A dual-branch generative network heterogeneous image change detection method based on content and attribute decoupling
By using a dual-branch generative network based on content and attribute decoupling, the problems of large cross-domain distribution differences and spurious change interference in remote sensing image change detection are solved, achieving high-precision and stable change detection under weak supervision conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN ENG UNIV
- Filing Date
- 2026-04-02
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies for detecting changes in multi-temporal remote sensing images suffer from problems such as large cross-domain distribution differences, susceptibility to spurious changes, and strong dependence on a large number of accurate annotations, resulting in insufficient robustness of detection.
A dual-branch generative network based on content and attribute decoupling is adopted. Shared content features and domain-specific style features are extracted through a cross-domain image translation network to construct a cross-domain consistent intermediate representation. Adaptive updating of soft labels is achieved through cross-feature difference calculation and information entropy evaluation mechanisms, forming an end-to-end closed-loop optimization mechanism.
It significantly improves the accuracy and stability of change detection under weak or unsupervised conditions, reduces false change interference, and enhances detection performance in complex cross-domain scenarios.
Smart Images
Figure CN122347740A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of remote sensing image change detection and cross-domain image translation, specifically to a heterogeneous image change detection method based on a dual-branch generative network with content and attribute decoupling. It is particularly suitable for unsupervised or weakly supervised change detection tasks for multi-temporal and multi-modal remote sensing images. Background Technology
[0002] With the rapid development of remote sensing technology, multi-temporal remote sensing images play a crucial role in urban sprawl monitoring, disaster assessment, land use change analysis, and military reconnaissance. Change detection, as the core task of multi-temporal remote sensing image interpretation, aims to accurately identify areas of land cover change from images acquired at different times. However, due to differences in imaging conditions, sensor variations, and the influence of factors such as illumination and climate change, significant cross-domain differences often exist between multi-temporal images, posing considerable challenges to change detection.
[0003] Traditional change detection methods often rely on pixel differencing, ratio transformation, or manually designed features for analysis, making them highly sensitive to noise and spurious changes, and lacking robustness in complex scenarios. In recent years, deep learning-based change detection methods have made some progress, but most assume that the input images are in a uniform distribution space, performing only simple feature concatenation or differencing operations. They lack explicit modeling of cross-domain distributional differences, easily misclassifying imaging differences as real changes. Furthermore, existing methods typically depend on large amounts of precisely labeled data for supervised training, while pixel-level change labeling is costly and difficult to acquire on a large scale in practical remote sensing applications.
[0004] While some studies have attempted to introduce image translation networks to reduce inter-domain differences, translation and change detection networks are often trained independently, lacking end-to-end collaborative optimization mechanisms, making it difficult to simultaneously improve translation quality and detection performance. Furthermore, existing self-training or pseudo-labeling methods often use fixed thresholds for hard label updates, lacking effective evaluation of model prediction confidence, which can easily lead to error accumulation and training instability.
[0005] Therefore, how to establish a tight coupling relationship between cross-domain image translation and change detection, and how to achieve adaptive updating of soft labels through a confidence-aware mechanism to improve the accuracy and stability of change detection under weak or even unsupervised conditions, has become a key technical problem that urgently needs to be solved. Summary of the Invention
[0006] The purpose of this invention is to address the problems in existing technologies, such as large cross-domain distribution differences, susceptibility to pseudo-change interference, and strong dependence on a large number of accurate annotations. It proposes a heterogeneous image change detection method based on a dual-branch generative network with content and attribute decoupling.
[0007] This invention is achieved through the following technical solution: This invention proposes a heterogeneous image change detection method based on a dual-branch generative network with content and attribute decoupling, the method comprising: Step 1: Input the two-phase images to be detected into the cross-domain image translation network. Decouple the features of the images from different domains through the content encoder and style encoder, extract shared content features and domain-specific style features, and generate bidirectional translated images and self-reconstructed images through the cross-domain reconstruction mechanism to build an intermediate representation with cross-domain consistency. Step 2: The original dual-temporal images and the corresponding cross-domain translated images are feature-stitched together. The cross-feature difference response is constructed through the cross-feature difference calculation module and normalized to obtain the change probability response map, so as to characterize the potential change areas between images at different times. Step 3: Input the change probability response map into the change detection network for multi-scale feature extraction and discrimination, output the change prediction result through the convolution classification module, and construct a joint loss function of translation loss, reconstruction loss, adversarial loss and change detection loss to realize end-to-end collaborative training of translation network and detection network; Step 4: During the training process, a confidence evaluation mechanism based on information entropy is introduced to calculate the pixel-level entropy value of the change probability response map. When the average entropy is lower than the set threshold, the soft label is adaptively updated using an exponential moving average momentum update strategy, thereby forming a closed-loop optimization mechanism and finally outputting the change detection result.
[0008] Furthermore, in step 1, the cross-domain image translation network adopts a dual-generator dual-discriminator structure. Each generator contains a content encoder and a style encoder. Content features are aligned through a shared coding space, and style features of different domains are modeled independently. Cross-domain translated image generation is achieved through content-style reorganization.
[0009] Furthermore, in step 1, the cross-domain reconstruction mechanism includes two forms: self-reconstruction and cross-reconstruction. Self-reconstruction is used to maintain content consistency, while cross-reconstruction is used to enhance the mapping capability between domains. The reconstruction loss is constrained by the L1 loss function.
[0010] Further, in step 2, the cross-feature difference response is calculated as follows: First, the original two-phase images are concatenated along the channel dimension to obtain the feature representation. ; The corresponding translated images are concatenated to obtain the feature representation. ; Calculate the Euclidean distance between the two as the cross-feature difference response:
[0011] The change probability response map is mapped to the [0,1] interval through minimum-maximum normalization.
[0012] Furthermore, in step 3, the change detection network uses a multi-scale convolutional structure to perform local region discrimination on the change probability response map, constructs a change discrimination patch using a sliding window approach, and distinguishes between changed and non-changed regions through binary cross-entropy loss.
[0013] Furthermore, in step 3, the joint loss function is expressed as:
[0014] in, Indicates the losses incurred during reconstruction. Indicates translation loss, Indicating resistance to loss, Indicates the change detection loss. , , , These are the weighting coefficients.
[0015] Furthermore, in step 4, the confidence assessment mechanism uses information entropy as the discriminant metric, and the current probability response diagram shows the change. The entropy is calculated as follows:
[0016] When the average entropy value is lower than the set threshold, the model prediction is considered to have high confidence, triggering a soft label update.
[0017] Furthermore, soft tag updates employ an exponential moving average strategy:
[0018] in, This is the previous round of soft label images. This is the current probability response diagram. This is the momentum coefficient, used to control the smoothness of the update process.
[0019] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the heterogeneous image change detection method based on content and attribute decoupling dual-branch generative network.
[0020] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the heterogeneous image change detection method based on content and attribute decoupling in a dual-branch generative network.
[0021] The beneficial effects of this invention are: This invention proposes a heterogeneous image change detection method based on a dual-branch generative network with content and attribute decoupling. Cross-domain distribution alignment is achieved through content-style decoupling, a change response map is constructed across feature differences to reduce false change interference, and a confidence-aware soft-label momentum update mechanism based on information entropy is introduced to form an adaptive closed-loop training process. This method maintains high detection accuracy and stability even under weak supervision or without precise annotation, significantly improving change detection performance in complex cross-domain scenarios. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0023] Figure 1 This is a flowchart illustrating a heterogeneous image change detection method based on a dual-branch generative network with content and attribute decoupling, as described in this invention.
[0024] Figure 2 This is a framework diagram of a cross-domain image translation network.
[0025] Figure 3 This is a diagram of the network structure for detecting differences in two domains.
[0026] Figure 4 This is a schematic diagram of the unsupervised change detection results in the example. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] Specifically, in combination Figures 1-4 This invention proposes a method for detecting heterogeneous image changes based on a dual-branch generative network with content and attribute decoupling. The method includes: Step 1: Input the two-phase images to be detected into the cross-domain image translation network. Decouple the features of the images from different domains through the content encoder and style encoder, extract shared content features and domain-specific style features, and generate bidirectional translated images and self-reconstructed images through the cross-domain reconstruction mechanism. Ensure the consistency and authenticity of the cross-domain mapping through reconstruction constraints and adversarial constraints, and provide a unified feature space representation for subsequent change analysis. Step 2: The original dual-temporal image and its corresponding cross-domain translated image are concatenated in the channel dimension. The cross-feature difference response is constructed through the cross-feature difference calculation module and normalized to obtain the change probability response map, so as to characterize the potential change area between images at different times. Step 3: Input the change probability response map into the change detection network for multi-scale feature extraction and discrimination, output the change prediction result through the convolution classification module, and at the same time construct a joint loss function of translation loss, reconstruction loss, adversarial loss and change detection loss to realize end-to-end collaborative training of cross-domain translation network and change detection network; Step 4: During the training process, a confidence evaluation mechanism based on information entropy is introduced. The pixel-level entropy value is calculated for the change probability response map as a prediction uncertainty index. When the average entropy is lower than the set threshold, the soft label is adaptively updated using the exponential moving average momentum update strategy, thereby forming a closed-loop optimization mechanism to improve the stability and detection accuracy of the model under weak supervision, and finally output the change detection result.
[0029] Furthermore, in step 1, the cross-domain image translation network adopts a dual-generator dual-discriminator structure. Each generator contains a content encoder and a style encoder. Content features are aligned through a shared coding space, and style features of different domains are modeled independently. Cross-domain translated image generation is achieved through content-style reorganization.
[0030] Furthermore, in step 1, the cross-domain reconstruction mechanism includes two forms: self-reconstruction and cross-reconstruction. Self-reconstruction is used to maintain content consistency, while cross-reconstruction is used to enhance the mapping capability between domains. The reconstruction loss is constrained by the L1 loss function.
[0031] Cross-domain generation is achieved as follows: first, image content features and style features are extracted, cross-domain images are generated through content exchange, and structural consistency is maintained through self-reconstruction constraints.
[0032] Further, in step 2, the cross-feature difference response is calculated as follows: First, the original two-phase images are concatenated along the channel dimension to obtain the feature representation. ; The corresponding translated images are concatenated to obtain the feature representation. ; Calculate the Euclidean distance between the two as the cross-feature difference response:
[0033] The change probability response map is mapped to the [0,1] interval using minimum-maximum normalization to enhance the contrast between the change region and the non-change region.
[0034] Furthermore, in step 3, the change detection network uses a multi-scale convolutional structure to perform local region discrimination on the change probability response map, constructs a change discrimination patch using a sliding window approach, and distinguishes between changed and non-changed regions through binary cross-entropy loss.
[0035] Furthermore, in step 3, the joint loss function is expressed as:
[0036] in, Indicates the losses incurred during reconstruction. Indicates translation loss, Indicating resistance to loss, Indicates the change detection loss. , , , where represents the weight coefficients. Through joint optimization, the cross-domain translation network and the change detection network converge collaboratively within a shared feature space.
[0037] Furthermore, in step 4, the confidence assessment mechanism uses information entropy as the discriminant metric, and the current probability response diagram shows the change. The entropy is calculated as follows:
[0038] When the average entropy value is lower than the set threshold, the model prediction is considered to have high confidence, triggering a soft label update.
[0039] Furthermore, soft tag updates employ an exponential moving average strategy:
[0040] in, This is the previous round of soft label images. This is the current probability response diagram. The momentum coefficient is used to control the smoothness of the update. This mechanism avoids the error accumulation problem caused by hard labels, and achieves a smooth and stable self-supervised optimization process.
[0041] Example 1 To address the problems of large cross-domain distribution differences, severe spurious change interference, and strong dependence on accurate annotation in existing change detection technologies, this invention proposes a heterogeneous image change detection method based on a dual-branch generative network with content and attribute decoupling. By constructing a joint optimization framework of a cross-domain translation network and a change detection network, inter-domain alignment is achieved in the feature space, and a change probability response map is constructed through cross-feature differences. Simultaneously, an uncertainty measurement mechanism based on information entropy is introduced to achieve adaptive momentum updates for soft labels, thereby forming a stable self-supervised closed-loop optimization process.
[0042] Combination Figures 1-4 The method of the present invention includes the following steps: Step 1: Cross-domain image translation and feature decoupling: Translate dual-temporal images and Each generator is input into a cross-domain image translation network. Each generator contains a content encoder. With style encoder This achieves decoupling of content and style features:
[0043] Achieving cross-domain generation through content-style reorganization:
[0044] Simultaneously construct self-reconstruction constraints:
[0045] The reconstruction loss is subject to L1 constraints:
[0046] Adversarial loss is introduced to ensure the realism of the generated images. This step achieves cross-domain distribution alignment, constructing a unified feature space for change detection.
[0047] Step 2: Constructing a change response map based on cross-feature differences: The original dual-temporal images are stitched together by channel concatenation.
[0048] The translated image is then stitched together using channels:
[0049] Calculate cross-feature difference response:
[0050] in Indicates the channel index.
[0051] Minimum-maximum normalization is applied to the differential responses:
[0052] Obtain the probability response plot of the change .
[0053] To enhance spatial consistency, it is Gaussian smoothed:
[0054] The probability response map of this change is used as input to the change detection network.
[0055] Step 3: Change Detection Network and Joint Optimization: Input the change probability map into the change detection network for multi-scale convolutional feature extraction, and output the prediction results through the classification module.
[0056] The change detection loss uses binary cross-entropy:
[0057] The joint total loss is defined as:
[0058] in : Reconstruction losses, Translation loss Combating losses Change detection loss.
[0059] The parameters of the translation network and the detection network are updated simultaneously through end-to-end backpropagation.
[0060] Step 4: Confidence-aware soft label update based on information entropy: To solve the problem of pseudo-label error accumulation, this invention introduces a confidence evaluation mechanism.
[0061] Response diagram to change probability Calculate information entropy:
[0062] Calculate the average entropy:
[0063] when:
[0064] in If the preset threshold is used, the model prediction is considered to have a high confidence level.
[0065] Triggering Exponential Moving Average (EMA) soft tag update:
[0066] in: The previous round of soft labeling, : Current probability response diagram Momentum coefficient. This mechanism enables smooth updates, avoiding unstable oscillations caused by hard thresholds.
[0067] Example 2 This embodiment uses multi-temporal remote sensing data as input and trains the network on a GPU. During training, the Adam optimizer is used to optimize the translation network generator, discriminator, and change detection network. Confidence testing and soft label updates are performed every few rounds, and the following evaluation metrics are calculated during the validation phase: Overall accuracy:
[0068] F1 score:
[0069] Kappa coefficient:
[0070] in:
[0071] Experimental results show that the present invention can significantly reduce the false change detection rate and improve the accuracy of change region identification under conditions of no precise manual annotation or weak supervision.
[0072] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the heterogeneous image change detection method based on content and attribute decoupling dual-branch generative network.
[0073] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the heterogeneous image change detection method based on content and attribute decoupling in a dual-branch generative network.
[0074] The memory in this application embodiment can be volatile memory or non-volatile memory, or it can include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.
[0075] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).
[0076] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.
[0077] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above methods.
[0078] The above provides a detailed description of the heterogeneous image change detection method based on content and attribute decoupling using a dual-branch generative network proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. A method for detecting heterogeneous image changes based on a dual-branch generative network with content and attribute decoupling, characterized in that, The method includes: Step 1: Input the two-phase images to be detected into the cross-domain image translation network. Decouple the features of the images from different domains through the content encoder and style encoder, extract shared content features and domain-specific style features, and generate bidirectional translated images and self-reconstructed images through the cross-domain reconstruction mechanism to build an intermediate representation with cross-domain consistency. Step 2: The original dual-temporal images and the corresponding cross-domain translated images are feature-stitched together. The cross-feature difference response is constructed through the cross-feature difference calculation module and normalized to obtain the change probability response map, so as to characterize the potential change areas between images at different times. Step 3: Input the change probability response map into the change detection network for multi-scale feature extraction and discrimination, output the change prediction result through the convolution classification module, and construct a joint loss function of translation loss, reconstruction loss, adversarial loss and change detection loss to realize end-to-end collaborative training of translation network and detection network; Step 4: During the training process, a confidence evaluation mechanism based on information entropy is introduced to calculate the pixel-level entropy value of the change probability response map. When the average entropy is lower than the set threshold, the soft label is adaptively updated using an exponential moving average momentum update strategy, thereby forming a closed-loop optimization mechanism and finally outputting the change detection result.
2. The method according to claim 1, characterized in that, In step 1, the cross-domain image translation network adopts a dual generator and dual discriminator structure. Each generator contains a content encoder and a style encoder. Content features are aligned through a shared coding space, and style features of different domains are modeled independently. Cross-domain translated image generation is achieved through content-style reorganization.
3. The method according to claim 2, characterized in that, In step 1, the cross-domain reconstruction mechanism includes two forms: self-reconstruction and cross-reconstruction. Self-reconstruction is used to maintain content consistency, while cross-reconstruction is used to enhance the mapping capability between domains. The reconstruction loss is constrained by the L1 loss function.
4. The method according to claim 1, characterized in that, In step 2, the cross-feature difference response is calculated as follows: First, the original two-phase images are concatenated along the channel dimension to obtain the feature representation. ; The corresponding translated images are concatenated to obtain the feature representation. ; Calculate the Euclidean distance between the two as the cross-feature difference response: The change probability response map is mapped to the [0,1] interval through minimum-maximum normalization.
5. The method according to claim 1, characterized in that, In step 3, the change detection network uses a multi-scale convolutional structure to distinguish local regions of the change probability response map, constructs a change discrimination patch using a sliding window approach, and distinguishes between changed and non-changed regions through binary cross-entropy loss.
6. The method according to claim 1, characterized in that, In step 3, the joint loss function is expressed as: in, Indicates the losses incurred during reconstruction. Indicates translation loss, Indicating resistance to loss, Indicates the change detection loss. , , , These are the weighting coefficients.
7. The method according to claim 1, characterized in that, In step 4, the confidence assessment mechanism uses information entropy as the discriminant metric, and the current probability response diagram shows the change. The entropy is calculated as follows: When the average entropy value is lower than the set threshold, the model prediction is considered to have high confidence, triggering a soft label update.
8. The method according to claim 7, characterized in that, Soft tag updates employ an exponential moving average strategy: in, This is the previous round of soft label images. This is the current probability response diagram. This is the momentum coefficient, used to control the smoothness of the update process.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-8.
10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-8.