions of the image. When dealing with images with complex textures and various boundary shapes, the enhancement effect is limited, and even useful boundary details may be destroyed. Second, existing network models often use single-path encoder structures to mix spectral information (such as color, brightness) and geometric structure information (such as edges, textures) together for feature extraction, which may cause interference between features of different properties, limiting the ability to accurately identify boundaries. Third, commonly used loss functions such as cross-entropy loss or Dice loss, while optimizing the accuracy of the overall segmentation area, are insufficiently penalized for misclassification of boundary pixels with a very small area ratio, resulting in the trained model producing fuzzy, broken, or discontinuous segmentation results in the boundary area. SUMMARY

[0004] Therefore, it is necessary to propose a UAV image boundary segmentation method, system and storage medium to solve the above problems.

[0005] A UAV image boundary segmentation method, the method comprising:

[0006] Generating a structural element corresponding to each pixel position in the UAV disaster image based on the local gradient amplitude and texture complexity of each pixel position, performing an opening operation on the image using the generated structural element, and generating an edge-enhanced image.

[0007] Extracting morphological features of the edge-enhanced image.

[0008] Extracting spectral features of the UAV disaster image.

[0009] Fusing the spectral features and the morphological features through a gated attention mechanism to obtain a disaster boundary segmentation map.

[0010] According to the local gradient amplitude and the texture complexity of each pixel position in the unmanned aerial vehicle disaster image, a structure element corresponding to the pixel position is generated, and the structure element comprises the following steps:

[0011] According to Determine the local gradient amplitude of each pixel position in the image, wherein, is a horizontal gradient component, is a vertical gradient component, is a local gradient amplitude, is a pixel position.

[0012] According to Determine the texture complexity of each pixel position in the image, wherein, is the probability of the intensity value in the neighborhood, is the texture complexity.

[0013] According to Generate a radius corresponding to the pixel position, wherein, is the radius of the pixel position, is the local gradient amplitude, is the texture complexity, is a learnable parameter of the local gradient amplitude, is a learnable parameter of the texture complexity, is a rounding up.

[0014] Generate a structure element corresponding to the pixel position by the radius.

[0015] According to the generated structure element, an opening operation is performed on the image to generate an edge enhanced image, and the edge enhanced image comprises the following steps:

[0016] According to Perform an opening operation on the image to generate an edge enhanced image, wherein, is an edge enhanced image, is an unmanned aerial vehicle disaster image, is a structure element, is an opening operation.

[0017] According to the generated structure element, an opening operation is performed on the image to generate an edge enhanced image, and the edge enhanced image comprises the following steps:

[0018] Fuse the spectral feature and the morphological feature through the gated attention mechanism to obtain a disaster boundary segmentation map.

[0019] Upsample the fusion feature to obtain an intermediate feature.

[0020] The intermediate feature is subjected to a 1*1 convolution operation and a softmax activation function to generate a disaster boundary segmentation map.

[0021] The fusion feature is obtained by fusing the spectral feature and the morphological feature through the gated attention mechanism, and specifically includes:

[0022] According to The fusion feature is obtained by fusing the spectral feature and the morphological feature, and specifically includes: The fusion feature is obtained by fusing the spectral feature and the morphological feature, The spectral feature is obtained by fusing the spectral feature and the morphological feature, The morphological feature is obtained by fusing the spectral feature and the morphological feature, The weight matrix of the spectral feature is obtained by fusing the spectral feature and the morphological feature, The weight matrix of the morphological feature is obtained by fusing the spectral feature and the morphological feature, The sigmoid activation function is obtained by fusing the spectral feature and the morphological feature, The element-wise multiplication is obtained by fusing the spectral feature and the morphological feature.

[0023] After the disaster boundary segmentation map is obtained by fusing the spectral feature and the morphological feature through the gated attention mechanism, the method further includes:

[0024] The disaster boundary segmentation map is subjected to scale adaptive post-processing through morphological closing operation to obtain a final disaster boundary fine segmentation map.

[0025] An unmanned aerial vehicle image boundary segmentation system, the system comprising:

[0026] A morphological image preprocessing module is configured to generate a structural element corresponding to each pixel position in the unmanned aerial vehicle disaster image according to the local gradient amplitude and texture complexity of each pixel position, and perform an opening operation on the image using the generated structural element to generate an edge enhanced image.

[0027] A spectral path encoder is configured to extract a spectral feature of the unmanned aerial vehicle disaster image.

[0028] A morphological path encoder is configured to extract a morphological feature of the edge enhanced image.

[0029] A decoder is configured to fuse the spectral feature and the morphological feature through a gated attention mechanism to obtain a disaster boundary segmentation map.

[0030] The system further comprises:

[0031] A post-processing module is configured to perform scale adaptive post-processing on the disaster boundary segmentation map through morphological closing operation to obtain a final disaster boundary fine segmentation map.

[0032] The spectral path encoder comprises five down-sampling modules, each of which is composed of two 3*3 convolution layers, a batch normalization layer, a ReLU activation function layer and a 2*2 maximum pooling layer connected in sequence; and the morphological path encoder comprises three down-sampling stages, and the structure of the down-sampling stages is consistent with that of the down-sampling modules of the spectral path encoder.

[0033] The morphological image preprocessing module, the spectral path encoder module, the morphological path encoder and the decoder are trained and optimized according to a combined loss function, and the combined loss function is

[0034] , wherein, is a combined loss function, is a real disaster boundary segmentation map, is a predicted disaster boundary segmentation map, is a spatial gradient, is an indicator function, is a weight coefficient of a cross-entropy loss term, is a cross-entropy loss term, is a weight coefficient of an edge-aware loss term, is a pixel position, is a Euclidean norm.

[0035] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the method.

[0036] By adopting the embodiment of the present application, the following beneficial effects are achieved:

[0037] The present application first generates an adaptive structure element according to the local gradient amplitude and texture complexity of each pixel point, effectively suppresses noise and enhances the real boundary. The morphological opening operation is performed by using the structure element, different scale features are intelligently processed, and the global optimal edge enhancement effect is achieved. Further, the spectral features of the unmanned aerial vehicle disaster image and the morphological features of the edge enhanced image are extracted respectively, the spectral features and the morphological features are dynamically balanced through the gated attention mechanism fusion module, the boundary is sharpened while the regional consistency is maintained, so that the finally generated disaster boundary segmentation map can accurately locate and segment the disaster boundary.

[0038] The present application comprehensively utilizes adaptive morphological enhancement and deep learning to extract multi-size feature information, ensures accurate identification and segmentation of complex ground object information and real disaster boundary in the image, and significantly improves the accuracy and robustness of image boundary segmentation. BRIEF DESCRIPTION OF DRAWINGS

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] in:

[0041] Figure 1 This is a flowchart illustrating an embodiment of a UAV image boundary segmentation method provided by the present invention.

[0042] Figure 2 This is a flowchart illustrating another embodiment of the UAV image boundary segmentation method provided by the present invention;

[0043] Figure 3 This is a schematic diagram of an embodiment of an UAV image boundary segmentation system provided by the present invention;

[0044] Figure 4 A schematic diagram of the structure of an embodiment of the storage medium provided by the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] like Figure 1 As shown, Figure 1 This is a flowchart illustrating an embodiment of a UAV image boundary segmentation method provided by the present invention. The UAV image boundary segmentation method includes:

[0047] S101: Based on the local gradient magnitude and texture complexity of each pixel in the UAV disaster image, generate the structuring element for the corresponding pixel position, and use the generated structuring element to perform an opening operation on the image to generate an edge-enhanced image.

[0048] For example, acquire a drone disaster image to be processed. Determine the local gradient magnitude at each pixel location in the image according to the following formula:

[0049] ;

[0050] in, For horizontal gradient components, For vertical gradient components, is the local gradient magnitude, is the pixel position. The gradient magnitude can effectively reflect the degree of change of image gray or color, and usually takes a larger value at the edge of an object.

[0051] The texture complexity of each pixel position in the image is determined according to the formula as follows:

[0052] ;

[0053] wherein, is the probability of the intensity value in the neighborhood, is the texture complexity. The higher the texture complexity, the more complex the texture of the region and the greater the amount of information.

[0054] Further, according to the local gradient magnitude and the texture complexity of each pixel position in the UAV disaster image, a radius corresponding to the pixel position is generated, as shown in the following formula:

[0055] ;

[0056] wherein, is the radius of the pixel position, is a learnable parameter of the local gradient magnitude, is a learnable parameter of the texture complexity, used to adjust the contribution proportion of gradient information and texture information, is the upward rounding.

[0057] Further, a dynamic structure element is generated for the current pixel position based on the calculated dynamic radius. An opening operation is performed on the image using the generated structure element to generate an edge-enhanced image, and the edge-enhanced image is as shown in the following formula:

[0058] ;

[0059] wherein, is the edge-enhanced image, is the UAV disaster image, is the structure element, is the opening operation.

[0060] It should be noted that the opening operation, i.e., performing a corrosion operation first and then performing a dilation operation, can disconnect fine connections, eliminate small bright spots, and smooth the outline of an object. Since the structure element is adaptive, this operation can intelligently process image features of different scales, thereby achieving the optimal edge enhancement effect in the global range. After all pixel positions are processed, a complete edge-enhanced image is obtained.

[0061] And, the adaptive morphology proposed by the present application is different from the traditional morphology method. The network autonomously learns the optimal scaling factor in the training process and rather than relying on the empirically set numerical value. The gradient term focuses on optimizing the edge region, while the texture component can dynamically adjust the processing strength according to the characteristics of the heterogeneous region such as vegetation or debris region.

[0062] S102: Extracting morphological features of the edge-enhanced image.

[0063] Exemplarily, the morphological path processes the edge-enhanced image subjected to adaptive enhancement through a U-Net encoder architecture containing three down-sampling modules. Each module is composed of two 3x3 convolution layers, adopts batch normalization and ReLU activation function, and is connected to two 2x2 max-pooling layers.

[0064] S103: Extracting spectral features of the UAV disaster image.

[0065] Exemplarily, the UAV disaster image is processed through a classic U-Net encoder architecture containing five down-sampling modules. Each module is composed of two 3x3 convolution layers, adopts batch normalization and ReLU activation function, and is connected to two 2x2 max-pooling layers.

[0066] S104: Fusing the spectral features and the morphological features through a gated attention mechanism to obtain a disaster boundary segmentation map.

[0067] Exemplarily, the spectral features and the morphological features are fused according to the formula shown below to obtain fusion features:

[0068] ;

[0069] wherein, is the fusion feature, is the spectral feature, is the morphological feature, is the weight matrix of the spectral feature, is the weight matrix of the morphological feature, is a sigmoid activation function, is an element-wise multiplication.

[0070] Through this gating mechanism, the network can intelligently and dynamically weigh and selectively fuse the two features, thereby retaining the regional consistency provided by the spectral information while fully utilizing the morphological information to sharpen the boundaries.

[0071] Further, the fused features are up-sampled to obtain intermediate features; and the intermediate features are subjected to 1*1 convolution operation and softmax activation function processing to generate a disaster boundary segmentation map.

[0072] As can be seen from the above description, the present application first generates a self-adaptive structure element according to the local gradient amplitude and texture complexity of each pixel point, effectively suppresses noise and enhances the real boundary. The morphological opening operation is performed by using the structure element, different scale features are intelligently processed, and the global optimal edge enhancement effect is realized. Further, the spectral features of the unmanned aerial vehicle disaster image and the morphological features of the edge enhanced image are extracted respectively, the spectral features and the morphological features are dynamically balanced by the gating attention mechanism fusion module, the boundary is sharpened while the region consistency is maintained, so that the finally generated disaster boundary segmentation map can accurately locate and segment the disaster boundary.

[0073] The present application comprehensively utilizes adaptive morphological enhancement and deep learning to extract multi-size feature information, ensures accurate identification and segmentation of complex ground information and real disaster boundary in the image, and significantly improves the accuracy and robustness of image boundary segmentation.

[0074] As shown in Figure 2 , Figure 2 is a flowchart of another embodiment of the unmanned aerial vehicle image boundary segmentation method provided by the present application. The unmanned aerial vehicle image boundary segmentation method comprises the following steps:

[0075] S201: According to the local gradient amplitude and texture complexity of each pixel position in the unmanned aerial vehicle disaster image, a structure element corresponding to the pixel position is generated, and the generated structure element is used to perform opening operation on the image to generate an edge enhanced image.

[0076] Exemplarily, a to-be-processed unmanned aerial vehicle disaster image is obtained. The image can be a high-resolution aerial photograph image containing complex ground information and a disaster boundary to be accurately identified. For each pixel position in the unmanned aerial vehicle disaster image, a structure element used for morphological operation is dynamically and independently generated according to the local characteristics of the original image. This adaptability enables the morphological operation to better match the local changes of the image content, thereby effectively suppressing noise and maximizing the enhancement of the real target boundary, providing high-quality input for subsequent feature extraction.

[0077] The calculation of two kinds of local image features is performed for any pixel point in the image. The first kind is gradient calculation. Specifically, the local gradient amplitude of each pixel position in the image is determined according to the formula as shown below:

[0078] ;

[0079] Among them, is a horizontal gradient component, is the vertical gradient component, is the local gradient magnitude, is the pixel position. The gradient magnitude can effectively reflect the degree of change of image gray or color, usually taking a larger value at the edge of the object.

[0080] The second is the entropy value calculation, which is used to quantify the texture complexity of the local region. Specifically, the texture complexity of each pixel position in the image is determined according to the formula as follows:

[0081] ;

[0082] wherein, is the probability of the intensity value in the neighborhood, is the texture complexity. The higher the texture complexity, the more complex the texture of the region and the more information.

[0083] Further, based on the ceiling operation, according to the local gradient magnitude and the texture complexity of each pixel position in the unmanned aerial vehicle disaster image, the radius of the corresponding pixel position is generated, as shown in the following formula:

[0084] ;

[0085] wherein, is the radius of the pixel position, is the local gradient magnitude, is the texture complexity, is a learnable parameter of the local gradient magnitude, is a learnable parameter of the texture complexity, used to adjust the contribution ratio of gradient information and texture information, is the upward rounding.

[0086] Further, based on the calculated dynamic radius, a dynamic structure element is generated for the current pixel position. The shape of the structure element can be various, such as circular, diamond or the preferred elliptical shape in this embodiment, and the size is determined by the radius. For example, in a flat area, the gradient and entropy value are low, and the generated structure element size is small, which can effectively filter out noise; while in the contour and complex texture area, the gradient and entropy value are high, and the generated structure element size is large, which can better connect the broken edges and preserve the structural details.

[0087] Further, the generated structure element is used to perform an opening operation on the image to generate an edge enhanced image. The ceiling operation ensures that the generated integer radius is suitable for discrete morphological operation, and the structure element generated thereby can guide adaptive opening operation to suppress noise while preserving edge features, and the edge enhanced image is as follows:

[0088] ;

[0089] wherein, is an edge enhanced image, is a UAV disaster image, is a structure element, is an opening operation.

[0090] S202: morphological features of the edge enhanced image are extracted.

[0091] S203: spectral features of the UAV disaster image are extracted.

[0092] It should be noted that steps S202-S203 have been discussed in detail in the implementation scenario shown in FIG. 2, and will not be repeated here. Figure 1

[0093] S204: the spectral features and the morphological features are fused through a gated attention mechanism to obtain fused features.

[0094] Exemplarily, the spectral features and the morphological features are input into a gated attention fusion module, and the spectral features and the morphological features are fused through a gated attention mechanism to obtain fused features. Specifically, the spectral features and the morphological features are fused according to the formula shown below to obtain the fused features:

[0095] ;

[0096] wherein, is the fused feature, is the spectral feature, is the morphological feature, is a weight matrix of the spectral feature, is a weight matrix of the morphological feature, is a sigmoid activation function, is an element-wise multiplication.

[0097] Through this gating mechanism, the network can intelligently and dynamically weigh and selectively fuse the two features, thereby retaining the regional consistency provided by the spectral information while fully utilizing the morphological information to sharpen the boundaries.

[0098] S205: an up-sampling operation is performed on the fused features to obtain intermediate features.

[0099] S206: a 1x1 convolution operation and a softmax activation function processing are performed on the intermediate features to generate a disaster boundary segmentation map.

[0100] ​For example, the fused features are upsampled to obtain intermediate features. After processing the intermediate features with a 1×1 convolution and a softmax activation function, an initial disaster boundary segmentation map is generated.

[0101] S207: The disaster boundary segmentation map is post-processed by scaling adaptation through morphological closure operation to obtain the final detailed disaster boundary segmentation map.

[0102] For example, the disaster boundary segmentation map is optimized through morphological closure operations:

[0103] ;

[0104] in, A detailed map of the disaster boundary. This is a disaster boundary segmentation map. As a structural element, This is a closure operation.

[0105] Among them, for each connection area structural elements Proportional to the perimeter / area ratio:

[0106] ;

[0107] in, This is a proportionality coefficient used to adjust structural components. The size controls the strength of the closure operation. Let R be the perimeter of the connected region, reflecting the complexity of the boundary morphology. The area of ​​the connected region R represents the spatial extent of the region. This is for rounding up.

[0108] This adaptive closure operation can fill small gaps in elongated regions while maintaining a compact feature shape. (Scale factor) During training, it learns together with other network parameters and can automatically adapt to different disaster scenarios.

[0109] like Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the structure of an embodiment of a UAV image boundary segmentation system provided by the present invention. The UAV image boundary segmentation system 10 is an AMU-Net network (Adaptive Morphological U-Net), and the system includes:

[0110] The morphological image preprocessing module 11 is configured to generate a structuring element for each pixel position in the UAV disaster image according to a local gradient magnitude and a texture complexity of the pixel position, and perform an opening operation on the image using the generated structuring element to generate an edge-enhanced image.

[0111] For example, in the morphological image preprocessing module 11, a local gradient magnitude of each pixel position in the image is first determined according to the following formula:

[0112] ;

[0113] wherein, is a horizontal gradient component, is a vertical gradient component, is the local gradient magnitude, is the pixel position;

[0114] Further, a texture complexity of each pixel position in the image is determined according to the following formula:

[0115] ;

[0116] wherein, is a probability of an intensity value in a neighborhood, is the texture complexity;

[0117] Further, a radius of the corresponding pixel position is generated according to the following formula:

[0118] ;

[0119] wherein, is the radius of the pixel position, is the local gradient magnitude, is the texture complexity, is a learnable parameter of the local gradient magnitude, is a learnable parameter of the texture complexity, used to adjust a contribution ratio of gradient information and texture information, is a ceiling function.

[0120] Finally, a structuring element for the corresponding pixel position is generated by the radius.

[0121] The spectral path encoder 12 is configured to extract spectral features of the UAV disaster image.

[0122] The morphological path encoder 13 is configured to extract morphological features of the edge-enhanced image.

[0123] Exemplarily, the spectral path encoder comprises five down-sampling modules, each of which is sequentially connected by two 3x3 convolution layers, a batch normalization layer, a ReLU activation function layer, and a 2x2 max-pooling layer. The spectral features of the UAV disaster image are extracted by the spectral path encoder 12; the morphological path encoder comprises three down-sampling stages, and the structure thereof is consistent with the down-sampling module of the spectral path encoder. The morphological features of the edge-enhanced image are extracted by the morphological path encoder 13.

[0124] The decoder 14 is configured to fuse the spectral features and the morphological features by a gated attention mechanism to obtain a disaster boundary segmentation map.

[0125] Exemplarily, by setting a skip connection between the encoder path and the corresponding decoding level, the system can retain high-resolution boundary information. The AMU-Net uses a cross-channel attention gate mechanism to preferentially extract spatially aligned edge features. Finally, the decoder 14 outputs a disaster boundary segmentation map after being processed by a 1x1 convolution layer and being activated by a softmax function.

[0126] The UAV image boundary segmentation system composed of the morphological image preprocessing module, the spectral path encoder module, the morphological path encoder, and the decoder is trained and optimized by using a combined loss function combining a standard cross-entropy and an edge-aware term, and the combined loss function is as follows:

[0127] ;

[0128] wherein, is the combined loss function, is a real disaster boundary segmentation map, is a predicted disaster boundary segmentation map, is a spatial gradient, is an indicator function, is a weight coefficient of a cross-entropy loss term, is the cross-entropy loss term, is a weight coefficient of an edge-aware loss term. The edge term imposes a more stringent penalty on the misclassification of boundary pixels, while and balance the contributions of each term, is a pixel position, is a Euclidean norm.

[0129] A UAV image boundary segmentation system 10 further comprises:

[0130] A post-processing module (not shown in the figure) is configured to perform scale-adaptive post-processing on the disaster boundary segmentation map by a morphological closing operation to obtain a final disaster boundary fine segmentation map.

[0131] Exemplarily, the disaster boundary segmentation graph output by the post-training system is subjected to scale self-adaptive post-processing through morphological closing operation to obtain a final disaster boundary fine segmentation graph; the scale self-adaptive post-processing specifically refers to filling gaps in the segmentation result through morphological closing operation. The scale self-adaptive post-processing improves the spatial consistency of the result, so that the model output is closer to the real boundary shape, and meets the GIS (Geographic Information System) integration and risk quantification requirements.

[0132] As can be known from the above description, the unmanned aerial vehicle image boundary segmentation system provided by the application realizes end-to-end processing of the input image, and the gradient is back-propagated through each component of the neural network and the differentiable approximate calculation of the morphological operation. This joint optimization mechanism enables the AMU-Net to regard the morphological processing as an independent preprocessing or post-processing step. This integrated design enables the network to learn morphological parameters that complement feature representation, thereby achieving optimal performance in the specific task of detecting geological change boundaries.

[0133] As Figure 4 shown, Figure 4 is a structural schematic diagram of an embodiment of the storage medium provided by the application. The storage medium 20 stores at least one computer program 21, and the computer program 21 is executed by a processor to implement the method as shown in Figure 1 and Figure 2 . For details, refer to the above description, which will not be repeated here. In an embodiment, the storage medium 20 can be a storage chip, a hard disk or a mobile hard disk or an optical disc, or other readable and writable storage tools, and can also be a server, etc.

[0134] In addition, the processes depicted in the drawings do not necessarily have to be implemented in the specific order shown or in a continuous order to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0135] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the device, equipment and non-volatile computer readable storage medium embodiments, since they are basically similar to the method embodiments, they are described more simply, and the relevant parts can be referred to the part of the method embodiment.

[0136] The apparatus, device, non-volatile computer-readable storage medium and method provided in the embodiments of this specification are corresponding. Therefore, the apparatus, device and non-volatile computer storage medium also have similar beneficial technical effects as the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the corresponding apparatus, device and non-volatile computer storage medium will not be repeated here.

[0137] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0138] For ease of description, the above apparatus is described by dividing it into various functional units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components. Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0140] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1one or more processes and / or blocks Figure 1 the function specified in the one or more blocks

[0141] These computer program instructions can also be loaded into computer or other programmable data processing devices to cause a series of operational steps to be performed on the computer or other programmable devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable devices provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more processes and / or blocks Figure 1 the function specified in the one or more blocks

[0142] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0143] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) about which the computer stores information about an operating system, application software, and / or the like. Memory is an example of computer readable media.

[0144] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0145] It should also be noted that the terms "comprising", "comprises", "including", "includes" or any other variation thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not include only those elements recited, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article or apparatus that includes the recited element.

[0146] The specification can be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules can be located in local and remote computer storage media including memory storage devices.

[0147] The various embodiments in the specification are described in progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, the system embodiments are described simply because they are basically similar to the method embodiments, and the relevant parts can be referred to the description of the method embodiments.

[0148] The above only describes the preferred embodiments of the present application, and of course cannot limit the scope of the present application. Any equivalent changes made according to the claims of the present application are still within the scope of the present application.

Claims

1. A method for unmanned aerial vehicle image boundary segmentation, characterized in that, The method comprises: According to the local gradient amplitude and the texture complexity of each pixel position in the unmanned aerial vehicle disaster image, a structure element corresponding to the pixel position is generated, and an open operation is performed on the image by using the generated structure element to generate an edge enhanced image. determining the local gradient amplitude of each pixel position in the image, wherein, is a horizontal gradient component, is a vertical gradient component, is a local gradient amplitude, is a pixel position; and determining the texture complexity of each pixel position in the image, wherein, is the probability of the intensity value in the neighborhood, is a texture complexity; and generating a radius corresponding to the pixel position, wherein, is the radius of the pixel position, is a local gradient amplitude, is a texture complexity, is a learnable parameter of the local gradient amplitude, is a learnable parameter of the texture complexity, is a rounding up; and generating a structure element corresponding to the pixel position by using the radius. extracting morphological features of the edge enhanced image; extracting spectral features of the unmanned aerial vehicle disaster image; fusing the spectral features and the morphological features through a gated attention mechanism to obtain a disaster boundary segmentation map, specifically comprising: 2.The method of claim 1, wherein, the open operation of the generated structural elements on the image to generate an edge enhanced image, specifically comprising: According to performing an opening operation on the image to generate an edge-enhanced image, wherein the edge-enhanced image, the UAV disaster image, the structural element, the opening operation. 3.The method of claim 1, wherein, the fusion of the spectral features and the morphological features through a gated attention mechanism to obtain fusion features, specifically comprising: According to fusing the spectral features and the morphological features, obtaining fused features, wherein, the fused features, the spectral features, the morphological features, a weight matrix of the spectral features, a weight matrix of the morphological features, sigmoid activation function, element-wise multiplication. 4.The method of claim 1, wherein, after the fusion of the spectral features and the morphological features through a gated attention mechanism to obtain a disaster boundary segmentation map, it still specifically comprises: performing scale adaptive post-processing on the disaster boundary segmentation map through morphological closing operation to obtain a final disaster boundary fine segmentation map. 5.A system for UAV image boundary segmentation, the system comprising: The system comprises: a morphological image preprocessing module for generating a structural element corresponding to each pixel position in the unmanned aerial vehicle disaster image according to the local gradient amplitude and texture complexity of each pixel position, performing an open operation on the image using the generated structural element to generate an edge enhanced image; a spectral path encoder for extracting spectral features of the unmanned aerial vehicle disaster image, the spectral path encoder comprising five down-sampling modules, each down-sampling module being composed of two 3x3 convolution layers, a batch normalization layer, a ReLU activation function layer and a 2x2 maximum pooling layer connected in sequence; a morphological path encoder for extracting morphological features of the edge enhanced image, the morphological path encoder comprising three down-sampling stages, the structure of which is consistent with the down-sampling module of the spectral path encoder; a decoder for fusing the spectral features and the morphological features through a gated attention mechanism to obtain a disaster boundary segmentation map; training and optimizing the morphological image preprocessing module, the spectral path encoder module, the morphological path encoder and the decoder according to a combined loss function, the combined loss function being wherein, is a combined loss function, is a real disaster boundary segmentation map, is a predicted disaster boundary segmentation map, is a spatial gradient, is an indicator function, is a weight coefficient of a cross-entropy loss term, is a cross-entropy loss term, is a weight coefficient of an edge-aware loss term, is a pixel position, is a Euclidean norm. 6.The UAV image boundary segmentation system of claim 5, wherein, The system further comprises: a post-processing module for performing scale adaptive post-processing on the disaster boundary segmentation map through morphological closing operation to obtain a final disaster boundary fine segmentation map.

7. A computer-readable storage medium, characterized in that, A computer program is stored, and when executed by a processor, causes the processor to perform the steps of the method of any one of claims 1 to 4.

Images