A high-quality change detection method and device based on iterative enhancement of difference features

Abstract

The application discloses a high-quality change detection method and device based on difference feature iterative enhancement, comprising the following steps: determining the influence of feature difference quality on change detection results by noise simulation interference; extracting multi-scale convolution features by using U-Net; constructing an IDET module to generate difference information from two input multi-scale convolution features, and iteratively refining to obtain high-quality feature difference maps of each scale; inputting the high-quality feature difference maps of each scale into a coarse-to-fine strategy constructed by multiple convolution modules to generate change maps of each scale, and then fusing the change maps of each scale to generate a final change map. The device comprises a processor and a memory. The application optimizes feature differences, suppresses invariant regions, and obtains high-quality change maps.

Description

Technical Field

[0001] This invention relates to the field of change detection, and in particular to a high-quality change detection method and apparatus based on iterative difference-enhanced transformers (IDET). Background Technology

[0002] Change detection (CD) aims to detect areas of change caused by changes in objects (such as newly appearing objects in a scene) from a set of images taken at different times. It has wide applications in urban development, disaster prediction, and resource monitoring. However, change detection remains very challenging because two images may be taken from different angles and under different lighting conditions, and the background scene may undergo unknown changes.

[0003] In reference [1], a simple idea is to compute the difference between features of two images and map the feature difference to the final change map, using features and feature differences to overcome environmental changes and positional misalignment. A representative work in reference [2] proposes cross-encoded features to compute the feature difference between image pairs and generate a change map. The fluid pyramid integration (FPIN) network in reference [3] uses a complex feature extraction network to improve CD accuracy by enhancing feature representation capabilities. Another work in reference [4] proposes position-related attention, channel-related attention, and change difference modules to capture the correlation between two images. The aforementioned state-of-the-art (SOTA) methods focus on designing advanced network architectures to map feature differences to the final change map. While these methods achieve impressive detection accuracy, they neglect the quality impact of feature differences. Intuitively, it would be easier to achieve high-quality change maps if feature differences could clearly highlight the major changes while suppressing unchanged regions.

[0004] Furthermore, since its introduction in reference [5], this variation network has been widely used in most natural language processing (NLP) tasks due to its long-term dependency. Recently, it has shown good performance in computer vision tasks, such as image classification in reference [6], object detection in reference [7], and semantic segmentation in reference [8]. The excellent performance of the variation network model has inspired research on its application in CD. To date, there is still very little work on applying variation methods to CD tasks.

[0005] The bitemporal image transformer (BiT) proposed in reference [9] can model the semantic information of the changing object using a change network encoder to refine the coarse change mapping predicted by the decoder. However, when the changing object is small and the changes are diverse, learning semantic labels is difficult. The ChangeFormer network in reference

[10] consists of a series of hierarchical change network encoders used to extract long-range features to generate change masks. Due to the low resolution of its encoder, it cannot capture detailed changes in the boundaries.

[0006] References

[0007] [1]Huang, R.; Zhou, M.;

[0008] [2] Huang, R.; Zhou, M.; Zhao, Q.; and Zou, Y. 2020. Change detection with absolute difference of multiscale deep features. Neurocomputing, 418: 102–113.

[0009] [3] Zhao, J.-X.; Cao, Y.; Fan, D.-P.; Cheng, M.-M.; Li,

[0010] [4]Zhang,L.;Hu,X.;Zhang,M.;Shu,Z.;and Zhou,H.2021.Object-level changedetection with a dual correlation attention-guided detector.ISPRS Journal ofPhotogrammetry and Remote Sensing,177:147–160.

[0011] [5]Vaswani,A.;Shazeer,N.;Parmar,N.;Uszkoreit,J.;Jones,L.;Gomez,A.N.;Kaiser, and Polosukhin,I.2017.Attention is all you need.In Advances inneural information processing systems,5998–6008.

[0012] [6]Dosovitskiy,A.;Beyer,L.;Kolesnikov,A.;Weissenborn,D.;Zhai,X.;Unterthiner,T.;Dehghani,M.;Minderer,M.;Heigold,G.;Gelly,S.;et al.2020.Animage is worth 16x16 words:Transformers for image recognition at scale.arXivpreprint arXiv:2010.11929.

[0013] [7]Carion,N.;Massa,F.;Synnaeve,G.;Usunier,N.;Kirillov,A.;andZagoruyko,S.2020.End-to-end object detection with transformers.In EuropeanConference on Computer Vision,213–229.Springer.

[0014] [8]Zheng, S.; Lu, J.; Zhao, H.; Zhu, Recognition,6881–6890.

[0015] [9]Hao Chen, ZQ; and Shi, Z.2021.Remote Sensing Image Change Detectionwith Transformers.IEEE Transactions on Geoscience and Remote Sensing,1–14

[0016]

[10] Bandara,WGC; and Patel,VM2022.A transformerbased siamesenetwork for change detection.arXiv preprint arXiv:2201.01293 Summary of the Invention

[0017] This invention provides a high-quality change detection method and apparatus based on iterative enhancement of differential features. First, multi-scale features of two input images are extracted using a U-Net containing an encoder and decoder. Second, the reference change network and query change network in IDET are used to extract long-range information at each scale from the two feature maps, respectively, to calculate the feature difference map. Then, the difference enhancement network in IDET iteratively enhances the difference information, obtaining refined feature differences at each scale. Finally, a coarse-to-fine fusion method is designed to combine the multi-scale results, optimizing the feature differences, suppressing invariant regions, and obtaining a high-quality change map. See the description below for details.

[0018] A high-quality change detection method based on differential feature iteration enhancement, the method comprising:

[0019] The impact of characteristic difference quality on change detection results was determined by simulating noise interference.

[0020] Extracting multi-scale convolutional features using U-Net;

[0021] The IDET module is constructed to generate difference information from the multi-scale convolutional features of two inputs, and iteratively refines it to obtain high-quality feature difference maps at each scale.

[0022] High-quality feature difference maps at each scale are input into a coarse-to-fine strategy constructed from multiple convolutional modules to generate scale change maps, and then the scale change maps are merged to generate the final change map.

[0023] The IDET module is constructed as follows:

[0024] τ * (R * ) = MLP(Norm(Z)) + Z,

[0025] Z = MSA(Norm(R) * ))+R *

[0026] Where, τ * It refers to τ ref (·), τ que (·) Network, R * R refers to multi-scale features x R y The input multi-scale features are first processed by a residual unit composed of Norm and MSA, and the result is denoted as Z. Then, the residual unit composed of Norm and MLP is input to obtain R. x R y In the long-range information, MLP stands for Multilayer Perceptron, Norm is the normalization layer, and MSA is the multi-head self-attention mechanism;

[0027] Dissimilarity information of two long-range features is generated using MLP and Norm. The procedure function is as follows:

[0028]

[0029] Where, τ ref (R x R is the multi-scale feature of input x. x Long-range information obtained through the reference change network, τ que (R y R is the multi-scale feature of input y. y Long-distance information obtained by querying the change network;

[0030] In τ diff (·) contains:

[0031]

[0032]

[0033] The entire IDET process can be described as follows:

[0034]

[0035] in, To obtain the refined feature differences through iterative updates.

[0036] Specifically, the step of inputting high-quality feature difference maps at each scale into a coarse-to-fine strategy constructed by multiple convolutional modules to generate scale variation maps is as follows:

[0037]

[0038] Where l represents the scale number. It refers to the feature differences obtained through iterative updates at the l-scale. Conv1 represents the result of fusion at scale l. * () refers to the convolutional module corresponding to each scale, and "[·]" indicates a connection operation, which inputs the initial feature difference D into the convolutional layer to extract the change map M at scale l. l :

[0039]

[0040] Among them, Conv2 l () is the second convolutional module at scale l. Conv3() is used to generate the initial change map M from the initial feature differences D. 0 The third convolutional module.

[0041] Furthermore, the final transformation diagram is as follows:

[0042] M = Conv4([M 0 ,...,M L ])

[0043] Conv4() is a fourth convolutional module consisting of 12×3×3×2 convolutional filters; the input to this module is M. * The output shows the changes at various scales.

[0044] A high-quality change detection device based on differential feature iteration enhancement, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to execute method steps.

[0045] A computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform method steps.

[0046] The beneficial effects of the technical solution provided by this invention are:

[0047] 1. This invention constructs IDET for high-quality change detection, using mainstream image datasets for change detection as training samples to improve the quality of feature differences, and obtains a change detector that is unaffected by changes in lighting, angle, and other conditions.

[0048] 2. This invention investigates the impact of feature difference quality on detection accuracy by adding random noise of varying severity to the traditional CD method, proving that feature difference quality is key to high-quality detection.

[0049] 3. This invention uses a U-Net module, which includes an encoder and a decoder, to extract features from the input image and output multi-scale representations of each layer from the decoder.

[0050] 4. This invention combines an effective multi-head self-attention (MSA) module with a multilayer perceptron (MLP) module and a normalization layer (Norm) to design a reference change network, a query change network, and a difference enhancement network in IDET, thereby achieving the function of iterative enhancement of difference information and improving the quality of feature differences.

[0051] 5. This invention designs a coarse-to-fine fusion method, which integrates the initial feature differences with the enhanced feature differences at each scale obtained after iteration to obtain the final transformation map. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the multi-scale change detection model with embedded IDET structure proposed in this invention;

[0053] Figure 2 This is a schematic diagram of the structure of the IDET proposed in this invention;

[0054] Figure 3 This is a schematic diagram of an existing feature-difference-based change detection system.

[0055] Figure 4 This is a schematic diagram illustrating the evaluation results based on the VL-CMU-CD dataset with noise of varying severity added to feature difference D.

[0056] Figure 5 A visual comparison of eight CD methods on the VL-CMU-CD, PCD, CDnet, LEVIR-CD, CDD, and AICD datasets;

[0057] Figure 6 Display the original feature difference, noise feature difference D, and enhancement feature difference respectively. D and A schematic diagram showing the detection result M and the ground truth.

[0058] Figure 7 This is a comparison chart of ablation experimental results with and without the IDET structure. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0060] This method addresses the problem of low feature difference quality affecting change detection results due to issues such as lighting and angle. It proposes an iterative difference enhancement change method and device for high-quality change detection, which can continuously iterate and update during the calculation of feature differences, thereby improving the quality of feature differences and making the change detection results more accurate.

[0061] Example 1

[0062] A high-quality change detection method based on differential feature iterative enhancement, see [link to relevant documentation]. Figure 1 , Figure 2 The method includes the following steps:

[0063] I. By simulating noise interference, the impact of characteristic difference quality on the change detection results is determined.

[0064] According to the embodiments of the present invention Figure 3 The importance of the quality of feature differences to the final transformation map was studied. Given images X and Y, they are input into a Convolutional Neural Network (CNN) to obtain a series of features. Then, the feature differences of each layer can be calculated and fused using convolutional layers, as shown in formula (1):

[0065]

[0066] in, These are the features of X and Y obtained by CNN at scale l, respectively, and correspondingly, D l This represents the feature difference between two images at scale l. After... Figure 3The Fuse structure shown fuses the feature differences from l=1 to L to obtain the total feature difference D. This feature difference D can then be input into two other convolutional layers to predict the final change map M, as shown. Figure 3 As shown. Previous work mainly focused on designing more advanced architectures to replace basic CNNs. In contrast, this invention focuses on the impact of total feature difference D quality.

[0067] Let D be the difference in original features obtained from the input image after the above process. o Differences in original features D o Random noise is added to the invariant region to simulate feature differences of varying quality. Specifically, based on the ground truth of the changed region, the average value of the feature difference between the inside and outside of the changed region is first calculated, denoted as […]. and Then, a zero-mean random noise with a standard deviation of is set. Add this noise to the original feature difference D o In the region where the mean remains unchanged, a low-quality feature difference is obtained, denoted as D1. This method obtains a low-quality feature difference each time by adjusting the zero-mean noise parameter α from 5 to 100 in intervals of 5, and evaluates it using a comprehensive evaluation index (F1-measure).

[0068] Specifically, this noise is added to the original feature difference D. o The process of obtaining a lower quality feature difference from an unchanging region is well known to those skilled in the art, and will not be described in detail in the embodiments of the present invention.

[0069] Depend on Figure 4 As can be seen from the star-shaped curve, the basic feature detection method is highly sensitive to the quality of feature differences. Clearly, as noise increases (α value continuously increases), the F1 value decreases significantly. Therefore, this embodiment of the invention demonstrates that the accuracy of the change detection method can be improved by altering the quality of feature differences.

[0070] 2. Extracting multi-scale convolutional features using U-Net.

[0071] Specifically, this step involves inputting reference images and query images of the same scene with unchanged external environments (lighting, angles, etc.) into a U-Net containing an encoder and a decoder. The U-Net has 5 compression modules that extract abstract features by reducing spatial resolution; and 5 extension modules that restore the resolution of features by fusing high-resolution and low-resolution features from the compression modules, thereby extracting multi-scale convolutional features as representations to obtain multi-scale representations of each layer of the decoder.

[0072] The input images in this embodiment of the invention are taken from popular change detection datasets, namely VL-CMU-CD, PCD, and CDnet, as well as three remote sensing change detection datasets, namely LEVIR-CD, CDD, and AICD (all of the above datasets are well-known in the art). For images in VL-CMU-CD, 80% of the image pairs were randomly selected for training, and 20% were used for testing. The training images were enhanced by random cropping and flipping. For PCD, images of size 224×224 were cropped, and the training and testing sets were divided in an 8:2 ratio. The training images were flipped to increase the number of training images. For CDnet, the proposed training and testing datasets were used in this experiment. For LEVIR-CD, the images were cropped into small, non-overlapping 256×256 blocks, and the training images were enhanced by rotation and color transformation. Finally, 15835 and 4675 image pairs were obtained for training and testing, respectively. For CDD, the original images were cropped and rotated to generate 10000 training images and 3000 test image pairs. For AICD, image pairs are randomly cropped, and training images are enhanced through rotation and color vibration. This embodiment of the invention uses the above values ​​as examples for illustration; however, in specific implementations, this embodiment of the invention does not impose limitations.

[0073] The U-Net structure in this embodiment of the invention can be found in [reference needed]. Figure 1 The first part of the multi-scale IDET structure is shown in the image input section. The Encoder and Decoder parts of U-Net share parameters and achieve feature fusion through concatenation. The overall structure can improve the accuracy of feature extraction while using fewer training images, thus obtaining multi-scale representations of each layer of the decoder.

[0074] Third, construct the IDET module to generate difference information from two input multi-scale features and perform iterative refinement.

[0075] The structure of the IDET module in this embodiment of the invention is as follows: Figure 2 As shown. The module's input mainly consists of three parts, including R. x R y All of these are multi-scale features output by U-Net in step two, and D represents the initial feature difference.

[0076] IDET consists of three modules: the reference change network τ ref (·), Query the change network τ que (·) and differential enhancement network τ diff (·), the network structures are as follows Figure 2 As shown. Where τ ref (·), τ que (·) is used to extract the above R respectively. x Ry The long-distance information in the data is used to calculate the feature difference map, the specific form of which is shown in formula (2):

[0077]

[0078] Where, τ * It refers to τ ref (·), τ que (·) Network, R * R refers to multi-scale features x R y The input multi-scale features are first processed through a residual unit composed of Norm and MSA, and the result is denoted as Z. Then, the residual unit is input again, consisting of Norm and MLP, to obtain R. x R y In the long-range information, MLP stands for Multilayer Perceptron, Norm is the normalization layer, and MSA is the multi-head self-attention mechanism.

[0079] Furthermore, difference information between two long-range features is generated using MLP and Norm. The procedure function is as follows:

[0080]

[0081] Where, τ ref (R x R is the multi-scale feature of input x. x Long-range information obtained through the reference change network, τ que (R y R is the multi-scale feature of input y. y Long-distance information obtained by querying the changing network.

[0082] The key to the embodiments of the present invention is τ diff (·) Use the difference information obtained from equation (3) To guide the refinement process of the initial feature difference D, in τ diff (·) contains:

[0083]

[0084] The idea is to The data is embedded into MSA and MLP to extract long-range information, and the embedded residual structure is used to update the initial feature difference value D.

[0085] In summary, the entire IDET process can be described as follows:

[0086]

[0087] in, To obtain the refined feature differences through iterative updates, The calculation process is shown in formula (3).

[0088] 4. Use the convolution module to construct a coarse-to-fine strategy to generate change maps at various scales.

[0089] This invention's embodiments include the design of multiple convolutional modules (see [link]). Figure 1 As shown. The refined feature differences obtained in the IDET module are then input into the first convolutional module Conv1, which contains three parts. * () in (Conv1) * () represents the convolutional modules corresponding to each scale, such as Conv11(), Conv12(), ..., Conv1 l (), where l is the scale number). Conv1 * Each part in parentheses contains a convolutional layer and a ReLU layer. The first part reduces the number of channels to half the number of channels in the input features; the second part has the same number of input and output channels to further refine the features; the third part produces features with a given number of output channels. Through the combined action of these three parts, results at all scales are fused using a coarse-to-fine strategy, i.e.:

[0090]

[0091] Where l represents the scale number. It is the feature difference obtained by iterative update at the l-scale, and its calculation process is shown in formula (5); Conv1 represents the result of fusion at scale l. * () represents the convolutional module corresponding to each scale, and "[·]" represents the connection operation. Empty.

[0092] get Then, the difference between the initial feature D and the feature D is input into a convolutional layer to extract the change map M at the l-scale. l :

[0093]

[0094] Conv2 here l () is the second convolutional module at scale l, with a size of N × (N represents the number of input channels); The calculation process is shown in formula (6); Conv3() is used to generate the initial change map M from the initial feature difference D. 0 The third convolutional module.

[0095] 5. Integrate the change maps at various scales to generate the final change map.

[0096] The process of generating the final change diagram M in this design example is as follows:

[0097] M = Conv4([M 0 ,...,M L (8)

[0098] Conv4() is a fourth convolutional module consisting of 12×3×3×2 convolutional filters; the input to this module is M. * The calculation process for the transformation maps output at each scale is shown in formula (7). By inputting each transformation map into the convolution module, the final transformation map M can be obtained by fusion.

[0099] VI. Network Training and Testing

[0100] This design model was implemented in PyTorch and trained using a single NVIDIA GeForce 2080Ti GPU. The base learning rate was set to 1e-3, and the batch size was set to 4. The Adam algorithm was used to update the parameters, with a momentum of 0.9 and a weight decay of 0.999. During training, the same training parameters were used on all six datasets, and all change detection methods were trained on the general scene dataset and the remote sensing dataset for 20 and 200 epochs respectively.

[0101] This design compares the IDET-based method with seven other change detection methods: FCNCD, ADCDnet, CSCDnet, BIT, IFN, STAnet, and ChangeFormer. All comparison methods are trained under the same settings, and different CD methods are evaluated using precision (P), recall (R), F1-measure (F1), overall accuracy (OA), and intersection over union (IoU).

[0102] By adopting this design, one can... Figure 5 This document provides a visual comparison of eight CD methods across different datasets, and also includes... Figure 6 The graph showing the changes in comparable feature differences before and after enhancement (columns 6 and 7) is provided by... Figure 7 The detection results with and without IDET can be compared (columns 4 and 6). Clearly, IDET enhances feature differentiation and has a significant advantage in the final evaluation.

[0103] In summary, this design example significantly improves the accuracy of change detection by introducing U-Net to extract multi-scale features from images, using IDET to iteratively enhance feature differences, and then fusing features through multiple convolutional modules. This also demonstrates the importance of feature difference quality.

[0104] Example 2

[0105] The feasibility of the scheme in Example 1 is verified below with reference to the attached diagram, as detailed in the following description:

[0106] Depend on Figure 4 , Figure 6 The presented content leads to the conclusion that the quality of feature differences can affect the accuracy of change detection, and based on this conclusion, a [system / mechanism] can be designed. Figure 1 , Figure 2 The network structure shown is used to build and train the IDET-based change detection model in this embodiment of the invention; and Figure 3 Compared to previous change detection methods, the embodiments of the present invention improve change detection accuracy, such as... Figure 5 As shown, a visual comparison of eight CD methods on the VL-CMU-CD, PCD, CDnet, LEVIR-CD, CDD, and AICD datasets is presented. Figure 7 The ablation experiment results shown, with and without the IDET module, lead to the conclusion that the design method of this invention is superior to all basic methods, and further illustrate the importance of the quality of feature differences.

[0107] Figure 5 This document contains representative CD examples of state-of-the-art methods on different datasets. Specifically, FCNCD can capture the main parts of the changes but loses details of changes in rows 1-3. ADCDnet detects changes in image pairs well on the commonsense dataset but fails to capture overall changes in rows 4 and 6 of the remote sensing dataset. CSCDnet fails to detect changes in buildings in rows 5 and 6. IFN is susceptible to lighting differences, resulting in poor detection quality in rows 1 and 6. ChangeFormer fails to detect fine-grained changes in rows 1 and 2. BIT is susceptible to shadow and lighting differences, leading to false detections in rows 1 and 6. STAnet performs poorly when lighting differences are large (rows 2, 5, and 6). In contrast, the results of this embodiment (last column) are not only complete but also accurate at the boundaries of changes.

[0108] Figure 4 , Figure 6 The importance of feature difference enhancement is visualized in the image. Figure 6The graph showing the changes in comparable feature differences before and after enhancement (columns 6 and 7) is provided by... Figure 7 The detection results with and without IDET can be compared (columns 4 and 6). Clearly, IDET enhances the expected changes in feature differences, giving it a significant advantage in the final evaluation.

[0109] A high-quality change detection device based on differential feature iteration enhancement includes a processor and a memory. The memory stores program instructions, and the processor calls the program instructions stored in the memory to cause the device to perform the following method steps:

[0110] The impact of characteristic difference quality on change detection results was determined by simulating noise interference.

[0111] Extracting multi-scale convolutional features using U-Net;

[0112] The IDET module is constructed to generate difference information from the multi-scale convolutional features of two inputs, and iteratively refines it to obtain high-quality feature difference maps at each scale.

[0113] High-quality feature difference maps at each scale are input into a coarse-to-fine strategy constructed from multiple convolutional modules to generate scale change maps, and then the scale change maps are merged to generate the final change map.

[0114] The IDET module is constructed as follows:

[0115] τ * (R * ) = MLP(Norm(Z)) + Z,

[0116] Z = MSA(Norm(R) * ))+R *

[0117] Where, τ * It refers to τ ref (·), τ que (·) Network, R * R refers to multi-scale features x R y The input multi-scale features are first processed by a residual unit composed of Norm and MSA, and the result is denoted as Z. Then, the residual unit composed of Norm and MLP is input to obtain R. x R y In the long-range information, MLP stands for Multilayer Perceptron, Norm is the normalization layer, and MSA is the multi-head self-attention mechanism;

[0118] Dissimilarity information of two long-range features is generated using MLP and Norm. The procedure function is as follows:

[0119]

[0120] Where, τ ref (R x R is the multi-scale feature of input x. x Long-range information obtained through the reference change network, τ que (R y R is the multi-scale feature of input y. y Long-distance information obtained by querying the change network;

[0121] In τ diff (·) contains:

[0122]

[0123]

[0124] The entire IDET process can be described as follows:

[0125]

[0126] in, To obtain the refined feature differences through iterative updates.

[0127] Specifically, the process of inputting high-quality feature difference maps at each scale into a coarse-to-fine strategy constructed from multiple convolutional modules to generate scale variation maps is as follows:

[0128]

[0129] Where l represents the scale number. It refers to the feature differences obtained through iterative updates at the l-scale. Conv1 represents the result of fusion at scale l. * () refers to the convolutional module corresponding to each scale, and "[·]" indicates a connection operation, which inputs the initial feature difference D into the convolutional layer to extract the change map M at scale l. l :

[0130]

[0131] Among them, Conv2 l () is the second convolutional module at scale l. Conv3() is used to generate the initial change map M from the initial feature differences D. 0 The third convolutional module.

[0132] Furthermore, the final change diagram is as follows:

[0133] M = Conv4([M 0 ,...,M L ])

[0134] Conv4() is a fourth convolutional module consisting of 12×3×3×2 convolutional filters; the input to this module is M. * The output shows the changes at various scales.

[0135] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0136] The execution entities of the processor 1 and memory 2 mentioned above can be devices with computing functions such as computers, microcontrollers, and microcontrollers. In specific implementation, the embodiments of the present invention do not limit the execution entity and can select it according to the needs of actual application.

[0137] Data signals are transmitted between memory 2 and processor 1 via bus 3, which will not be described in detail in this embodiment of the invention.

[0138] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0139] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0140] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0141] 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. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0142] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A 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 medium can be magnetic or semiconductor, etc.

[0143] Unless otherwise specified, the model numbers of the various devices in this invention are not limited; any device capable of performing the above functions is acceptable.

[0144] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the above-mentioned example numbers are for descriptive purposes only and do not represent the superiority or inferiority of the examples.

[0145] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-quality change detection method based on difference feature iterative enhancement, characterized in that, The method comprises: Determine the influence of feature difference quality on change detection results by noise simulation interference; Extract multi-scale convolution features by using U-Net; Construct an IDET module to generate difference information from two input multi-scale convolution features and perform iterative refinement to obtain high-quality feature difference maps of each scale; Input the high-quality feature difference maps of each scale into a coarse-to-fine strategy constructed by multiple convolution modules to generate change maps of each scale, and then fuse the change maps of each scale to generate a final change map; The construction of the IDET module is: ； wherein, denotes , network, denotes multi-scale feature R x , R y , the input multi-scale feature is first input into a residual unit composed of Norm and MSA, and the result is denoted as Z, and then input into a residual unit composed of Norm and MLP to obtain R x , R y , the long-distance information in R x , MLP is a multi-layer perceptron, Norm is a normalization layer, and MSA is a multi-head self-attention mechanism. Difference information of two long-distance features is generated by MLP and Norm The process function is as follows: ； wherein, a multi-scale feature R for input x x long-distance information obtained by the reference change network, a multi-scale feature R for input y y long-distance information obtained by the query change network; In which is: ； The whole process of IDET is described as: ； wherein to obtain an iteratively updated refined feature difference.

2. The high-quality change detection method based on difference feature iterative enhancement according to claim 1, characterized in that the input of the high-quality feature difference maps of each scale into a coarse-to-fine strategy constructed by multiple convolution modules to generate change maps of each scale is specifically:

3. The high-quality change detection method based on difference feature iterative enhancement according to claim 1, characterized in that the final change map is: ； wherein l denotes the scale number, is the feature difference of the iterative update refinement obtained at the scale l, denotes the result of fusion at the scale l, denotes the convolution module corresponding to each scale, and "[·]" denotes the connection operation, which is input into the convolution layer with the initial feature difference D to extract the change map at the scale l : ； wherein, is a second convolutional module in the l-scale, for generating an initial variation map from the initial feature difference D is a third convolutional module. The device comprises a processor and a memory, the memory stores program instructions, and the processor invokes the program instructions stored in the memory to make the device execute the method steps of any one of claims 1-3. The computer readable storage medium stores a computer program, the computer program comprises program instructions, and the program instructions are executed by the processor to make the processor execute the method steps of any one of claims 1-3. ； wherein is composed of fourth convolution module; the module input M * is the change map output for each scale.

4. An apparatus for high quality change detection based on iterative enhancement of difference features, characterized in that, ​ 5. A computer readable storage medium, characterized in that, ​

Images