Dual-mask output lightweight remote sensing image change detection system and method, electronic device and storage medium

Abstract

The application discloses a lightweight remote sensing image change detection system and method based on double-mask output, an electronic device and a storage medium, and belongs to the technical field of remote sensing image processing. The method acquires double-temporal remote sensing images of the same geographical area, inputs the double-temporal remote sensing images into a shared-weight twin lightweight encoder after preprocessing, aligns cross-sensor and cross-season feature distribution by using a cross-sensor adaptive normalization layer, enhances real changes and suppresses pseudo changes caused by seasons, light and textures by using a difference amplification and pseudo change suppression module at each scale, respectively recovers change boundaries and cloud areas by using a double-path decoupling decoder, and outputs a binary change mask and a cloud mask by using a cloud decision module. In the training, a loss design of masking and layered punishment is used for the cloud area to improve the robustness under complex conditions such as cloud interference. The system model parameter quantity is not more than 10.5M, not more than 21.0M, supports export and quantization, and is suitable for lightweight deployment of edge devices.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to a lightweight remote sensing image change detection system, method, electronic device, and storage medium with dual-mask output, suitable for lightweight edge deployment. Background Technology

[0002] Remote sensing image change detection is a core technology for identifying surface changes in the same geographic area across different time periods. Existing methods mainly include traditional difference / ratio methods, threshold segmentation methods based on image registration, and end-to-end methods based on deep learning (e.g., , , , (Network-like structures). Currently, in practical applications of remote sensing image change detection, it is often necessary to deal with remote sensing data of different resolutions collected by different satellites. At the same time, spurious changes caused by factors such as imaging angle, weather conditions, seasonal changes, and objects with very chaotic appearances in remote sensing images are still difficult to distinguish. Therefore, under these circumstances, the robustness, accuracy, real-time performance, and suppression of spurious changes of existing algorithms cannot achieve ideal results. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a lightweight and robust remote sensing image change detection method. This method achieves high precision, high efficiency, and high robustness in detecting false changes in remote sensing images from different satellites and at different resolutions, as well as in various scenarios such as different imaging angles, complex weather conditions, and seasonal changes. The method is adaptable to different satellites and resolutions, achieves high accuracy and high efficiency in complex scenarios such as cloud interference, and supports deployment on edge devices. Quantification.

[0004] The technical solution adopted in this invention is: a lightweight remote sensing image change detection system with dual mask output, the system including an image preprocessing module, a twin encoder with cross-sensor phase invariance feature extraction capability, a differential amplification and pseudo-change suppression module, a dual-path decoupling decoder and a cloud layer decision module;

[0005] The preprocessing module processes the input and The image undergoes radiometric calibration, atmospheric correction, normalization, and subpixel-level geometric registration.

[0006] Twin encoder and The layers employ a lightweight backbone network with shared weights, and an adaptive normalization layer across sensors is set after each convolutional block;

[0007] For the channel Features Standardization should be carried out first. ,in Let be the statistics for this channel; It is a vector composed of global statistics of the input image, processed by a micro-network. Generate affine parameters ,get

[0008]

[0009] according to Dynamically generated To align differences in distribution across sensors / seasons;

[0010] The differential amplification and pseudo-change suppression module is implemented in parallel at each scale of the encoder and contains two branches. The differential amplification branch subtracts the two features element by element and enhances the small real changes through a depthwise separable convolutional residual network. The pseudo-change suppression branch uses cross-channel self-attention to calculate the internal correlation and generate a suppression mask to attenuate the differential signal in periodic or highly textured similar regions.

[0011] The dual-path decoupled decoder and cloud decision module employ an upsampling structure with spatial attention and cross-layer connections with shallow features to recover boundary details; the cloud path uses lightweight decoding blocks to outline the contours of cloud-occluded areas; the cloud decision module is small. The combined decoding features of the two paths and local statistical outputs three classification results: "change", "unchanged" and "cloud".

[0012] Furthermore, the system meets edge deployment constraints, with the total number of network parameters not exceeding [a certain threshold]. Parameters; when the input size is Under pixel conditions, the inference computation does not exceed Sub-floating-point operations; supports Export and Quantitative deployment.

[0013] A detection method for a lightweight remote sensing image change detection system with dual-mask output, the method comprising the following steps:

[0014] S1. Acquire the first and second phase remote sensing images of the same geographic area;

[0015] S2. Perform radiometric calibration, atmospheric correction, normalization, and subpixel-level geometric registration on the image;

[0016] S3. Use a shared-weighted twin encoder to extract features from two temporal images at multiple scales and perform feature difference and fusion at at least one scale; input the difference-fused features into the decoder to generate change detection results;

[0017] S4. After each convolutional block, a cross-sensor adaptive normalization layer is set to dynamically generate affine parameters based on the input global statistics to align the feature distributions across sensors and seasons; and the global statistical vector is used as the basis for this normalization. The fusion weights at various scales are obtained through micro-network mapping. Used for local-global feature fusion control;

[0018] S5, the vector Further adaptive adjustment of thresholds and loss weights for cloud decision-making modules; differential amplification branches and pseudo-change suppression branches are set in parallel at each scale to enhance real changes and suppress pseudo-changes caused by seasons, illumination and texture;

[0019] S6. A dual-path decoupled decoder and cloud decision module are used to output a binary change mask map that identifies surface changes and a cloud mask map that indicates areas that cannot be reliably judged due to occlusion or interference. During the training phase, a loss design of masking and layering penalty is used for cloud areas.

[0020] Furthermore, the dual-path decoupled decoder simultaneously generates a binary transformation mask and a cloud mask; the output range of the binary transformation mask is dimensionless [0,1]; the output range of the cloud mask is dimensionless [0,1].

[0021] Furthermore, in step S4, the cross-sensor adaptive normalization layer generates channel affine parameters through a micro-network based on the mean, variance, and histogram statistics of the input image, satisfying... Dimensionless Dimensionless boundary conditions; for each channel Standardization characteristics Perform affine transformation To achieve distribution alignment;

[0022] The fusion weights for each scale are derived from the global statistical vector. Through two layers and Generate, satisfy ,in For lightweight networks, for function.

[0023] Furthermore, in step S6, the layered penalty weights of the clouds are determined by... Exported cloud occlusion metrics Perform adaptive weighting to satisfy , Cloud Decision Threshold exist Within range Monotonically increasing;

[0024] Furthermore, the training loss consists of two parts: the main cross-entropy loss for sharp regions and the cloud layer stratification penalty loss; the main loss is only applied to the sharp region mask. Time calculation; cloud penalty loss in cloud region masking Weights are applied to predictions of "change" and "invariance". Dimensionless, weighting the "cloud" prediction in the clear region. Dimensionless.

[0025] Furthermore, the twin encoder employs a lightweight structure, incorporating depthwise separable convolution and a lightweight attention mechanism; the total number of network parameters does not exceed [a certain threshold]. Parameters; when the input size is Under pixel conditions, the inference computation does not exceed This is a floating-point operation.

[0026] An electronic device includes a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the above-described change detection method.

[0027] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described change detection method.

[0028] a. Technical Solution: The method includes: image preprocessing; multi-scale feature extraction using a shared-weighted twin lightweight encoder; feature differencing and fusion at the same scale, employing local-global semantic fusion to suppress pseudo-changes; dual-path decoupled decoder generating binary change masks and cloud masks; loss function applying masking and layered penalty to cloud regions; and support for... Export and Quantification.

[0029] b. Core Innovation: Proposes a statistically driven cross-sensor adaptive normalization method. Collaborative framework with dual-mask decoding, using global statistical vectors of the input image At its core, in addition to performing channel affine alignment after each convolutional block, it also obtains the fusion weights for each scale through a small network mapping. Control local-global feature fusion; and It participates in cloud layer decision-making and loss weighting to achieve robust three-class classification output under complex conditions such as clouds and fog. This statistical coupling mechanism unifies normalization, fusion, and decision / loss under a single control variable. Under these circumstances, a tightly coupled core innovation point is formed. Attached Figure Description

[0030] Figure 1 This is a flowchart of the change detection method in complex scenarios described in this invention;

[0031] Figure 2 This is a detailed flowchart of the change detection network of the present invention;

[0032] Figure 3 This is the training framework for the lightweight remote sensing image change detection network with dual mask output described in this invention. Detailed Implementation

[0033] A lightweight remote sensing image change detection method with dual-mask output includes: acquiring a first-temporal remote sensing image and a second-temporal remote sensing image of the same geographic area; performing radiometric calibration, atmospheric correction, normalization, and subpixel-level geometric registration on the images; extracting multi-scale features from the two-temporal images using a shared-weight twin encoder and performing feature difference and fusion at at least one scale; and inputting the difference-fused features into a decoder to generate change detection results.

[0034] A cross-sensor adaptive normalization layer is set after each convolutional block to dynamically generate affine parameters based on the input global statistics to align the feature distributions across sensors and seasons; and the global statistical vector is used as the basis for this normalization. The fusion weights at various scales are obtained through micro-network mapping. Used for local-global feature fusion control; the vector Further adaptive adjustments to the threshold and loss weights of the cloud decision module are made; differential amplification branches and pseudo-change suppression branches are set in parallel at each scale to enhance real changes and suppress pseudo-changes caused by seasons, illumination, and texture; a dual-path decoupled decoder and cloud decision module are used to output a binary change mask map that identifies surface changes and a cloud mask map that identifies areas that cannot be reliably judged due to occlusion or interference; during the training phase, a loss design of masking and layering penalty is used for cloud areas to improve robustness under complex meteorological conditions.

[0035] In some specific embodiments, the decoder simultaneously generates a binary transformation mask and a cloud mask; wherein the output range of the binary transformation mask is dimensionless [0,1]; and the output range of the cloud mask is dimensionless [0,1].

[0036] In some specific implementations, the cross-sensor adaptive normalization layer generates channel affine parameters through a micronetwork based on the mean, variance, and histogram statistics of the input image, satisfying... Dimensionless Dimensionless boundary conditions; for each channel Standardization characteristics Perform affine transformation To achieve distribution alignment.

[0037] In some specific implementations, the fusion weights for each scale are derived from the global statistical vector. Through two layers and Generate, satisfy ,in For lightweight networks, for function.

[0038] In some specific implementations, the cloud layer stratification penalty weights are determined by... Exported cloud occlusion metrics Perform adaptive weighting to satisfy , Cloud Decision Threshold exist Within range Monotonically increasing.

[0039] In some specific implementations, the training loss consists of two parts: the main cross-entropy loss for sharp regions and the cloud layer stratification penalty loss; wherein the main loss is only applied to the sharp region mask. Time calculation; cloud penalty loss in cloud region masking Weights are applied to predictions of "change" and "invariance". Dimensionless, weighting the "cloud" prediction in the clear region. Dimensionless.

[0040] In some specific implementations, training employs a course-based learning strategy. Initially, only cloudless samples are used; in the later stages, cloud-interference samples are gradually introduced, and the cross-sensor adaptive normalization layer and cloud decision-making module are trained jointly. The total number of training rounds is [number missing]. .

[0041] In some specific embodiments, the twin encoder employs a lightweight structure, incorporating depthwise separable convolution and a lightweight attention mechanism; the total number of network parameters does not exceed [a certain threshold]. Parameters; when the input size is Under pixel conditions, the inference computation does not exceed This is a floating-point operation.

[0042] In some specific implementations, the model supports Export and Quantitative deployment; under standard test set conditions, The absolute decrease in overall precision after quantization shall not exceed (Dimensionless).

[0043] The change detection method as described in claim 1 is characterized in that: cross-resolution consistency is achieved through scale normalization or dynamic stride mechanism combined with pyramid feature representation, applicable to ground sampling distances of [missing information]. to Remote sensing images (including boundaries).

[0044] A remote sensing image change detection system includes: an encoder module, a feature fusion module, a decoder module, and a cloud decision module; the encoder module is used to extract multi-scale features from dual-temporal remote sensing images; the feature fusion module is used to perform feature difference and fusion at at least one scale; and the decoder module is used to generate change detection results.

[0045] The encoder module sets a cross-sensor adaptive normalization layer after each convolutional block; the feature fusion module includes a differential amplification branch and a pseudo-change suppression branch in parallel; the decoder module is a dual-path structure that recovers the change boundary and the cloud region respectively; the cloud decision module integrates the two-path decoding features and local statistical features to output three categories: "change", "unchanged", and "cloud".

[0046] In some specific implementations, the system meets edge deployment constraints, and the total number of network parameters does not exceed [a certain limit]. Parameters; when the input size is Under pixel conditions, the inference computation does not exceed Sub-floating-point operations; supports Export and Quantitative deployment.

[0047] An electronic device includes a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the above-described change detection method.

[0048] A computer-readable storage medium having a computer program stored thereon, characterized in that: when the computer program is executed by a processor, it implements the above-described change detection method.

[0049] Example 1

[0050] a. System Overall Architecture

[0051] The system includes: an image preprocessing module, a twin encoder with cross-sensor phase-invariant feature extraction capability, a differential amplification and pseudo-change suppression module, a dual-channel decoupled decoder, and a cloud decision module.

[0052] The preprocessing module processes the input and The image undergoes radiometric calibration, atmospheric correction, normalization, and subpixel-level geometric registration.

[0053] Twin encoder and The layers employ a lightweight backbone network with shared weights, and cross-sensor adaptive normalization is set after each convolutional block. Layer. For channels Features Standardization should be carried out first. ,in Let be the statistics for this channel; The vector is composed of global statistics (mean, variance, histogram) of the input image, processed by a micro-network. Generate affine parameters ,get

[0054]

[0055] according to Dynamically generated To align with differences in distribution across sensors / seasons and improve temporal invariance.

[0056] The differential amplification and pseudo-change suppression module comprises two branches operating in parallel at each scale of the encoder. The differential amplification branch subtracts the two features element-wise and then passes them through a depthwise separable convolutional residual network to enhance subtle real changes; the pseudo-change suppression branch utilizes cross-channel self-attention to calculate internal correlations and generate a suppression mask to attenuate the differential signal in periodic or highly textured similar regions.

[0057] The dual-path decoupled decoder and the cloud decision change boundary path employ an upsampling structure with spatial attention and cross-layer connections with shallow features to recover boundary details; the cloud path uses lightweight decoding blocks (bilinear interpolation and...) (Convolution) outlines the contours of the cloud-occupied area; the cloud decision module is small. The system combines the decoding features from both channels with local statistics (brightness, texture entropy) to output three classification results: "change", "unchanged", and "cloud".

[0058] b. Training and Reasoning

[0059] The learning strategy employs a course-based approach, first training basic skills with clear samples, and then introducing cloud interference and spurious change samples for joint training. With the decision module. Make the clear area mask. Cloud area mask Three types of soft maximum prediction True value The main loss is calculated by multiplying the cross-entropy in the clear region:

[0060]

[0061] Encourage the output of "cloud" in cloud areas, and suppress the output of "cloud" in clear areas:

[0062]

[0063] The total loss is

[0064]

[0065] Parameter updates are performed using stochastic gradient descent with momentum:

[0066]

[0067] During inference, a single forward propagation directly yields the final classification graph containing three types of information, supporting... and Quantitative deployment.

[0068] c. Terminology Definitions and Symbol Conventions

[0069] The following terms and symbols are used in this manual: ground sampling distance This represents the distance to the ground per pixel, in units of... Pyramid hierarchy These correspond to the feature map resolution and the number of channels in each layer, respectively. kernel size stride (in the context of convolution) For stride, in a statistical context (denoted as the global statistical vector); upsampling operation Bilinear interpolation is used; For channel feature, As a standardized feature, Channel statistics; affine parameters Affine output Difference characteristics With fusion features satisfy ,in Lightweight through function Generation; Three types of probabilities Clear area mask Cloud area mask Image size ;loss and Total loss Threshold ; Optimization and updates in progress For momentum, For learning rate, For parameters.

[0070] d. Detailed parameters of the model architecture

[0071] The encoder consists of four stages, each containing two depthwise separable convolutional blocks and one cross-sensor adaptive normalization layer, with the number of channels adjusted accordingly. Incremental; block structure is First block of the phase Downsampling. The differential amplification branch performs element-wise subtraction on the features of the same layer at two different times, and then connects them with the residuals through two depthwise separable convolutions. The pseudo-change suppression branch employs cross-channel self-attention, first performing global average pooling to obtain the channel description. Through two layers and Obtain weights Suppress output as Local-global fusion follows

[0072]

[0073] in It is a learnable scalar, and is composed of statistical vectors. generate: , ,in Lightweight , for Function. The decoder has a dual-path structure: the changing boundary path executes at each layer. and Convolution + Spatial Attention; Cloud Pathway A minimal block of convolution. Ultimately passed through... Convolution and The three-class probability map is obtained.

[0074] e. Lightweight strategy and compression ratio

[0075] A lightweight combination of depthwise separable convolutions, pointwise convolutions, group convolutions, and cross-channel attention is used to replace heavy convolutions; while maintaining the baseline model... Under the premise of parameter accuracy, compress to no more than Parameters, compression ratio not less than ;use Quantization and quantization perception training, with a calibration sample size of no less than 1000 images, and the absolute decrease in accuracy after quantization not exceeding [a certain value]. .

[0076] f. Data preprocessing and augmentation

[0077] Radiometric calibration and atmospheric correction are performed on each pair of temporal images; histogram matching is used to achieve brightness alignment; and the following methods are employed. Subpixel-level registration is performed using this method, with the termination condition being that the iteration gain is less than [a certain value]. Or the number of iterations reaches Next; normalized to Augmentation includes random rotation, horizontal / vertical flipping, scaling, and color jitter.

[0078] g. Training configuration and hyperparameters

[0079] The optimizer is a driving force Initial learning rate 0.01, momentum 0.9, weight decay Batch size 16, training rounds Learning rate scheduling uses cosine annealing; random seed; mixed precision training with a gradient pruning threshold of 1.0. Total loss. ,in and The definition is the same as before. Two types of change detection thresholds are used: a fixed threshold and an adaptive threshold. Used for binary masks; cloud thresholding Used for cloud masking; the adaptive scheme can target clear areas. Obtained by applying Otsu's method And retain Unchanged. Under complex weather conditions, based on... Exported cloud occlusion metrics Adaptive weighting; cloud threshold It can be set within the range of [0.5, 0.7]. Monotonically increasing.

[0080] h. Hardware and Software Environment

[0081] Training and inference are performed in a single-GPU environment, where the GPU is... Or of the same level, with video memory not less than ; 8 cores or more; memory not less than The software environment is .

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lightweight remote sensing image change detection system with dual-mask output, characterized in that, The system includes an image preprocessing module, a twin encoder with cross-sensor phase invariance feature extraction capability, a differential amplification and pseudo-change suppression module, a dual-path decoupling decoder, and a cloud decision module. The preprocessing module processes the input first-temporal remote sensing image. With second-phase remote sensing images Perform radiometric calibration, atmospheric correction, normalization, and subpixel-level geometric registration; Twin encoder and The layers employ a lightweight backbone network with shared weights, and a cross-sensor adaptive normalization layer is set after each convolutional block. layer; For the channel Features Standardization should be carried out first. ,in For this channel The statistics; let It is a vector composed of global statistics of the input image, processed by a micro-network. Generate affine parameters ,get ； Layer according to Dynamically generated To align differences in distribution across sensors / seasons; The differential amplification and pseudo-change suppression module is implemented in parallel at each scale of the encoder and contains two branches. The differential amplification branch subtracts the two features element by element and enhances the small real changes through a depthwise separable convolutional residual network. The pseudo-change suppression branch utilizes cross-channel self-attention to calculate internal correlations and generate a suppression mask to attenuate differential signals in periodic or highly textured similar regions; The dual-path decoupled decoder and cloud decision module employ an upsampling structure with spatial attention and cross-layer connections with shallow features to recover boundary details; the cloud path uses lightweight decoding blocks to outline the contours of cloud-occluded areas; the cloud decision module is small. The combined decoding features of the two paths and the local statistical output result are classified into three categories: "change", "no change" and "cloud".

2. The detection system as described in claim 1, characterized in that: The system meets edge deployment constraints, and the total number of network parameters does not exceed [a certain limit]. Parameters; when the input size is Under pixel conditions, the inference computation does not exceed Sub-floating-point operations; supports Export and Quantitative deployment.

3. The detection method of a lightweight remote sensing image change detection system with dual-mask output according to claim 1, characterized in that, The method includes the following steps: S1. Acquire the first and second phase remote sensing images of the same geographic area; S2. Perform radiometric calibration, atmospheric correction, normalization, and subpixel-level geometric registration on the first and second temporal remote sensing images. S3. Use a shared-weighted twin encoder to extract features from two temporal images at multiple scales and perform feature difference and fusion at at least one scale; input the difference-fused features into the decoder to generate change detection results; S4. After each convolutional block, a cross-sensor adaptive normalization layer is set to dynamically generate affine parameters based on the input global statistics to align the feature distributions across sensors and seasons; and a global statistical vector is used. The fusion weights at each scale are obtained through micro-network mapping. Used for local-global feature fusion control; S5, the vector Further adaptive adjustment of thresholds and loss weights for cloud decision-making modules; differential amplification branches and pseudo-change suppression branches are set in parallel at each scale to enhance real changes and suppress pseudo-changes caused by seasons, illumination and texture; S6. A dual-path decoupled decoder and cloud decision module are used to output a binary change mask map that identifies surface changes and a cloud mask map that indicates areas that cannot be reliably judged due to occlusion or interference. During the training phase, a loss design of masking and layering penalty is used for cloud areas.

4. The detection method as described in claim 3, characterized in that: The dual-channel decoupled decoder simultaneously generates a binary transformation mask and a cloud mask; the output range of the binary transformation mask is dimensionless [0,1]; the output range of the cloud mask is dimensionless [0,1].

5. The detection method as described in claim 3, characterized in that: In step S4, the cross-sensor adaptive normalization layer generates channel affine parameters through a micro-network based on the mean, variance, and histogram of the input image, satisfying... Dimensionless Dimensionless boundary conditions; for each channel Standardization characteristics Perform affine transformation To achieve distribution alignment; The fusion weights for each scale are derived from the global statistical vector. Through two layers and Generate, satisfy ,in For lightweight networks, for function.

6. The detection method as described in claim 3, characterized in that: In step S6, the layered penalty weights of the cloud layer are determined according to... Exported cloud obstruction metrics Perform adaptive weighting to satisfy the weight Weight Cloud Decision Threshold exist Within range Monotonically increasing.

7. The detection method as described in claim 3, characterized in that: The training loss consists of two parts: the main loss of cross-entropy in the clear regions and the cloud layer penalty loss; the main loss is only applied to the mask in the clear regions. Time calculation; cloud penalty loss in cloud region masking Weights are applied to predictions of "change" and "invariance". Dimensionless, weighting the "cloud" prediction in the clear region. Dimensionless.

8. The detection method as described in claim 3, characterized in that: The twin encoder employs a lightweight structure, incorporating depthwise separable convolution and a lightweight attention mechanism; the total number of network parameters does not exceed [a certain threshold]. Parameters; when the input size is Under pixel conditions, the inference computation does not exceed This is a floating-point operation.

9. An electronic device comprising a processor and a memory, characterized in that: The memory stores a computer program, and when the processor executes the computer program, it implements the detection method as described in any one of claims 3 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the detection method as described in any one of claims 3 to 8.

Images