Remote sensing image segmentation method and system based on Mama architecture

Abstract

The invention discloses a remote sensing image segmentation method and system based on Mama architecture, and relates to the technical field of computer vision, and the method comprises the steps: obtaining a ship remote sensing image, and carrying out the preprocessing of the ship remote sensing image, and obtaining an optimized image set; training the segmentation model based on the optimized image set to obtain a trained segmentation model; obtaining a to-be-processed image, and sequentially inputting the to-be-processed image into a partition layer and a first convolution unit in the trained segmentation model to obtain a first feature sequence; inputting the first feature sequence into a down-sampling unit to obtain a second feature sequence and a third feature sequence; inputting the third feature sequence into a bottleneck unit to obtain a fourth feature sequence; inputting the second feature sequence, the third feature sequence and the fourth feature sequence into an upper sampling unit to obtain a fifth feature sequence; and inputting the first feature sequence and the fifth feature sequence into a second convolution unit for processing, and performing linear projection to obtain a ship segmentation result. And the segmentation extraction precision of the small target ship in the complex image is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more particularly to a remote sensing image segmentation method and system based on Mamba architecture. Background Art

[0002] At present, ship extraction technology is the core foundation of ocean monitoring, maritime traffic management and marine resource development. With the popularization of high-resolution satellite remote sensing technology, ship detection and segmentation based on remote sensing images have become key means, providing rich data support for dynamic monitoring of the ocean. Among the existing ship extraction methods, they can be mainly divided into two categories: target detection and semantic segmentation.

[0003] Among them, target detection technology can quickly identify the target area by locating it, and has high timeliness. However, traditional target detection technology has weak ability to accurately depict the boundaries of ships, and pays more attention to target positioning rather than accurate description of the target's geometric shape. In tasks that require fine-grained boundary information, this feature often cannot meet actual needs; semantic segmentation technology provides new ideas for ship extraction through pixel-level prediction and accurate depiction of the boundaries of objects. It can not only locate the target, but also fully describe its shape and details, thus showing unique advantages in detailed tasks. However, existing semantic segmentation technology is sensitive to complex ocean noise, has a high missed detection rate of small target ships, and relies on limited data sets, resulting in insufficient generalization ability.

[0004] In recent years, the Mamba structure has attracted widespread attention due to its high precision and powerful long-range dependency modeling capabilities. The existing Mamba-Transformer backbone network combines convolution with Transformer features, breaking through the bottleneck of traditional methods that suffer from reduced segmentation performance in complex backgrounds. The existing technology proposes a universal network U-Mamba for biomedical image segmentation, which opens up a new path for long-range dependency modeling in biomedical image analysis. Although these methods have made important progress in some areas, ship images face great challenges in improving model performance and generalization capabilities for practical applications due to their diversity, random location, and uncertain quantity. In addition, existing research lacks high-quality datasets specifically for ship segmentation and high-precision, high-resolution remote sensing image extraction networks, and there are still significant gaps in the field of ship research.

[0005] Therefore, how to accurately obtain the shape of small target ships in complex images and effectively improve the accuracy of ship segmentation is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0006] In view of this, the present invention provides a remote sensing image segmentation method based on the Mamba architecture, constructs a high-quality data set specifically for ship segmentation, and improves the segmentation and extraction accuracy of small target ships in complex images.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A remote sensing image segmentation method based on Mamba architecture, comprising:

[0009] Acquire multiple ship remote sensing images and perform segmentation and optimization processing to obtain an optimized image set;

[0010] Training the segmentation model based on the optimized image set to obtain a trained segmentation model;

[0011] Obtaining the image to be processed and inputting it into the trained segmentation model;

[0012] The image to be processed is input into a partitioning layer to obtain a plurality of image blocks;

[0013] Based on the multiple image blocks, input into a first convolution unit to obtain a first feature sequence;

[0014] The first feature sequence is input into a downsampling unit to obtain a second feature sequence and a third feature sequence;

[0015] The third feature sequence is input into the bottleneck unit to obtain a fourth feature sequence;

[0016] The second feature sequence, the third feature sequence and the fourth feature sequence are input into an upsampling unit to obtain a fifth feature sequence;

[0017] The first feature sequence and the fifth feature sequence are input into the second convolution unit to obtain the sixth feature sequence and linear projection is performed to obtain the ship segmentation result.

[0018] Preferably, the segmentation specifically includes:

[0019] Frame the ship remote sensing image based on the ship tag coordinates to obtain a framed area;

[0020] Performing mask processing based on the framed area and cutting to generate a corresponding segmentation mask;

[0021] The ship remote sensing image is segmented based on the segmentation mask to obtain a segmented image.

[0022] Preferably, the optimization process specifically includes:

[0023] Applying a morphological opening operation based on the segmented image to remove small white noise in the image to obtain a denoised image;

[0024] Based on the denoised image, a morphological closing operation is applied to fill holes to obtain an optimized image;

[0025] The optimized image set is obtained based on all optimized images.

[0026] Preferably, the first convolution unit includes: a first depthwise convolution and a first processing unit;

[0027] The multiple image blocks are input into the first deep convolution to obtain multiple corresponding short sequences;

[0028] Based on all the short sequences, splicing is performed to obtain a one-dimensional long sequence;

[0029] The one-dimensional long sequence is input into the first processing unit to obtain the first feature sequence.

[0030] Preferably, the down-sampling unit comprises: a second processing unit and a third processing unit;

[0031] The first feature sequence is subjected to a patch merging operation and then input into the second processing unit to obtain the second feature sequence;

[0032] The second feature sequence is subjected to a patch merging operation and then input into the third processing unit to obtain the third feature sequence;

[0033] The bottleneck unit includes: a fourth processing unit and a fifth processing unit;

[0034] The third feature sequence is sequentially input to the fourth processing unit and the fifth processing unit to obtain the fourth feature sequence.

[0035] Preferably, the up-sampling unit comprises: a sixth processing unit and a seventh processing unit;

[0036] The fourth feature sequence is input into the sixth processing unit to obtain a first processing sequence;

[0037] The first processing sequence and the third feature sequence are fused and then patch-extended to obtain a first extended sequence;

[0038] The first extended sequence is input into the seventh processing unit to obtain a second processing sequence;

[0039] The second processing sequence and the second feature sequence are fused and then patch-extended to obtain the fifth feature sequence.

[0040] Preferably, the second convolution unit comprises: an eighth processing unit and a second depthwise convolution;

[0041] The fifth feature sequence is input into the eighth processing unit to obtain a third processing sequence;

[0042] The third processing sequence and the first feature sequence are fused and input into the second deep convolution to obtain the sixth feature sequence.

[0043] Preferably, all processing units have the same structure, including two ResMambaBlock modules connected in series;

[0044] The ResMambaBlock modules all include: a first normalization layer, a first linear transformation layer, a first standardization layer, an activation layer, a MambaLayer module, a second normalization layer, a third normalization layer and a second linear transformation layer;

[0045] The input sequence is input into the first normalization layer to obtain a first intermediate sequence;

[0046] The first intermediate sequence is sequentially input into the first linear transformation layer, the first normalization layer and the activation layer to obtain a second intermediate sequence;

[0047] The second intermediate sequence is input into the MambaLayer module to obtain a third intermediate sequence;

[0048] The third intermediate sequence is input into the second normalization layer to obtain a fourth intermediate sequence;

[0049] The first intermediate sequence is input into the second linear transformation layer to obtain a fifth intermediate sequence;

[0050] The fifth intermediate sequence and the fourth intermediate sequence are fused and input into the third normalization layer to obtain a sixth intermediate sequence;

[0051] The sixth intermediate sequence and the input sequence are fused to obtain an output sequence.

[0052] Preferably, the MambaLayer module includes: a fourth normalization layer, a second standardization layer, an SSM model, a fifth normalization layer and a third linear transformation layer;

[0053] The second intermediate sequence is sequentially input into the fourth normalization layer, the second standardization layer and the SSM model to obtain a process feature sequence;

[0054] The process feature sequence is fused with the learnable parameters and then sequentially input into the fifth normalization layer and the third linear transformation layer to obtain the third intermediate sequence.

[0055] A remote sensing image segmentation system based on Mamba architecture, comprising: a data acquisition module, a model training module, an image segmentation module, a feature extraction module and a result output module;

[0056] The data acquisition module is used to acquire multiple ship remote sensing images and perform segmentation and optimization processing to obtain an optimized image set;

[0057] The model training module is used to train the segmentation model based on the optimized image set to obtain a trained segmentation model;

[0058] The image segmentation module is used to obtain the image to be processed and input it into the trained segmentation model; the image to be processed is input into the partition layer to obtain multiple image blocks;

[0059] The feature extraction module is used to input the plurality of image blocks into a first convolution unit to obtain a first feature sequence; the first feature sequence is input into a downsampling unit to obtain a second feature sequence and a third feature sequence; the third feature sequence is input into a bottleneck unit to obtain a fourth feature sequence; the second feature sequence, the third feature sequence and the fourth feature sequence are input into an upsampling unit to obtain a fifth feature sequence;

[0060] The result output module is used to input the first feature sequence and the fifth feature sequence into the second convolution unit to obtain the sixth feature sequence and perform linear projection to obtain the ship segmentation result.

[0061] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a remote sensing image segmentation method based on the Mamba architecture, which has the following beneficial effects:

[0062] 1. The dataset constructed by the present invention not only restores the image data of the original size, but also ensures the integrity and clarity of the ship boundary during the cutting process. Its high-quality annotation and detailed segmentation provide a reliable foundation for subsequent research and provide strong support for improving the performance of the ship target detection model.

[0063] 2. The image segmentation model disclosed in the present invention can effectively restore the spatial resolution of the image, avoiding the gradient vanishing problem in traditional networks, and aims to comprehensively improve the performance of the model in long-range dependency modeling and complex scene processing.

[0064] 3. The present invention effectively improves the feature representation capability and spatial detail recovery capability by constructing a state-space-based ResMambaBlock module, thereby enhancing the model's ability to process complex images and small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] 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 or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0066] Figure 1 A flow chart of a remote sensing image segmentation method based on the Mamba architecture provided by the present invention.

[0067] Figure 2 This is a flow chart of the optimized image set construction method provided by the present invention.

[0068] Figure 3 This is a schematic diagram of the segmentation model structure provided by the present invention.

[0069] Figure 4 This is a schematic diagram of the ResMambaBlock module structure provided by the present invention.

[0070] Figure 5 This is a schematic diagram of the MambaLayer module structure provided by the present invention.

[0071] Figure 6 This is a schematic diagram of the segmentation results of other target remote sensing images processed by the segmentation model provided by the present invention.

[0072] Figure 7 A schematic diagram of the structure of a remote sensing image segmentation system based on the Mamba architecture provided by the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] Example 1

[0075] like Figure 1 As shown, the embodiment of the present invention discloses a remote sensing image segmentation method based on the Mamba architecture, comprising:

[0076] Acquire multiple ship remote sensing images and perform segmentation and optimization processing to obtain an optimized image set;

[0077] The segmentation model is trained based on the optimized image set to obtain a trained segmentation model;

[0078] Obtain the image to be processed and input it into the trained segmentation model;

[0079] The image to be processed is input into the partitioning layer to obtain multiple image blocks;

[0080] Based on the multiple image blocks, the first convolution unit is input to obtain a first feature sequence;

[0081] The first feature sequence is input into a downsampling unit to obtain a second feature sequence and a third feature sequence;

[0082] The third feature sequence is input into the bottleneck unit to obtain the fourth feature sequence;

[0083] The second feature sequence, the third feature sequence and the fourth feature sequence are input to an upsampling unit to obtain a fifth feature sequence;

[0084] The first feature sequence and the fifth feature sequence are input into the second convolution unit to obtain the sixth feature sequence and perform linear projection to obtain the ship segmentation result.

[0085] Example 2

[0086] The embodiment of the present invention discloses a remote sensing image segmentation method based on Mamba architecture, comprising:

[0087] like Figure 2 As shown, multiple ship remote sensing images are acquired and segmented and optimized to obtain an optimized image set.

[0088] Preferably, in this embodiment, satellite remote sensing images on Google Earth are selected. These satellite remote sensing image datasets cover a variety of objects such as airplanes, bridges, chimneys, dams, ports, ships, and storage tanks. In order to focus on the ship extraction task, the images and labels in the dataset are screened, only images related to ships are retained, and those images that are unclear or do not contain ships are eliminated, ensuring that the screened dataset contains only accurate and high-quality ship images.

[0089] Preferably, the segmentation specifically includes:

[0090] The ship remote sensing image is framed based on the ship tag coordinates to obtain a framed area;

[0091] Perform masking based on the framed area and cut to generate the corresponding segmentation mask;

[0092] The ship remote sensing image is cut based on the segmentation mask to obtain a segmented image.

[0093] Preferably, the ship target can be effectively extracted through the segmentation process, but the cut image may still have defects and noise, which will directly affect the subsequent training effect. Therefore, a morphological patching operation is introduced to optimize the segmentation result.

[0094] Preferably, in this embodiment, the SAM model is used to segment the ship remote sensing image.

[0095] Preferably, the optimization process specifically includes:

[0096] Based on the segmented image, the morphological opening operation is applied to remove the small white noise in the image to obtain the denoised image;

[0097] Based on the denoised image, morphological closing operation is applied to fill the holes and obtain the optimized image;

[0098] Based on all the optimized images, an optimized image set is obtained as the Boat R dataset.

[0099] Preferably, the above-mentioned segmentation and optimization processing effectively improves the quality of the image, and can generate a set of high-quality ship segmentation data sets, providing a solid foundation for subsequent ship segmentation model training.

[0100] Preferably, there is a dilemma in the field of satellite remote sensing ships that there is a lack of high-quality datasets specifically for ship segmentation. Some existing public remote sensing image segmentation datasets are shown in Table 1:

[0101] Table 1 Comparison of Boat R dataset and some public remote sensing image segmentation datasets

[0102]

[0103] As can be seen from Table 1, most remote sensing ship target datasets are limited to the visible light band, and only a few combine data from other bands. In these datasets, the original remote sensing images are often cropped to a smaller size, resulting in incomplete boundaries and unclear images in the cropped ship images, which seriously affects the accuracy of subsequent analysis and model training.

[0104] In contrast, the Boat R dataset established by the present invention not only restores the image data of the original size, but also ensures the integrity and clarity of the ship boundary during the cutting process, making it of great application value in the training of ship segmentation models. Its high-quality annotation and detailed segmentation provide a reliable foundation for subsequent research and provide strong support for improving the performance of ship target detection and segmentation models.

[0105] The segmentation model is trained based on the optimized image set to obtain a trained segmentation model, such as Figure 3 shown.

[0106] Preferably, the segmentation model inputs a 3D color image of size H×W×C, each image is first segmented into multiple image patches and converted into a 1D sequence of dimensions H / 4×W / 4×16. The feature dimension is adjusted to an arbitrary size through an initial linear embedding layer for subsequent processing.

[0107] Preferably, during the training process, the patch tokens processed by the embedding layer are gradually extracted with features through multiple ResMambaBlock and patch merging layers. The segmentation model MambaSegNet gradually reduces the spatial resolution and increases the feature dimension through downsampling operations.

[0108] Each ResMambaBlock module can efficiently learn the features of complex images, especially in the processing of small objects and complex scenes. The output resolution gradually decreases from H×W×C to H / 4×W / 4×C, H / 8×W / 8×2C, H / 16×W / 16×4C and finally H / 32×W / 32×8C. The spatial resolution of the image is gradually restored through upsampling operations, and the shallow features are combined with the deep features using residual connections to restore the lost spatial details. The residual connection effectively avoids the gradient vanishing problem and improves the stability and convergence speed during training.

[0109] Preferably, the convolution layer of the segmentation model MambaSegNet uses a depthwise convolution operation, which significantly reduces the number of parameters and improves computational efficiency. Finally, the high-dimensional features are mapped to the target number of channels through a 1×1 convolution to generate the final segmentation result.

[0110] Preferably, in the training process, this embodiment adopts a combined loss function that combines the cross-entropy loss (Cross-Entropy Loss) and the intersection-over-union loss (IoU Loss). CE for:

[0111] ;

[0112] Among them, y i Indicates the true label value, which is 0 or 1, p i Represents the probability value predicted by the model, and N represents the total number of pixels

[0113] Intersection-over-combination loss L IoU for:

[0114] ;

[0115] Among them, P represents the predicted area, and G represents the real area;

[0116] Combined loss function formula:

[0117] ;

[0118] Among them, α and β control the weights of the cross entropy loss and the intersection-over-union loss respectively. This combined strategy balances the influence of the two, so that the model can accurately classify and effectively capture the target area during training, thereby ensuring a balance between classification accuracy and regional overlap.

[0119] Preferably, the Adam optimizer is used in the optimization process of this embodiment, the learning rate is set to 0.0001, the batch size is 4, and the ratio of the training data set to the verification data set is 7:2.

[0120] The optimally trained segmentation model MambaSegNet has achieved excellent results in the ship segmentation task in high-resolution remote sensing images, especially in the segmentation of complex backgrounds and small targets, showing higher accuracy and computational efficiency compared to existing models. The fully trained model can accurately extract the boundaries and details of ship targets, has high robustness, and is suitable for a variety of practical application scenarios such as marine monitoring and ship emission monitoring.

[0121] Preferably, the accuracy evaluation index is constructed by calculating the loss function:

[0122] Precision evaluation indicators are key criteria for evaluating the segmentation performance of classification models on test sets, and are particularly important in classification tasks. Common evaluation indicators include recall, precision, accuracy, F1-Score, and intersection over union (IoU). Each indicator has unique characteristics and can comprehensively evaluate the performance of the model from different angles, and is suitable for different task requirements.

[0123] Precision measures the proportion of samples that are actually positive among those predicted by the model to be positive. It focuses on evaluating the model's predictive ability on the positive class and is particularly suitable for tasks with more false positives.

[0124] ;

[0125] Among them, TP represents the number of pixels predicted to be positive and actually positive, and FP represents the number of pixels predicted to be positive but actually negative.

[0126] The recall rate focuses on how many samples that are actually positive are correctly identified as positive, emphasizing the model's ability to capture positive samples. It plays an important role when it is necessary to identify as many positive samples as possible.

[0127] ;

[0128] Here, FN represents the number of pixels that are predicted to be negative but are actually positive.

[0129] Accuracy reflects the overall classification accuracy of the model for all samples. Although it can provide an overview of the overall performance of the model, it may not fully reflect the actual ability of the model when the samples are unevenly distributed.

[0130] .

[0131] The F1 value, as the harmonic mean of precision and recall, is suitable for tasks that require a balance between the two, and is particularly important in small object segmentation tasks.

[0132] .

[0133] The IoU indicator measures the degree of overlap between the predicted area and the true area. It is particularly suitable for target detection and segmentation tasks, and can evaluate the accuracy of the model in locating and segmenting the target area.

[0134] .

[0135] These evaluation indicators complement each other and provide the ability to consider model performance from multiple dimensions, ensuring that the evaluation results of the model are more comprehensive and accurate in different tasks and scenarios.

[0136] Get the image to be processed and input it into the trained segmentation model.

[0137] The image to be processed is input into the partition layer to obtain multiple image blocks.

[0138] Based on multiple image blocks input into the first convolution unit, a first feature sequence is obtained.

[0139] Preferably, the first convolution unit includes: a first depthwise convolution and a first processing unit;

[0140] Multiple image blocks are input into the first depth convolution to obtain multiple corresponding short sequences;

[0141] All short sequences are concatenated to obtain a one-dimensional long sequence;

[0142] The one-dimensional long sequence is input into the first processing unit to obtain a first feature sequence.

[0143] The first feature sequence is input to the down-sampling unit to obtain the second feature sequence and the third feature sequence.

[0144] Preferably, the down-sampling unit comprises: a second processing unit and a third processing unit;

[0145] The first feature sequence is subjected to a patch merging operation and then input into a second processing unit to obtain a second feature sequence;

[0146] The second feature sequence is subjected to a patch merging operation and then input into a third processing unit to obtain a third feature sequence.

[0147] The third feature sequence is input into the bottleneck unit to obtain a fourth feature sequence.

[0148] Preferably, the bottleneck unit comprises: a fourth processing unit and a fifth processing unit;

[0149] The third feature sequence is sequentially input to the fourth processing unit and the fifth processing unit to obtain a fourth feature sequence.

[0150] The second feature sequence, the third feature sequence and the fourth feature sequence are input to the up-sampling unit to obtain a fifth feature sequence.

[0151] Preferably, the up-sampling unit comprises: a sixth processing unit and a seventh processing unit;

[0152] The fourth feature sequence is input into the sixth processing unit to obtain a first processing sequence;

[0153] The first processing sequence and the third feature sequence are fused and then patch-extended to obtain a first extended sequence;

[0154] The first extended sequence is input into the seventh processing unit to obtain a second processing sequence;

[0155] The second processing sequence and the second feature sequence are fused and then patch-extended to obtain the fifth feature sequence.

[0156] The first feature sequence and the fifth feature sequence are input into the second convolution unit to obtain the sixth feature sequence and perform linear projection to obtain the ship segmentation result.

[0157] Preferably, the second convolution unit comprises: an eighth processing unit and a second depthwise convolution;

[0158] The fifth feature sequence is input into the eighth processing unit to obtain a third processing sequence;

[0159] The third processing sequence and the first feature sequence are fused and input into the second deep convolution to obtain the sixth feature sequence.

[0160] Preferably, all the processing units have the same structure, that is, the first processing unit, the second processing unit, the third processing unit, the fourth processing unit, the fifth processing unit, the sixth processing unit, the seventh processing unit and the eighth processing unit have the same structure, and all include two ResMambaBlock modules connected in series;

[0161] like Figure 4As shown, the ResMambaBlock modules include: a first normalization layer, a first linear transformation layer, a first standardization layer, an activation layer, a MambaLayer module, a second normalization layer, a third normalization layer and a second linear transformation layer;

[0162] The input sequence is input into the first normalization layer to obtain the first intermediate sequence;

[0163] The first intermediate sequence is sequentially input into the first linear transformation layer, the first normalization layer and the activation layer to obtain a second intermediate sequence;

[0164] The second intermediate sequence is input into the MambaLayer module to obtain a third intermediate sequence;

[0165] The third intermediate sequence is input into the second normalization layer to obtain a fourth intermediate sequence;

[0166] The first intermediate sequence is input into the second linear transformation layer to obtain a fifth intermediate sequence;

[0167] The fifth intermediate sequence and the fourth intermediate sequence are fused and input into the third normalization layer to obtain the sixth intermediate sequence;

[0168] The sixth intermediate sequence and the input sequence are fused to obtain an output sequence.

[0169] Preferably, unlike the traditional visual Transformer, the ResMambaBlock module used in the present invention abandons positional embedding, selects a simpler structure, removes the MLP stage, and enables more modules to be stacked under the same computational budget, thereby improving computational efficiency and feature expression capabilities. In addition, by introducing the MambaLayer module, the network can effectively learn complex image features when processing small targets and complex scenes in remote sensing images.

[0170] Preferably, by constructing a state-space based ResMambaBlock module, the feature representation capability and spatial detail recovery capability are effectively improved, thereby enhancing the model's ability to process complex images and small objects.

[0171] Preferably, Figure 5 As shown, the MambaLayer module includes: a fourth normalization layer, a second standardization layer, an SSM model, a fifth normalization layer and a third linear transformation layer;

[0172] The second intermediate sequence is sequentially input into the fourth normalization layer, the second standardization layer and the SSM model to obtain a process feature sequence;

[0173] The process feature sequence is fused with the learnable parameters and then input into the fifth normalization layer and the third linear transformation layer in sequence to obtain the third intermediate sequence.

[0174] Preferably, the SSM model is a Mamba state space model, and the learning parameters are used to adjust the ratio of skip connections.

[0175] Preferably, by training and testing the road and building datasets, the segmentation model MambaSegNet of the present invention can be applied to the segmentation and extraction of roads and buildings. Figure 6 The results show that the segmentation model MambaSegNet of the present invention has strong adaptability and stability in both the extraction of road networks and the precise segmentation of building forms.

[0176] Example 3

[0177] like Figure 7 As shown, a remote sensing image segmentation system based on Mamba architecture includes: a data acquisition module, a model training module, an image segmentation module, a feature extraction module and a result output module;

[0178] A data acquisition module is used to acquire multiple ship remote sensing images and perform segmentation and optimization processing to obtain an optimized image set;

[0179] A model training module is used to train the segmentation model based on the optimized image set to obtain a trained segmentation model;

[0180] The image segmentation module is used to obtain the image to be processed and input it into the trained segmentation model; the image to be processed is input into the partition layer to obtain multiple image blocks;

[0181] A feature extraction module is used to input a first convolution unit based on multiple image blocks to obtain a first feature sequence; the first feature sequence is input to a downsampling unit to obtain a second feature sequence and a third feature sequence; the third feature sequence is input to a bottleneck unit to obtain a fourth feature sequence; the second feature sequence, the third feature sequence and the fourth feature sequence are input to an upsampling unit to obtain a fifth feature sequence;

[0182] The result output module is used to input the first feature sequence and the fifth feature sequence into the second convolution unit to obtain the sixth feature sequence and perform linear projection to obtain the ship segmentation result.

[0183] Preferably, the function implementation method of each module in this embodiment corresponds to the above method one by one, and will not be repeated here.

[0184] Example 4

[0185] Based on the same inventive concept, the present invention also provides a computer device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0186] Memory, used to store computer programs;

[0187] The processor, when used to execute the program stored in the memory, can implement a remote sensing image segmentation method based on the Mamba architecture as in Example 1 or 2.

[0188] The electronic device may include: a processor, a communications interface, a memory and a communications bus, wherein the processor, the communications interface and the memory communicate with each other via the communications bus.

[0189] The processor can call the logic instructions in the memory to execute a remote sensing image segmentation method based on the Mamba architecture in Example 1 or 2.

[0190] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0191] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a remote sensing image segmentation method based on the Mamba architecture, which has the following beneficial effects:

[0192] 1. The dataset constructed by the present invention not only restores the image data of the original size, but also ensures the integrity and clarity of the ship boundary during the cutting process. Its high-quality annotation and detailed segmentation provide a reliable foundation for subsequent research and provide strong support for improving the performance of the ship target detection model.

[0193] 2. The image segmentation model disclosed in the present invention can effectively restore the spatial resolution of the image, avoiding the gradient vanishing problem in traditional networks, and aims to comprehensively improve the performance of the model in long-range dependency modeling and complex scene processing.

[0194] 3. The present invention effectively improves the feature representation capability and spatial detail recovery capability by constructing a state-space-based ResMambaBlock module, thereby enhancing the model's ability to process complex images and small targets.

[0195] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0196] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image segmentation method based on Mamba architecture, characterized in that: include: Acquire multiple ship remote sensing images and perform segmentation and optimization processing to obtain an optimized image set; Training the segmentation model based on the optimized image set to obtain a trained segmentation model; Obtaining the image to be processed and inputting it into the trained segmentation model; The image to be processed is input into a partitioning layer to obtain a plurality of image blocks; Based on the multiple image blocks, input into a first convolution unit to obtain a first feature sequence; The first feature sequence is input into a downsampling unit to obtain a second feature sequence and a third feature sequence; The third feature sequence is input into the bottleneck unit to obtain a fourth feature sequence; The second feature sequence, the third feature sequence and the fourth feature sequence are input into an upsampling unit to obtain a fifth feature sequence; The first feature sequence and the fifth feature sequence are input into the second convolution unit to obtain the sixth feature sequence and linear projection is performed to obtain the ship segmentation result.

2. The remote sensing image segmentation method based on Mamba architecture according to claim 1, characterized in that: The segmentation specifically includes: Frame the ship remote sensing image based on the ship tag coordinates to obtain a framed area; Performing mask processing based on the framed area and cutting to generate a corresponding segmentation mask; The ship remote sensing image is segmented based on the segmentation mask to obtain a segmented image.

3. The remote sensing image segmentation method based on Mamba architecture according to claim 2, characterized in that: The optimization process specifically includes: Applying a morphological opening operation based on the segmented image to remove small white noise in the image to obtain a denoised image; Based on the denoised image, a morphological closing operation is applied to fill holes to obtain an optimized image; The optimized image set is obtained based on all optimized images.

4. The remote sensing image segmentation method based on Mamba architecture according to claim 1, characterized in that: The first convolution unit includes: a first depth convolution and a first processing unit; The multiple image blocks are input into the first deep convolution to obtain multiple corresponding short sequences; Based on all the short sequences, splicing is performed to obtain a one-dimensional long sequence; The one-dimensional long sequence is input into the first processing unit to obtain the first feature sequence.

5. The remote sensing image segmentation method based on Mamba architecture according to claim 4, characterized in that: The down sampling unit comprises: a second processing unit and a third processing unit; The first feature sequence is subjected to a patch merging operation and then input into the second processing unit to obtain the second feature sequence; The second feature sequence is subjected to a patch merging operation and then input into the third processing unit to obtain the third feature sequence; The bottleneck unit includes: a fourth processing unit and a fifth processing unit; The third feature sequence is sequentially input to the fourth processing unit and the fifth processing unit to obtain the fourth feature sequence.

6. The remote sensing image segmentation method based on Mamba architecture according to claim 1, characterized in that: The up-sampling unit comprises: a sixth processing unit and a seventh processing unit; The fourth feature sequence is input into the sixth processing unit to obtain a first processing sequence; The first processing sequence and the third feature sequence are fused and then patch-extended to obtain a first extended sequence; The first extended sequence is input into the seventh processing unit to obtain a second processing sequence; The second processing sequence and the second feature sequence are fused and then patch-extended to obtain the fifth feature sequence.

7. The remote sensing image segmentation method based on Mamba architecture according to claim 6, characterized in that: The second convolution unit includes: an eighth processing unit and a second depth convolution; The fifth feature sequence is input into the eighth processing unit to obtain a third processing sequence; The third processing sequence and the first feature sequence are fused and input into the second deep convolution to obtain the sixth feature sequence.

8. A remote sensing image segmentation method based on Mamba architecture according to claim 5 or 7, characterized in that: All processing units have the same structure, consisting of two ResMambaBlock modules connected in series; The ResMambaBlock modules all include: a first normalization layer, a first linear transformation layer, a first standardization layer, an activation layer, a MambaLayer module, a second normalization layer, a third normalization layer and a second linear transformation layer; The input sequence is input into the first normalization layer to obtain a first intermediate sequence; The first intermediate sequence is sequentially input into the first linear transformation layer, the first normalization layer and the activation layer to obtain a second intermediate sequence; The second intermediate sequence is input into the MambaLayer module to obtain a third intermediate sequence; The third intermediate sequence is input into the second normalization layer to obtain a fourth intermediate sequence; The first intermediate sequence is input into the second linear transformation layer to obtain a fifth intermediate sequence; The fifth intermediate sequence and the fourth intermediate sequence are fused and input into the third normalization layer to obtain a sixth intermediate sequence; The sixth intermediate sequence and the input sequence are fused to obtain an output sequence.

9. The remote sensing image segmentation method based on Mamba architecture according to claim 8, characterized in that: The MambaLayer module includes: a fourth normalization layer, a second standardization layer, an SSM model, a fifth normalization layer and a third linear transformation layer; The second intermediate sequence is sequentially input into the fourth normalization layer, the second standardization layer and the SSM model to obtain a process feature sequence; The process feature sequence is fused with the learnable parameters and then sequentially input into the fifth normalization layer and the third linear transformation layer to obtain the third intermediate sequence.

10. A remote sensing image segmentation system based on Mamba architecture, applied to a remote sensing image segmentation method based on Mamba architecture as claimed in any one of claims 1 to 9, characterized in that: include: Data acquisition module, model training module, image segmentation module, feature extraction module and result output module; The data acquisition module is used to acquire multiple ship remote sensing images and perform segmentation and optimization processing to obtain an optimized image set; The model training module is used to train the segmentation model based on the optimized image set to obtain a trained segmentation model; The image segmentation module is used to obtain the image to be processed and input it into the trained segmentation model; the image to be processed is input into the partition layer to obtain multiple image blocks; The feature extraction module is used to input the plurality of image blocks into a first convolution unit to obtain a first feature sequence; the first feature sequence is input into a downsampling unit to obtain a second feature sequence and a third feature sequence; the third feature sequence is input into a bottleneck unit to obtain a fourth feature sequence; the second feature sequence, the third feature sequence and the fourth feature sequence are input into an upsampling unit to obtain a fifth feature sequence; The result output module is used to input the first feature sequence and the fifth feature sequence into the second convolution unit to obtain the sixth feature sequence and perform linear projection to obtain the ship segmentation result.

Images