Image segmentation method based on lightweight feature extraction and multi-receptive-field feature interaction

Abstract

The invention discloses an image segmentation method based on lightweight feature extraction and multi-receptive-field feature interaction, and belongs to the technical field of remote sensing image processing. The method comprises the steps of building an image segmentation model based on a plurality of lightweight Mama blocks; each lightweight Mama block comprises a first lightweight feature extraction module, a multi-receptive-field feature interaction module and a second lightweight feature extraction module which are connected in sequence; performing remote sensing image segmentation training on the built image segmentation model; and performing image segmentation on the target remote sensing image by using the trained image segmentation model. According to the method, the calculation complexity can be effectively reduced by utilizing lightweight feature extraction and multi-receptive-field feature interaction, and meanwhile, the feature extraction capability, especially the segmentation precision of complex ground features in a remote sensing image, is improved. In addition, a composite loss function is designed, the loss function can effectively solve the problem of class imbalance in remote sensing image segmentation, and the segmentation precision is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction. Background Art

[0002] With the rapid development of remote sensing technology, remote sensing images, as a vital information source, play a crucial role in a variety of fields, including environmental monitoring, land use, urban planning, and agricultural surveys. Remote sensing image segmentation, a core problem in image processing, aims to divide an image into multiple meaningful regions for further analysis and application. However, due to the unique characteristics of remote sensing images, such as large scale, high resolution, and multispectral information, remote sensing image segmentation faces numerous challenges.

[0003] First, remote sensing images typically have complex scene structures, containing a variety of ground objects (such as buildings, roads, water bodies, and farmland), with significant scale variations. Traditional image segmentation methods, such as pixel-based classification, are often ineffective in extracting objects of varying scales. Especially in large-scale remote sensing imagery, where image resolution is often very high and processing is computationally expensive, existing segmentation algorithms often struggle to strike a good balance between computational efficiency and accuracy.

[0004] Secondly, existing remote sensing image segmentation methods often rely on traditional convolutional neural networks (CNNs) for feature extraction. However, standard convolution operations have limitations when capturing the complex textures and multi-scale features found in remote sensing images. In particular, when faced with long-range dependencies in remote sensing images, traditional CNNs often fail to effectively capture this long-range information. Therefore, effectively extracting and fusing global and local features remains a pressing challenge in current remote sensing image segmentation technology.

[0005] Furthermore, data imbalance in remote sensing images is a key factor affecting segmentation accuracy. Certain ground objects (such as water bodies and forests) occupy a small area in remote sensing images and may have very subtle texture differences within the image. Traditional pixel-level loss functions (such as cross-entropy loss) often have poor recognition performance for these small objects. Therefore, designing effective loss functions to address the class imbalance problem is a key issue in remote sensing image segmentation.

[0006] To address these issues, remote sensing image segmentation methods based on deep learning have gained widespread attention in recent years. In particular, models such as convolutional neural networks (CNNs) and Transformer networks have become a mainstream approach for remote sensing image segmentation. However, existing methods still have limitations in terms of computational complexity, feature representation capabilities, and multi-scale perception capabilities.

[0007] While existing lightweight convolutional neural networks (such as MobileNet and EfficientNet) have improved computational efficiency to a certain extent, they still struggle to balance computational efficiency and segmentation accuracy when working with high-resolution remote sensing images. More importantly, these methods often fail to fully utilize the multi-scale information and global features in remote sensing images, resulting in low segmentation accuracy in complex scenes.

[0008] Furthermore, while Transformer networks have demonstrated powerful capabilities in visual tasks in recent years, their computational complexity is high, particularly in remote sensing image segmentation, where processing large amounts of data often creates a bottleneck. Therefore, ensuring efficient computation while maintaining strong feature representation capabilities has become an important area of research in remote sensing image segmentation. Summary of the Invention

[0009] In order to at least partially solve the technical problems existing in the above-mentioned background technology, the present invention provides an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction, which can not only ensure high-precision feature extraction while reducing computational complexity and adapt to the processing requirements of large-scale remote sensing image data.

[0010] To achieve the above object, the technical solution adopted by the present invention is:

[0011] In a first aspect, an embodiment of the present invention provides an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction, the method comprising the following steps:

[0012] S1. Building an image segmentation model based on multiple lightweight Mamba blocks; each lightweight Mamba block includes: a first lightweight feature extraction module, a multi-receptive field feature interaction module, and a second lightweight feature extraction module connected in sequence;

[0013] S2. Performing remote sensing image segmentation training on the constructed image segmentation model;

[0014] S3. Use the trained image segmentation model to perform image segmentation on the target remote sensing image.

[0015] Furthermore, the first lightweight feature extraction module uses depth-separable convolution, normalization, lightweight feedforward neural network and residual connection to perform lightweight feature extraction and fusion processing on the input features, and sends the processed features to the multi-receptive field feature interaction module.

[0016] Furthermore, the multi-receptive field feature interaction module uses a multi-receptive field feature interaction mechanism to process the features output by the first lightweight feature extraction module to extract local features and global features. The processing flow includes:

[0017] The features output by the first lightweight feature extraction module are divided into three sub-features along the channel dimension and sent to three different feature extraction branches respectively. The three different feature extraction branches use identity mapping, multi-scale convolution, and wavelet transform Mamba to process the sub-features respectively, and then splice the processed sub-features to generate the final output of the multi-receptive field feature interaction module and send it to the second lightweight feature extraction module.

[0018] Furthermore, the second lightweight feature extraction module uses depth-separable convolution, normalization, lightweight feedforward neural network and residual connection to perform lightweight feature extraction and fusion processing on the features output by the multi-receptive field feature interaction module to generate the final output of the lightweight Mamba block.

[0019] Furthermore, the lightweight feedforward neural network only contains two fully connected layers, and the hidden layer dimension is adjustable.

[0020] Furthermore, in S2, a composite loss function is used to train the image segmentation model. The composite loss function is:

[0021] In S2, a composite loss function is used to train the image segmentation model. The composite loss function is:

[0022]

[0023] Among them, y i is the true label of pixel i, p i is the predicted probability of pixel i, λ1, λ2, λ3 are hyperparameters used to adjust the weights of each loss term, and N represents the total number of pixels in the image.

[0024] In a second aspect, the present invention also provides an electronic device comprising a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the above-mentioned image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction.

[0025] Compared with the prior art, the present invention has at least the following beneficial effects:

[0026] 1. The present invention provides an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction. Through lightweight feature extraction and multi-receptive field feature interaction, the present invention can effectively reduce computational complexity while improving feature extraction capabilities, especially helping to improve the segmentation accuracy of complex objects in remote sensing images.

[0027] 2. This paper reduces computational overhead through deep separable convolution and lightweight feedforward neural network (FFN), while the multi-receptive field feature interaction module enhances the model's multi-scale perception capability through identity mapping, multi-scale convolution, and wavelet transform technologies, and can effectively capture local and global features in remote sensing images.

[0028] 3. The present invention also constructs a composite loss function, which combines pixel-level cross entropy loss, Dice loss and IoU loss, which can effectively solve the category imbalance problem in remote sensing image segmentation and further improve the segmentation accuracy.

[0029] Compared to existing technologies, this method improves segmentation accuracy and robustness while ensuring efficient computation, making it applicable to a variety of remote sensing image segmentation tasks. The method can better address complex objects, objects of varying scales, and class imbalance in remote sensing images, and has broad application prospects.

[0030] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0031] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0033] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0034] Figure 1 A flow chart of an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction provided by an embodiment of the present invention.

[0035] Figure 2 A schematic diagram of the working principle of the image segmentation model provided by an embodiment of the present invention.

[0036] Figure 3 Schematic diagram of the working principle of the multi-receptive field feature interaction module provided in an embodiment of the present invention.

[0037] Figure 4 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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.

[0039] In describing the present invention, it should be noted that some processes described in this specification and accompanying drawings include multiple operations that appear in a specific order. However, it should be understood that these operations may be performed in a different order than the order in which they appear, or may be performed in parallel. Furthermore, the use of various sequence numbers is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0040] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0041] See also Figure 1 As shown, the present invention provides an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction, which mainly includes the following steps:

[0042] S1. Building an image segmentation model based on multiple lightweight Mamba blocks; each lightweight Mamba block includes: a first lightweight feature extraction module, a multi-receptive field feature interaction module, and a second lightweight feature extraction module connected in sequence;

[0043] S2. Perform remote sensing image segmentation training on the constructed image segmentation model;

[0044] S3. Use the trained image segmentation model to perform image segmentation on the target remote sensing image.

[0045] The following combination Figure 2-Figure 3 As shown, the working principle and specific implementation methods of the method of the present invention are described in detail:

[0046] In the remote sensing image segmentation task, it is crucial to build an efficient and accurate network architecture. In this embodiment, the structure of the image segmentation model is as follows Figure 2As shown in the figure, by combining the lightweight Mamba block (including two lightweight feature extraction modules and a multi-receptive field feature interaction module) with the downsampling operation, an efficient multi-level feature extraction and fusion mechanism is formed, and the network performance is optimized through a composite loss function. Specific implementation methods include:

[0047] 1. Lightweight Mamba Block:

[0048] In the embodiment of the present invention, Figure 2 As shown in the figure, in order to improve the computational efficiency of remote sensing image segmentation while maintaining high-precision feature extraction capabilities, a lightweight Mamba block (a visual state space model Mamba) is proposed. Each lightweight Mamba block includes: a first lightweight feature extraction module, a multi-receptive field feature interaction module and a second lightweight feature extraction module; wherein:

[0049] The first and second lightweight feature extraction modules use technologies such as deep separable convolution, normalization, lightweight feedforward neural networks (FFNs), and residual connections to reduce computational complexity while enhancing model expressiveness. The multi-receptive field feature interaction module uses a multi-receptive field feature interaction mechanism to achieve the fusion of local and global information through channel decomposition, identity mapping, multi-scale convolution, wavelet transform, and Mamba state space modeling (SSM), thereby improving the accuracy and robustness of remote sensing image segmentation. The specific working principle is as follows:

[0050] 1. Lightweight feature extraction:

[0051] 1.1: Lightweight feature preprocessing. Assuming the input feature is X, in the first lightweight feature extraction module, the depth-wise separable convolution layer is first used to extract spatial features, and the normalization layer is combined to normalize the features to obtain feature X. n :

[0052] X n =Norm(DWConv(X))

[0053] Here, "Norm" represents a normalization layer, and "DWConv" represents a depthwise separable convolution layer. Depthwise separable convolution reduces computational complexity, making it suitable for large-scale remote sensing image processing. The normalization layer improves training stability and ensures a balanced distribution of features.

[0054] 1.2: Lightweight Feedforward Neural Network (FFN). In order to improve the nonlinear expression ability of features while keeping the computation lightweight, a feedforward neural network (FFN) consisting of only two fully connected layers is used:

[0055] X′ n =FFN(X n )=Linear2(ReLU(Linear1(Xn )))

[0056] Here, Linear represents the linear layer, and ReLU represents the ReLU activation function. Lightweight Advantages: Computational Efficiency: Contains only two fully connected layers, and reduces computational overhead through adjustable intermediate dimensions. Fewer Parameters: Compared to the multi-head attention mechanism of the Transformer structure, this FFN has lower computational complexity, making it suitable for lightweight remote sensing tasks. Improved Feature Representation: The ReLU nonlinear transformation ensures that the model can learn complex patterns.

[0057] 1.3: Efficient residual connection and multi-receptive field feature interaction. In order to avoid information loss and reduce computational redundancy, residual connection is used:

[0058]

[0059] in, represents the matrix addition operation. Then, X1 进 Enter the multi-receptive field feature interaction module to capture remote sensing ground object information of different scales, and obtain (The specific working principle of the multi-receptive field feature interaction module is introduced in detail in the following two sections).

[0060] 1.4: Further lightweight feature fusion. In order to ensure efficient calculation and enhance feature expression capabilities, a simplified residual path is used for feature fusion:

[0061]

[0062] Then, After the second lightweight feature extraction module (depthwise separable convolution + normalization + FFN), we get X2:

[0063]

[0064] Finally, X2 and Perform residual connections to generate the final output of the lightweight Mamba block:

[0065]

[0066] The advantages of the lightweight feature extraction module are summarized as follows:

[0067] 1) Reduce the amount of computation: Use depth-wise separable convolution instead of standard convolution, reducing the amount of computation by more than 80%. Use lightweight FFN to avoid computational redundancy in the Transformer structure.

[0068] 2) Reduce the number of parameters: Depthwise separable convolution reduces the number of parameters to 1 / K of the original standard convolution 2, which is suitable for high-resolution remote sensing data processing. In this embodiment, the FFN only contains two fully connected layers, and the hidden layer dimension is adjustable, further reducing the number of parameters.

[0069] 3) Improved training stability: Multiple residual connections ensure smooth gradient flow, enabling efficient model training even with reduced parameters. Normalization layers reduce internal covariate shift during training, making the lightweight Mamba block easier to optimize.

[0070] 4) Suitable for low-computing power devices: Its lightweight design enables it to be adapted to edge computing devices, drone remote sensing systems, or mobile terminals, reducing dependence on GPU resources.

[0071] 2. Multi-receptive field feature interaction:

[0072] Remote sensing images typically feature complex scene structures, objects of varying scales, and rich texture information. Traditional single-receptive-field feature extraction methods struggle to effectively capture this information. Therefore, this module employs a multi-receptive-field feature interaction mechanism. Through channel decomposition, identity mapping, multi-scale convolution, wavelet transform, and Mamba state-space modeling (SSM), it achieves the fusion of local and global information, improving the accuracy and robustness of remote sensing image segmentation.

[0073] In the embodiment of the present invention, the specific structure and workflow of the multi-receptive field feature interaction module are as follows: Figure 3 The specific working principle is as follows:

[0074] 2.1: Feature channel segmentation, building multi-receptive field branches. Assume that the input feature is Where H and W represent the spatial dimensions of the image, and C represents the number of channels.

[0075] Divide the feature F into three sub-features along the channel dimension They are fed into three different feature extraction branches:

[0076] The first branch (identity mapping): maintains the original features and improves computational efficiency.

[0077] The second branch (multi-scale convolution): extracts local features and enhances the model's perception of objects of different scales.

[0078] The third branch (wavelet transform Mamba): combines wavelet transform and Mamba state space modeling to extract global features and high-frequency edge information.

[0079] This branch structure enables feature interaction across different receptive fields, enabling the model to capture local details of remote sensing images while modeling long-range dependencies, thereby improving segmentation accuracy. The specific advantages and workflow of the three branches are as follows:

[0080] 2.2: The first branch is the identity mapping. Lightweight identity mapping reduces computational redundancy. Directly As output, no complex transformation is performed. Advantages:

[0081] High computational efficiency: no additional calculation is required, reducing computational complexity.

[0082] Lossless information transmission: retain key ground feature information in remote sensing images, such as building edges and roads.

[0083] Avoid overfitting: Directly transferring features makes the model easier to train and reduces the risk of overfitting.

[0084] This design ensures that the original feature information will not be lost due to excessive calculations, and also provides a basis for subsequent fusion.

[0085] 2.3: Second branch (multi-scale convolution module). Multi-scale receptive field, adapting to objects of different scales. Objects in remote sensing images (such as buildings, farmland, water bodies, etc.) have different spatial scales. The convolution operation of a single receptive field is difficult to capture the features of large and small targets at the same time. Therefore, this branch uses a multi-scale convolution module to extract local feature information at different receptive fields. Specific steps:

[0086] 1) Channel division: Input features Divide into n groups along the channel dimension:

[0087]

[0088] 2) Convolution kernels of different sizes capture multi-scale information:

[0089] For each sub-feature Apply convolution kernels of different scales:

[0090]

[0091] Where k i Represents different convolution kernel sizes, for example:

[0092] ① Small kernel (3×3): captures fine textures such as roads or rivers;

[0093] ② Medium core (5×5): identifies medium-sized objects, such as buildings;

[0094] ③ Large core (7×7): perceives large-scale targets such as forests and lakes.

[0095] 3) Feature Fusion:

[0096] Concatenate all convolutional features:

[0097]

[0098] This module combines the efficiency of multi-scale receptive field and deep convolution, effectively improving the local feature extraction capability of remote sensing image segmentation.

[0099] 2.4: The third branch: Processing is performed through the Mamba module enhanced by long-range wavelet transform to extract global features and enhance the extraction of high-frequency edge information. First, the feature channel is divided and the input feature Divided into two parts along the channel dimension: the first part Used by the Mamba module to extract global features.

[0100] Part 2 Used for wavelet transform to enhance high-frequency edge information. Assuming the channel ratio is ξ, then:

[0101]

[0102] Next, Apply the Mamba module to extract global features. The Mamba module captures long-range dependencies through bidirectional scanning and state-space modeling (SSM). The specific operation is as follows:

[0103]

[0104] Among them, SSM represents the state space model, ReLU represents the ReLU nonlinear activation function, and Conv represents the convolution operation. Next, wavelet transform processing is performed on Apply wavelet transform to extract high-frequency edge information. The specific steps are as follows:

[0105] Wavelet transform (WT): Perform wavelet transform to obtain feature representations at different frequency scales.

[0106] Local convolution operation: Perform local convolution operation on the features after wavelet transformation to extract high-frequency edge information.

[0107] Inverse wavelet transform (IWT): The convolution feature is restored to its original size through inverse wavelet transform

[0108] Then, feature fusion is performed to fuse the global features extracted by the Mamba module and the high-frequency edge information enhanced by wavelet transform to obtain the output feature

[0109] Finally, the outputs of each branch are spliced together to get the final output

[0110] In this embodiment, the advantages of the multi-receptive field feature interaction module are summarized as follows:

[0111] (1) Global feature extraction:

[0112] The Mamba module captures long-range dependencies through a state-space model (SSM), effectively extracting global features. The core concept of the SSM is to map the input sequence to hidden states and propagate information through state transition equations. This design enables the Mamba module to capture long-range dependencies in the image, thereby enhancing the model's understanding of global information.

[0113] (2) High-frequency edge information enhancement:

[0114] The wavelet transform is a multiresolution analysis method that decomposes an image into components at different frequency scales. High-frequency components (such as LH, HL, and HH) contain edge and texture information. By convolving these high-frequency components and restoring them to their original size, the WTE-Mamba module enhances the ability to extract high-frequency edge information. This design not only preserves global features but also significantly improves the model's ability to perceive detailed information.

[0115] (3) Computational efficiency:

[0116] Wavelet transforms and inverse wavelet transforms have low operational complexity and can be efficiently computed using fast wavelet transform algorithms (such as the Haar wavelet transform). Furthermore, the linear time complexity of the Mamba module enables the entire WTE-Mamba module to effectively extract global and high-frequency features while maintaining efficient computation.

[0117] 3. Image segmentation workflow:

[0118] 3.1: Preprocessing of original remote sensing images:

[0119] In the embodiment of the present invention, the remote sensing images collected by the original equipment usually have different resolutions, noise interference, illumination changes and other problems, and therefore need to undergo a preprocessing step to adapt to the subsequent model input.

[0120] The pre-processing process includes:

[0121] Denoising: Apply image denoising techniques, such as Gaussian filtering and non-local means (NLM) denoising, to remove noise from the image.

[0122] Geometric correction: Perform geometric correction on remote sensing images to ensure that the spatial information in the image is consistent with the spatial relationship of the actual objects.

[0123] Normalization: Normalize the image so that the pixel value range adapts to the model input. A common practice is to map the pixel value range to [0, 1] or [-1, 1].

[0124] Enhancement: Use image augmentation techniques (such as rotation, translation, and mirroring) to increase data diversity and improve model robustness.

[0125] After the above preprocessing, a standardized, clear, and geometrically accurate input image is obtained, which serves as the basis for subsequent feature extraction and segmentation tasks.

[0126] 3.2: Image feature extraction. The preprocessed image is used as input to the first part of the network, which is mainly used for preliminary feature extraction.

[0127] Patch Embedding and Convolution: The input image undergoes patch embedding, dividing the image into multiple small patches and converting them into vectors. This allows the subsequent network to efficiently process local regions. Preliminary feature extraction is performed through convolutional layers. The size and stride of the convolution kernel are designed based on the specific requirements of the remote sensing image to extract low-level texture features (such as edges and corners).

[0128] Normalization layers and activation functions: Use batch normalization or layer normalization to ensure greater stability during network training. Use the ReLU activation function (or other activation functions, such as Leaky ReLU) to enhance the model's nonlinear capabilities and increase the richness of feature expression.

[0129] Through these operations, the intermediate feature X is obtained. At this time, the feature already has some local information in the remote sensing image, laying the foundation for subsequent deeper feature fusion and spatial information modeling.

[0130] 3.3: Feature Extraction and Fusion via Lightweight Mamba Blocks and Downsampling Modules. In this step, the feature X obtained after preliminary feature extraction is passed through multiple Lightweight Mamba Blocks and downsampling modules to further extract high-level features and reduce the spatial resolution of the feature map. Feature X is then fed into the constructed Lightweight Mamba Block, outputting X′2. This is then subjected to downsampling, primarily consisting of convolutional layers and max pooling layers. Convolutional layers further compress the spatial resolution of the feature map while preserving deep spatial information. Max pooling layers reduce the size of the feature map and extract key feature patterns, playing a particularly crucial role in capturing larger features in the image, such as mountains and lakes. After further processing through Lightweight Mamba Blocks, downsampling, Lightweight Mamba Blocks, and average pooling layers, the resulting feature maps are further reduced in dimensionality and enriched in size. These feature maps are then fed into the segmentation head for final segmentation decision making, ultimately outputting predicted probabilities.

[0131] 3.4: In the remote sensing image segmentation task, model training requires the use of an appropriate loss function to guide network learning. To this end, this paper designs a composite loss function that combines pixel-level loss, Dice loss, and region-level loss to improve segmentation accuracy and stability. The details are as follows:

[0132] 1) Pixel-level cross-entropy loss: For each pixel, the cross-entropy loss between the predicted category and the true category is calculated to ensure that the segmentation result of each pixel is close to the true label. This loss function is particularly suitable for binary or multi-class segmentation tasks and can provide accurate feedback at the pixel level.

[0133] 2) Dice Loss: The Dice coefficient is a metric that measures the similarity between two samples and is particularly useful for segmentation tasks with imbalanced classes. In remote sensing images, certain feature classes (such as water bodies) often have a small but complex proportion. Dice loss can effectively alleviate the problem of class imbalance. It calculates the overlap between the predicted and true segmented regions, encouraging the model to improve its prediction accuracy for the target region.

[0134] 3) Region-level loss (such as Intersection over Union (IoU) loss): To further strengthen region segmentation, an Intersection over Union (IoU) loss is added. This loss can provide guidance on the accuracy of segmented regions, which is particularly important for the recognition of large object regions. The IoU loss encourages the model to focus more on the complete segmentation of the target region and reduce misclassification of background regions.

[0135] The general formula for the composite loss function is:

[0136]

[0137] Among them, y i is the true label of pixel i, p i is the predicted probability of pixel i, λ1, λ2, λ3 are hyperparameters used to adjust the weights of each loss term, and N represents the total number of pixels in the image, which can be adjusted according to the task requirements.

[0138] Finally, the trained image segmentation model can be used to efficiently and accurately segment the target remote sensing image.

[0139] From the description of the above embodiments, those skilled in the art can understand that the present invention provides an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction, which has the following main advantages:

[0140] 1) Improving the computational efficiency of remote sensing image segmentation: This paper significantly reduces computational complexity and improves processing speed while maintaining model accuracy by designing a lightweight feature extraction module and employing lightweight modules such as depthwise separable convolution and a feedforward neural network (FFN). This design enables the model to run on low-computing devices, adapting to the large-scale data processing requirements of remote sensing image segmentation.

[0141] 2) Enhanced Capture of Multi-Scale Features: The multi-receptive field feature interaction module of this invention utilizes a multi-branch structure, including identity mapping, multi-scale convolution, and a wavelet-enhanced Mamba module, to simultaneously capture local details and global dependencies in remote sensing images. Compared to traditional methods, it can more effectively handle complex scenes, objects of varying scales, and rich texture information in remote sensing images, thereby improving segmentation accuracy and robustness.

[0142] 3) Better segmentation accuracy and stability: The present invention combines pixel-level cross entropy loss, Dice loss, and IoU loss to construct a composite loss function, which can effectively solve the category imbalance problem in remote sensing image segmentation and further improve segmentation accuracy.

[0143] In addition, refer to Figure 4 As shown, an embodiment of the present invention also provides an electronic device, which may include a processor 10, a memory 11, a communication bus 12 and a communication interface 13, and may also include a computer program stored in the memory 11 and run on the processor 10. The processor executes the computer program to implement an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction in the above method embodiment.

[0144] In some embodiments, the processor 10 may be composed of an integrated circuit, such as a single packaged integrated circuit or a plurality of packaged integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and a combination of various control chips. The processor 10 is the control core of the electronic device, connecting the various components of the entire electronic device using various interfaces and circuits. It executes or runs programs or modules stored in the memory 11 and calls data stored in the memory 11 to perform various functions of the electronic device and process data.

[0145] The memory 11 may be, for example, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples of storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanical encoding device, and any suitable combination thereof.

[0146] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, electronic devices, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0147] It should be noted that the word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer.

[0148] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0149] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily 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 is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction, characterized in that: The method comprises the following steps: S1. Building an image segmentation model based on multiple lightweight Mamba blocks; each lightweight Mamba block includes: a first lightweight feature extraction module, a multi-receptive field feature interaction module, and a second lightweight feature extraction module connected in sequence; S2. Performing remote sensing image segmentation training on the constructed image segmentation model; S3. Use the trained image segmentation model to perform image segmentation on the target remote sensing image.

2. The image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction according to claim 1, characterized in that: The first lightweight feature extraction module uses depth-separable convolution, normalization, lightweight feedforward neural network and residual connection to perform lightweight feature extraction and fusion processing on the input features, and sends the processed features to the multi-receptive field feature interaction module.

3. The image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction according to claim 2, characterized in that: The multi-receptive field feature interaction module uses a multi-receptive field feature interaction mechanism to process the features output by the first lightweight feature extraction module to extract local features and global features. The processing flow includes: The features output by the first lightweight feature extraction module are divided into three sub-features along the channel dimension and sent to three different feature extraction branches respectively. The three different feature extraction branches use identity mapping, multi-scale convolution, and wavelet transform Mamba to process the sub-features respectively, and then splice the processed sub-features to generate the final output of the multi-receptive field feature interaction module and send it to the second lightweight feature extraction module.

4. The image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction according to claim 3, characterized in that: The second lightweight feature extraction module uses depth-separable convolution, normalization, lightweight feedforward neural network and residual connection to perform lightweight feature extraction and fusion processing on the features output by the multi-receptive field feature interaction module to generate the final output of the lightweight Mamba block.

5. The image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction according to claim 2 or 4, characterized in that: The lightweight feedforward neural network only contains two fully connected layers, and the hidden layer dimension is adjustable.

6. The image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction according to claim 1, characterized in that: In S2, a composite loss function is used to train the image segmentation model. The composite loss function is: Among them, y i is the true label of pixel i, p i is the predicted probability of pixel i, λ1, λ2, λ3 are the hyperparameters used to adjust the weights of each loss term, and N is the total number of pixels in the image.

7. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement an image segmentation method based on lightweight feature extraction and multi-receptive field feature interaction as described in any one of claims 1 to 6.

Images