Image change detection method based on parallel processing of detection and description

Abstract

The invention discloses an image change detection method based on parallel processing of detection and description. The method comprises the following steps: step 1, acquiring a data set; 2, an image change detection model comprises a feature extraction module, a time sequence encoder, and parallel change detection branches and change description branches; 3, training an image change detection model based on the data set; 4, obtaining a to-be-detected dual-temporal remote sensing image; and 5, inputting a to-be-detected dual-temporal remote sensing image into the trained image change detection model to obtain a change detection result and a change description result. According to the method, the parallel change detection branch and the change description branch are designed, the pixel position and the language description of the change area are output at the same time, and the hard association is established between the change area and the language description in the spatial dimension, so that the feature richness required for analyzing the complex scene is reserved, and the balance between the detection precision and the description quality is ensured through labor division; the method is suitable for complex ground feature change scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, and more particularly, to an image change detection method based on parallel processing of detection and description. BACKGROUND

[0002] Remote sensing image change detection and description technology has important value in land use monitoring, urban planning, ecological assessment and disaster warning fields. By accurately identifying feature changes and automatically generating natural language descriptions, this technology can convert massive remote sensing data into understandable change information, significantly improving decision-making efficiency. For example, in urban expansion monitoring, the system can automatically output structured descriptions such as "a certain area adds industrial plant" to provide real-time data support for land space planning and reduce the cost and subjective error of manual interpretation.

[0003] Current mainstream technical solutions can be divided into two categories: one is a serial processing solution based on CNN, which first generates a change mask through a convolutional network and then inputs a language model to generate a description, but is limited by the local receptive field of CNN, making it difficult to model global change correlations; the second is a hybrid architecture based on Transformer, which improves long-range dependency modeling capabilities through self-attention mechanisms, but the detection and description tasks are still processed independently, such as the pure Transformer architecture ViT, which is not sensitive enough to small target changes, and the CNN-Transformer hybrid solution does not achieve feature sharing between tasks.

[0004] The fundamental problem of existing technology is the use of a serial processing flow, which leads to the fragmentation of detection and description tasks: the detection results do not guide the description generation in real time, which may result in semantic deviation, such as misdescribing "building demolition" as "vegetation restoration"; at the same time, the semantic information of the description branch cannot be fed back to the detection branch to optimize it, missing the opportunity to improve accuracy. This one-way, isolated processing mode results in low data utilization and makes it difficult to deal with complex scenarios such as dense building areas and irregular feature changes, and there is an urgent need to break through the bottleneck through a task coordination architecture. SUMMARY

[0005] The technical problem to be solved by the present application is to address the deficiencies in the prior art, providing an image change detection method based on parallel processing of detection and description, which has a simple structure and reasonable design. The change detection branch and the change description branch are designed in parallel, and the pixel position and language description of the change area are output simultaneously, establishing a hard association between the change area and the language description in the spatial dimension. This not only retains the feature richness needed to analyze complex scenarios, but also ensures a balance between detection accuracy and description quality through division of labor, making it suitable for complex feature change scenarios.

[0006] To solve the above technical problems, the technical scheme adopted by the present application is: an image change detection method based on parallel processing of detection and description, characterized by comprising the following steps: Step one, obtaining a data set, which is a dual-time-phase remote sensing image with known change detection results and change description results; Step two, constructing an image change detection model: the image change detection model includes a feature extraction module, a time sequence encoder, and a parallel multi-task branch network; The feature extraction module extracts features from the dual-time-phase remote sensing image to obtain dual-time-phase remote sensing features; The time sequence encoder generates spatio-temporal difference features with spatial position and temporal change through position embedding and difference representation; The parallel multi-task branch network includes a change detection branch and a change description branch; the change detection branch outputs a change detection mask based on dual-time-phase remote sensing features and spatio-temporal difference features; the change description branch receives the change detection mask, crops the feature map block corresponding to the region from the spatio-temporal difference feature map according to the bounding box of the change detection mask, generates the weight of the feature map block based on the change detection mask, and fuses the feature map block and the weight to obtain mask-guided description features and generate description text; Step three, training the image change detection model based on the data set to obtain a trained image change detection model; Step four, obtaining a dual-time-phase remote sensing image to be detected; Step five, inputting the dual-time-phase remote sensing image to be detected into the trained image change detection model to obtain change detection results and change description results.

[0007] The image change detection method based on parallel processing of detection and description, characterized in that: the change description branch receives the change detection mask, crops the feature map block corresponding to the region from the spatio-temporal difference feature map according to the bounding box of the change detection mask, generates the weight of the feature map block based on the change detection mask, fuses the feature map block and the weight to obtain mask-guided description features, and generates description text, specifically including: RoIPool crops the feature map block corresponding to the region from the spatio-temporal difference feature map according to the bounding box of the change detection mask to obtain a description feature map; attention weights are generated based on the change detection mask through an attention mechanism; the description feature map and the attention weights are multiplied pixel by pixel to obtain mask-guided description features; and a Transformer decoder generates description text based on the mask-guided description features.

[0008] The image change detection method based on parallel processing of detection and description, characterized in that: the attention weights are generated based on the change detection mask, specifically including: the change detection mask is interpolated to adjust to a target size, and then a softmax activation function is used to generate dynamic attention weights.

[0009] The image change detection method based on parallel processing of detection and description, characterized in that: the change detection branch is based on dual-time-phase remote sensing features and spatio-temporal difference features, and outputs a change detection mask, specifically including: Fusing the dual-time-phase features and the temporal difference features to generate fusion features, performing nonlinear transformation on the fusion features to generate a probability map, and generating the change detection mask through dynamic threshold binarization.

[0010] The image change detection method based on parallel processing of detection and description, characterized in that: the feature extraction module extracts features from dual-time-phase remote sensing images to obtain dual-time-phase remote sensing features, specifically including: Using a CNN network to extract detailed features of the dual-time-phase remote sensing images to obtain a CNN feature map; Through the multi-scale asymmetric convolution of the edge information guided module and the edge detection, the CNN feature map is cooperatively acted on to strengthen the geometric boundaries of the change area, and an EIG enhanced feature map is obtained; The dynamic deformable Transformer adjusts the attention area through a dynamic offset mechanism adaptive to the image content, obtains the global semantics of irregular change targets in the EIG enhanced feature map, and outputs the dual-time-phase remote sensing features.

[0011] The image change detection method based on parallel processing of detection and description, characterized in that: the temporal encoder generates spatio-temporal difference features with spatial position and temporal change through position embedding and difference representation, specifically including: The position embedding component adds a learnable two-dimensional position encoding to the dual-time-phase remote sensing features to generate dual-time-phase position enhanced features; The difference representation component generates semantic-level difference features based on the dual-time-phase position enhanced features; The feature fusion component fuses the dual-time-phase position enhanced features and the semantic-level difference features in terms of time sequence and proportion to generate spatio-temporal difference features.

[0012] The image change detection method based on parallel processing of detection and description, characterized in that: the difference representation component generates semantic-level difference features based on the dual-time-phase position enhanced features, specifically including: calculating the basic difference of the dual-time-phase position enhanced features through a basic difference operation, quantifying the increase and decrease of the feature values, preliminarily marking the potential change area, obtaining the basic difference features, performing nonlinear transformation on the basic difference features through a learnable nonlinear convolution, suppressing irrelevant interference, enhancing the difference signal of the real change area, and obtaining processed difference features; performing joint feature extraction on the processed difference features, capturing the correlation rules between features, filtering isolated noise points, and generating semantic-level difference features.

[0013] The image change detection method based on parallel processing of detection and description has the characteristics that an image change detection model is trained based on a data set to obtain a trained image change detection model, and specifically includes the following steps: The data set is divided into a training set and a verification set; the training set is input into the image change detection model to train the image change detection model, and when the total loss function value of the image change detection model no longer decreases, the training is stopped to obtain the trained image change detection model; the verification set is used to verify the trained image change detection model, and if the verification is passed, the trained image change detection model is obtained; if the verification is not passed, the data set is replaced and the training is restarted; The image change detection method based on parallel processing of detection and description has the characteristics that the total loss function of the image change detection model , wherein represents the loss function of the change description branch, represents the loss function of the change detection branch, represents the first balancing factor, represents the second balancing factor.

[0014] The image change detection method based on parallel processing of detection and description has the characteristics that the loss function of the change description branch and the loss function of the change detection branch are both cross-entropy loss functions.

[0015] Compared with the prior art, the present application has the following advantages: 1. The structure of the present application is simple, reasonable in design, easy to implement and use.

[0016] 2. The present application designs parallel change detection branches and change description branches, simultaneously outputs the pixel positions and language descriptions of the change regions, and the two complement each other to form complete change information output, which can be understood by ordinary users and can better serve emergency, planning and other scenes, such as monitoring of newly added building groups in urban expansion, disappearance of buildings in demolition area monitoring, and planning changes in construction project acceptance.

[0017] 3. The present application simultaneously inputs the spatio-temporal difference features output by the time encoder into the change detection branch and the change description branch, simultaneously provides the change detection branch with the time dimension change clues and the change description branch with the time sequence context, avoids repeated modeling of time information, provides a unified feature basis for the two branches, and guarantees the time sequence logical consistency of the detection result and the description result.

[0018] 4、The application takes the binary change mask output by the change branch as a benchmark reference of spatial semantics, embeds the features of the change description branch through one-way fusion, establishes a hard association between the change area and the language description in the spatial dimension, makes the description text generation pay more attention to the change area marked by the binary change mask, maintains the objectivity of detection, and improves the pertinence of description.

[0019] 5、In the application, the change detection branch and the change description branch process the detection and description tasks in parallel, not only reduces the training redundancy caused by repeated calculation of features, but also finally has the pixel-level detection reliability and maintains the readability and intuitiveness of language description, effectively solves the change detection and change description problems in complex scenes such as dense building areas, industrial expansion, and suburban development.

[0020] In summary, the application designs the parallel change detection branch and change description branch, simultaneously outputs the pixel position of the change area and the language description, establishes a hard association between the change area and the language description in the spatial dimension, retains the feature richness required for analyzing complex scenes, and balances the detection accuracy and description quality through division, which is suitable for complex feature change scenes.

[0021] The technical solutions of the application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The method flowchart of the application.

[0023] Figure 2 The structure diagram of the image change detection model of the application. DETAILED DESCRIPTION

[0024] The application will be further described in detail below with reference to the drawings and embodiments.

[0025] It should be noted that the embodiments and features in the embodiments in the application can be combined with each other without conflict. The application will be described in detail below with reference to the drawings and embodiments.

[0026] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof.

[0027] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting, unless the contrary is indicated. For example, the description of the embodiments of the present application is intended to include both the described embodiments and equivalents thereof. The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense.

[0028] For the purposes of the description hereinafter, the terms "upper", "bottom", "lower", "over", "under" and "on" as may be used herein, are to be construed as taken with reference to the orientation of the figures. Relative terms, such as "horizontal", "vertical", "top", "bottom", "side" and the like, are used herein to describe the orientation of the components as shown in the figures. It will be understood that the orientation of the components as shown in the figures is merely for illustrative purposes and in no way is intended to limit the scope of the application. The terms "first", "second", "third", "fourth", etc. as used herein are meant to be labels for identifying various elements and are not intended to, and should not be construed to, have any other meaning or significance.

[0029] As shown in FIGS. 1 and 2, an image change detection method based on detecting and describing parallel processing according to the present application includes the following steps: Figure 1 and Figure 2 As shown in FIGS. 1 and 2, an image change detection method based on detecting and describing parallel processing according to the present application includes the following steps: Step one, obtaining a data set, the data set is a double-time-phase remote sensing image with known change detection results and change description results.

[0030] The data set is a double-time-phase remote sensing image set covering urban areas, including a plurality of double-time-phase remote sensing images containing typical features such as buildings and roads, and the annotation of the data set includes change detection results and change description results, i.e. pixel-level change mask and natural language description are annotated at the same time. In one possible embodiment, the double-time-phase remote sensing image set is the LEVIR-CD data set, the time span is 5-15 years, the image size is 256x256, and the T1 time phase and the T2 time phase in the double-time-phase remote sensing image can be 2018 and 2020.

[0031] Step two, constructing an image change detection model: the image change detection model includes a feature extraction module, a time series encoder, and a parallel multi-task branch network, and the parallel multi-task branch network includes a change detection branch and a change description branch. In one possible embodiment, the task of the image change detection model is to identify the increase or disappearance of buildings in urban areas and describe these changes in natural language, and is applied to building change scenarios in urban remote sensing images. Since the description results of natural language can be understood by ordinary users, it can better serve emergency, planning and other scenarios, such as newly added building groups in urban expansion monitoring, building disappearance in demolition area monitoring, and planning changes in construction project acceptance.

[0032] The feature extraction module extracts features from the dual-time-phase remote sensing images to obtain dual-time-phase remote sensing features, specifically including: using a CNN network to extract detailed features of the dual-time-phase remote sensing images to obtain a CNN feature map; through the multi-scale asymmetric convolution of the edge information guided module and the edge detection, the CNN feature map is cooperated to strengthen the geometric boundary of the change area to obtain an EIG enhanced feature map; the dynamic deformable Transformer adjusts the attention area through the dynamic offset mechanism adaptive to the image content to obtain the global semantics of the irregular change target in the EIG enhanced feature map, and outputs the dual-time-phase remote sensing features to provide the dual-time-phase remote sensing features with clear edges and complete targets for subsequent tasks.

[0033] In actual use, the CNN network adopts the classic architecture of ResNet to extract texture, color and other detailed features from the dual-time-phase remote sensing images respectively, to obtain CNN feature maps, which are used to determine information such as building surface material, roof structure, building structure, ground material, etc., for example, whether it is a ceramic tile wall or a concrete wall, a flat roof or a sloping roof, a low-rise dwelling or a high-rise building, an asphalt pavement or a cement pavement, etc. The edge information guidance module, i.e., the EIG module, Edge Information Guidance, captures different size change areas through multi-scale asymmetric convolution, and then strengthens the change edges through an edge detection module to extract edge features and obtain EIG enhanced feature maps, which are used to determine the building boundary, the boundary between buildings and open land, the boundary between buildings and roads, and the boundary between farmland and construction land, for example, the outline of a house wall, the boundary between a residential area and an undeveloped plot, the connecting edge between a building and a street, and the transition area between farmland and a factory building, etc. The multi-scale asymmetric convolution includes asymmetric kernels such as 3x3 and 5x1. The dynamically deformable Transformer, i.e., the DDaT module, Dynamically Deformable Transformer, effectively solves the target alignment problem in dual-time-phase remote sensing images through a dynamic offset attention mechanism that is adaptive to image content. The position of the convolution kernel is not fixed, and the attention area can be dynamically adjusted according to the image content, for example, if there is a displacement between the building group in the T1 time-phase remote sensing image and the T2 time-phase remote sensing image, the DDaT module automatically calculates the offset to achieve accurate feature alignment of the same target. Specifically, the DDaT module analyzes irregular change targets in the EIG enhanced feature maps to extract their global semantic information, and finally outputs dual-time-phase remote sensing features with stronger representation ability, providing more reliable feature support for subsequent change detection tasks.

[0034] It should be noted that the dynamic offset attention mechanism of the DDaT module effectively fuses local features and global context information, establishes long-range dependencies using a self-attention mechanism, and captures large-scale features such as the spatial distribution of building groups and the connectivity of road networks, thereby making up for the limitations of traditional CNN networks that can only handle local areas, and being suitable for change scenarios that require global understanding, for example: when it is necessary to identify scattered farmland in the T1 time-phase remote sensing image as a concentrated industrial park in the T2 time-phase remote sensing image, for this type of change pattern involving large-area spatial reorganization, the DDaT module can analyze both local architectural features and overall layout changes, accurately determine the regional functional transformation, and will not be misjudged due to similar local features, significantly improving the robustness of change detection in complex scenarios.

[0035] The time sequence encoder generates space-time difference features with spatial position and time variation through position embedding and difference representation, solves the problem of "long time sequence and less available training data" of double time phase remote sensing images, enhances the understanding of space-time variation of the image change detection model, and accurately extracts the time sequence difference, such as the change information of newly added or disappeared buildings between T1 time phase remote sensing images and T2 time phase remote sensing images.

[0036] In actual use, the position embedding component binds the corresponding spatial coordinate information for each pixel or feature point in the double time phase remote sensing feature, that is, adds a learnable two-dimensional position code, generates a double time phase position enhanced feature, and since the double time phase remote sensing feature is extracted from the T1 time phase image and the T2 time phase image selected from the time sequence data, the double time phase remote sensing feature itself has time node information, and after two-dimensional position coding, the double time phase remote sensing feature also obtains spatial information, enhancing the understanding of space and time of the image change detection model. The difference representation component generates semantic level difference features based on the double time phase position enhanced feature: calculates the basic difference of the double time phase position enhanced feature through the basic difference operation, quantifies the increase and decrease of the feature value, preliminarily marks the potential change area, obtains the basic difference feature, and performs nonlinear transformation on the basic difference feature through the learnable nonlinear convolution, suppresses irrelevant interference, enhances the difference signal of the real change area, and obtains the processed difference feature; joint feature extraction is performed on the processed difference feature, the semantic level difference feature is generated by capturing the correlation between features and filtering isolated noise points, and the difference between the double time phase remote sensing images is revealed. The feature fusion component fuses the double time phase position enhanced feature and the semantic level difference feature according to the time sequence and the proportion, enhances the image difference feature while retaining the continuity of the time sequence evolution, can reflect the feature change of different time phases, and can strengthen the performance of the time sequence difference in space and semantics, and finally outputs more discriminative space-time difference features. For example, through the time sequence encoder, newly added or disappeared buildings between T1 time phase and T2 time phase can be effectively captured, and the accuracy and time sequence consistency of change detection are ensured.

[0037] The parallel multi-task branch network includes a change detection branch and a change description branch. The space-time difference features output by the time sequence encoder are input into the change detection branch and the change description branch at the same time, the change detection branch is provided with time dimension change clues for the detection task, and the change description branch is provided with time sequence context for the description task, avoiding repeated modeling of time information, providing a unified feature basis for the two branches, and ensuring the time sequence logical consistency of the detection result and the description result.

[0038] The change detection branch outputs a change detection mask based on the dual-time remote sensing features and the spatio-temporal difference features. Specifically, the dual-time features and the temporal difference features are pixel by pixel aligned and fused by using a feature splicing or weighted fusion strategy, the channel dimension is compressed by 1x1 convolution, the unified fusion features are generated, the multi-layer perception is used to perform nonlinear transformation on the fusion features, the high-dimensional discriminative features are output, the 3x3 convolution layer + Softmax is used to form a classification head, the high-dimensional discriminative features are mapped into a single-channel probability map, the dynamic threshold is set, for example, 0.5, the probability Figure Two value is quantized, and a binary change mask is output, for example, 0 represents an unchanged pixel, and 1 represents a changed pixel, so as to mark the pixel position of the changed region.

[0039] The change description branch receives the change detection mask, cuts out the feature map block of the corresponding region from the spatio-temporal difference feature map according to the bounding box of the change detection mask, generates the weight of the feature map block based on the change detection mask, fuses the feature map block and the weight, obtains the mask-guided description feature, and generates the description text. Specifically, RoIPool cuts out the feature map block of the corresponding region from the spatio-temporal difference feature map according to the bounding box of the change detection mask, and obtains the description feature map. The attention weight is generated based on the change detection mask by using an attention mechanism: the change detection mask is interpolated to adjust to the target size, and then the softmax activation function is used to generate the dynamic attention weight. The description feature map and the attention weight are multiplied pixel by pixel to obtain the mask-guided description feature; the Transformer decoder generates the description text according to the mask-guided description feature.

[0040] In the generation process of the description text, the binary change mask output by the change branch serves as a benchmark reference for spatial semantics, provides accurate spatial coordinate anchoring, ensures that the text description strictly corresponds to the image change region, embeds the features of the change description branch through one-way fusion, establishes a hard association between the change region and the language description in the spatial dimension, makes the change description branch pay more attention to the change region marked by the binary change mask and ignore the unchanged background region during the generation of the description text, and obtains more accurate and vivid information. This one-way fusion mechanism ensures that the change description branch does not affect the change detection result in the reverse direction, maintains the objectivity of the detection, and improves the pertinence of the description.

[0041] The change detection branch and the change description branch process the detection and description tasks in parallel based on the shared spatio-temporal difference features, which not only guarantees the temporal logic consistency of the detection result and the description result, but also improves the overall efficiency. Specifically, in the analysis of building changes, the change detection branch can quickly locate the newly added building area and output a change detection mask to mark which pixels are the areas where the building has changed; the description branch can describe the changes in natural language. This parallel architecture not only reduces the training redundancy caused by repeated feature calculation, but also enhances the semantic accuracy of the description through the change detection mask, so that the description text will strictly focus on the change area marked by the mask, making the final result not only have pixel-level detection reliability, but also maintain the readability and intuitiveness of the language description. It effectively solves the problems of change detection and change description in complex scenarios such as dense building areas, industrial expansion, and suburban development, and realizes the connection from pixel changes to semantic understanding. For example, when the detection result of the change detection branch is that a certain area is building demolition, the change description branch generates corresponding expressions such as "the original warehouse area is now empty" corresponding to the visual changes.

[0042] Step three, training the image change detection model based on the dataset to obtain the trained image change detection model; the dataset is divided into a training set and a validation set; the training set is input into the image change detection model to train the image change detection model, and when the total loss function value of the image change detection model no longer decreases, the training is stopped to obtain the trained image change detection model; the validation set is used to verify the trained image change detection model, and if the verification is passed, the trained image change detection model is obtained; if the verification fails, the hyperparameters are adjusted and the training is restarted.

[0043] Total loss function of the image change detection model , wherein represents the loss function of the change description branch, represents the loss function of the change detection branch, represents the first balancing factor, represents the second balancing factor. The loss function of the change description branch and the loss function of the change detection branch are both cross-entropy loss functions.

[0044] It should be noted that the values of the first balancing factor λ1 and the second balancing factor λ2 need to be adjusted according to the task requirements and model performance. If the importance of the change description branch and the change detection branch is equivalent, λ1=λ2=0.5 can be set to ensure that the contributions of the two to the total loss are balanced. If more attention is paid to detection accuracy, i.e., building change positioning, λ2 can be increased, for example, λ1=0.3 and λ2=0.7. If more emphasis is placed on the accuracy of the description, λ1 can be increased, for example, λ1=0.7 and λ2=0.3.

[0045] Step four, obtaining the double-time-phase remote sensing images to be detected; the double-time-phase remote sensing images need to cover the same geographical area, have matched spatial resolution, and have been geometrically and radiometrically corrected. For example, images from the Gaofen series of satellites or unmanned aerial vehicle aerial images.

[0046] Step five, inputting the double-time-phase remote sensing images to be detected into the trained image change detection model to obtain the change detection result and the change description result.

[0047] In one possible embodiment, the processed double-time-phase remote sensing images to be detected are input into the trained image change detection model, first, the CNN network is used to extract the detail features of the double-time-phase remote sensing images respectively, then the EIG module is used to perform multi-scale edge enhancement on the detail features, highlighting the key details such as building boundaries and road profiles; subsequently, the DDaT module is used to adaptively establish the feature correspondence across time phases by using the dynamic offset attention mechanism, effectively aligning the feature displacement caused by the differences in view angle or season, obtaining the double-time-phase remote sensing features. The time sequence encoder obtains the spatio-temporal difference features based on the double-time-phase remote sensing features. The change detection branch generates the pixel-level binary change mask based on the spatio-temporal difference features, accurately marking the newly added or disappeared building areas; at the same time, the change description branch generates the description text by using the Transformer decoder to decode the description features guided by the mask.

[0048] The above is only an embodiment of the present application, and does not limit the present application in any way. Any simple modification, change and equivalent structural change made according to the technical essence of the present application to the above embodiment are still within the protection scope of the technical solution of the present application.

Claims

1. A method for image change detection based on detecting and describing parallel processing, characterized by: The method comprises the following steps: Step one, obtaining a data set, the data set being a dual-time-phase remote sensing image with known change detection results and change description results; Step two, constructing an image change detection model: the image change detection model comprising a feature extraction module, a time sequence encoder, and a parallel multi-task branch network; The feature extraction module extracts features from the dual-time-phase remote sensing image to obtain dual-time-phase remote sensing features; The time sequence encoder generates spatio-temporal difference features with spatial position and temporal change through position embedding and difference representation; The parallel multi-task branch network comprises a change detection branch and a change description branch; the change detection branch outputs a change detection mask based on the dual-time-phase remote sensing features and the spatio-temporal difference features; the change description branch receives the change detection mask, cuts out a feature map block corresponding to the region from the spatio-temporal difference feature map according to the bounding box of the change detection mask, generates a weight of the feature map block based on the change detection mask, fuses the feature map block and the weight to obtain mask-guided description features, and generates a description text; Step three, training the image change detection model based on the data set to obtain a trained image change detection model; Step four, obtaining a dual-time-phase remote sensing image to be detected; Step five, inputting the dual-time-phase remote sensing image to be detected into the trained image change detection model to obtain change detection results and change description results.

2. A method for image change detection based on detecting and describing parallel processing according to claim 1, characterized in that: The change description branch receives the change detection mask, cuts out a feature map block corresponding to the region from the spatio-temporal difference feature map according to the bounding box of the change detection mask, generates a weight of the feature map block based on the change detection mask, fuses the feature map block and the weight to obtain mask-guided description features, and generates a description text, specifically comprising: RoIPool cuts out a feature map block corresponding to the region from the spatio-temporal difference feature map according to the bounding box of the change detection mask to obtain a description feature map; an attention weight is generated based on the change detection mask through an attention mechanism; the description feature map and the attention weight are multiplied pixel by pixel to obtain mask-guided description features; and a Transformer decoder generates a description text according to the mask-guided description features.

3. A method for image change detection based on detecting and describing parallel processing according to claim 2, characterized in that: The attention weight is generated based on the change detection mask, specifically comprising: the change detection mask is interpolated to adjust to a target size, and then a softmax activation function is used to generate a dynamic attention weight.

4. A method for detecting changes in images based on detecting and describing parallel processing according to claim 1, characterized in that: The change detection branch outputs a change detection mask based on the dual-time-phase remote sensing features and the spatio-temporal difference features, specifically comprising: The dual-time-phase features and the time sequence difference features are fused to generate fusion features, the fusion features are subjected to nonlinear transformation to generate a probability map, and the change detection mask is generated through dynamic threshold binarization.

5. A method for detecting changes in images based on detecting and describing parallel processing according to claim 1, characterized in that: The feature extraction module extracts features from the dual-time-phase remote sensing image to obtain dual-time-phase remote sensing features, specifically comprising: A CNN network is used to extract detailed features of the dual-time-phase remote sensing image to obtain a CNN feature map; A multi-scale asymmetric convolution of an edge information guiding module and edge detection are cooperated with the CNN feature map to strengthen the geometric boundary of the change region to obtain an EIG enhanced feature map; The dynamic deformable Transformer adjusts the attention area through a dynamic offset mechanism adaptive to image content, obtains the global semantics of irregularly changing targets in the EIG enhanced feature map, and outputs dual-phase remote sensing features.

6. A method for detecting changes in images based on detecting and describing parallel processing according to claim 1, characterized in that: The timing encoder generates spatial and temporal difference features by position embedding and differential representation, and specifically includes: The position embedding component adds a learnable two-dimensional position encoding to the dual-phase remote sensing features to generate dual-phase position enhanced features. The differential representation component generates semantic-level differential features based on the dual-phase position enhanced features. The feature fusion component fuses the dual-phase position enhanced features and the semantic-level differential features in terms of timing and proportion to generate the spatial-temporal difference features.

7. A method for detecting changes in images based on detecting and describing parallel processing according to claim 6, characterized in that: The differential representation component generates semantic-level differential features based on the dual-phase position enhanced features, and specifically includes: calculating the basic difference of the dual-phase position enhanced features through a basic differential operation, quantifying the increase and decrease of the feature values, preliminarily marking the potential change area, obtaining the basic differential features, performing nonlinear transformation on the basic differential features through a learnable nonlinear convolution, suppressing irrelevant interference, enhancing the difference signal of the real change area, and obtaining the processed differential features; performing joint feature extraction on the processed differential features, filtering isolated noise points by capturing the correlation between features, and generating semantic-level differential features.

8. A method for detecting changes in images based on detecting and describing parallel processing according to claim 1, characterized in that: The image change detection model is trained based on the data set to obtain a trained image change detection model, and specifically includes: The data set is divided into a training set and a validation set; the training set is input into the image change detection model to train the image change detection model, and when the total loss function value of the image change detection model no longer decreases, the training is stopped to obtain a trained image change detection model; the trained image change detection model is verified using the validation set, and if the verification is passed, the trained image change detection model is obtained; if the verification is not passed, the data set is replaced and the training is performed again.

9. A method for image change detection based on detecting and describing parallel processing according to claim 8, characterized in that: Total loss function of the image change detection model wherein denotes the loss function of the change description branch, denotes the loss function of the change detection branch, denotes a first balancing factor, denotes a second balancing factor.

10. A method for detecting changes in images based on detecting and describing parallel processing according to claim 8, characterized in that: Loss function for the change description branch Loss function for the change detection branch are both cross-entropy loss functions.

Images