Low-altitude unmanned aerial vehicle image change detection method and system

Abstract

Provided in the present invention are a low-altitude unmanned aerial vehicle image change detection method and system. Sate features are extracted for each pair of key features; due to the varying importance of state parts within overall image features in different scenes, the importance of the state parts is analyzed; on the basis of the varying importance, the state features are removed from the overall features, so as to obtain corresponding essential parts. The present invention comprehensively considers the importance of input images and the state parts so as to obtain the essential parts containing complete and pure information that reflects the areas of actual change, uses essential pairs as inputs to a decoding module, and calculates a difference between the essential pairs, such that the decoding module can more focus on predicting change charts of useful information, thus improving the accuracy of change detection of low-altitude unmanned aerial vehicles.

Description

A low-altitude UAV image change detection method and system

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This invention claims priority to Chinese patent application No. 202310983540.5 filed with the State Intellectual Property Office of China on August 4, 2023, entitled “A Method and System for Detecting Image Changes of Low-Altitude UAVs,” the entire contents of which are incorporated by reference into this invention and constitute a part of this invention for all purposes. Technical Field

[0003] The present invention belongs to the field of computer vision technology, and in particular relates to a method and system for detecting changes in low-altitude unmanned aerial vehicle (UAV) images. Background Art

[0004] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0005] Change detection is a task that aims to locate and segment changed regions in pairs of images taken at different times. Change detection is of great significance to many fields, such as visual surveillance, disaster assessment, and desertification control. Due to the wide application and huge potential of this task, more and more methods have been proposed to solve the change detection task. Some early representative works used traditional methods to solve this task and gradually improved the performance. In the past decade, due to the high performance of modern chips, the application of deep learning models has become feasible, which has inspired researchers to design more large-scale models in the field of computer vision. Since 2015, deep learning-based change detection methods have developed rapidly. Some existing methods have introduced more innovative models and have successively refreshed the state-of-the-art performance of change detection or its subtasks.

[0006] The data for low-altitude drone image change detection is street view images captured by low-altitude imaging devices (such as drones). Compared to remote sensing change detection tasks, low-altitude drone change detection involves a greater diversity of objects and scenes, which translates to more diverse and complex changes, making it a more challenging task. In addition to meteorological conditions, low-altitude drone change detection also faces challenges such as building occlusion, shadows of varying shapes and depths, and complex and changing artificial lighting. To achieve high-quality predictions, the model must be able to discern essential changes and minimize these adverse effects.

[0007] Because each input set in low-altitude drone change detection consists of two images, previous work naturally applied twin convolutional neural networks to extract features. In a narrow sense, a twin convolutional neural network consists of two convolutional neural networks with shared weights. When two input images pass through each convolutional neural network, their features are extracted in the same form. Ideally, identical regions remain similar in the feature map, while different regions are represented as distinct features. This makes it easier to find the regions of change between image pairs, meeting the objective requirements of the task.

[0008] However, all previous work using twin convolutional neural networks has focused on how to correctly fuse feature pairs or expand the feature differences in the change regions to make them easier to locate. Although these ideas help predict higher quality change maps, and in fact they do improve performance, another key point has been overlooked. In the low-altitude UAV change detection task, finding the change regions does not mean caring about all the differences between the two images. As defined in existing datasets, the change regions that need to be found include two-dimensional changes in the surface of the object (such as murals on the wall) and three-dimensional changes (such as the appearance, disappearance or movement of objects). Changes caused by different lighting, shadows or color styles are regarded as different states of the same essence, that is, "state changes." Even if the essence is the same, different states will also lead to very large representation differences, which will interfere with the model's prediction of the correct change map. Therefore, how to get rid of the influence of state changes, focus only on the changes in the essence, and improve the accuracy of low-altitude UAV change detection is a problem that technicians in this field urgently need to solve.

[0009] Summary of the Invention

[0010] To overcome the shortcomings of the above-mentioned prior art, the present invention provides a low-altitude UAV image change detection method and system, which decouples the backbone features into the essential part and the state part, so that the decoding module can focus more on predicting the real change map and improve the accuracy of low-altitude UAV change detection.

[0011] To achieve the above objectives, the first aspect of the present invention provides a method for detecting changes in low-altitude UAV images, comprising:

[0012] Obtaining a time-phase image pair of the area to be detected, and inputting the obtained time-phase image pair into a trained change detection module to obtain an image change detection result of the area to be detected;

[0013] The change detection module includes a feature extraction module, an interactive decoupling module and a decoding module; the feature extraction module is used to extract backbone features from the phase image pairs respectively to obtain multi-level backbone feature pairs;

[0014] The interactive decoupling module is used to extract state features from the backbone feature pairs to obtain state feature pairs, and obtain state feature weights based on the importance of the state features in the time phase image; and obtain essential feature pairs based on the state feature pairs, backbone feature pairs, and state feature weights;

[0015] The decoding module is used to cascade and fuse different essential feature pairs to obtain the change detection result.

[0016] A second aspect of the present invention provides a low-altitude UAV image change detection system, comprising:

[0017] Acquisition module: acquires the time phase image pair of the area to be detected;

[0018] Change detection module: input the acquired time phase image pair into the trained change detection module to obtain the image change detection result of the area to be detected;

[0019] The change detection module includes a feature extraction module, an interactive decoupling module and a decoding module; the feature extraction module is used to extract backbone features from the phase image pairs respectively to obtain multi-level backbone feature pairs;

[0020] The interactive decoupling module is used to extract state features from the backbone feature pairs to obtain state feature pairs, and obtain state feature weights based on the importance of the state features in the time phase diagram; and obtain essential feature pairs based on the state feature pairs, the backbone feature pairs, and the state feature weights corresponding to the state feature pairs.

[0021] The decoding module is used to cascade and fuse different essential feature pairs to obtain a change detection result of the area to be detected.

[0022] The third aspect of the present invention provides a computer device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, a low-altitude drone image change detection method is performed.

[0023] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which executes a method for detecting changes in low-altitude drone images when the computer program is executed by a processor.

[0024] One or more of the above technical solutions have the following beneficial effects:

[0025] In the present invention, the state features are first extracted for each pair of backbone features. Since the importance of the state part to the entire image feature varies in different scenarios, the importance of each state part is analyzed, and the state feature is removed from the complete feature based on the difference in importance, thereby obtaining the corresponding essential part. Taking the importance of the input image and the state part into comprehensive consideration, the obtained essential part will contain complete and pure information and can reflect the area of ​​actual change. The essential pairs reflecting the area of ​​actual change are used as the input of the decoding module, and the difference between each pair of essential pairs is calculated, so that the decoding module can focus more on predicting the real change map, thereby improving the accuracy of low-altitude UAV image change detection.

[0026] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0028] FIG1 is a schematic diagram of a change detection network structure in a first embodiment of the present invention;

[0029] FIG2 is a partial display diagram of a virtual data set in Example 1 of the present invention;

[0030] FIG3 is a visualization comparison diagram in the first embodiment of the present invention. DETAILED DESCRIPTION

[0031] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0032] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0033] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0034] Explanation of terms:

[0035] Image essential changes: Image change areas include two-dimensional changes in the surface of an object, such as a mural on a wall, and three-dimensional changes, such as the appearance, disappearance, or movement of an object.

[0036] Image state changes: Changes caused by different lighting, shadows, or color styles are treated as changes to different states of the same essence.

[0037] Example 1

[0038] This embodiment discloses a method for detecting changes in low-altitude drone images, including:

[0039] Acquire a pair of time-phase images of the area to be detected;

[0040] Input the acquired time-phase image pairs into the trained change detection module to obtain the change detection results of the area to be detected;

[0041] The change detection module includes a feature extraction module, an interactive decoupling module and a decoding module; the feature extraction module is used to extract backbone features from the phase image pairs respectively to obtain multi-level backbone feature pairs;

[0042] The interactive decoupling module is used to extract state features from the backbone feature pairs to obtain state feature pairs. Based on the importance of the state features in the time phase image, the state feature weights are obtained. Based on the state feature pairs, the backbone feature pairs, and the state feature weights corresponding to the state feature pairs, the essential feature pairs are obtained.

[0043] The decoding module is used to cascade and fuse different essential feature pairs to obtain the change detection results of the area to be detected.

[0044] As shown in FIG1 , the input of the essence and state decoupling network proposed in this embodiment is a pair of images (x1, x2) taken at different times (t1, t2), and the photographed objects are scenes at the same location.

[0045] In this embodiment, first, a twin convolutional neural network consisting of two VGG16 networks is used as a feature extractor of the network to extract multi-level feature pairs. Among them, both VGG networks are pre-trained on ImageNet and share weights between them.

[0046] Then, the backbone feature pairs except the first layer will be sent to the interaction decoupling module, which will decouple them into "state" pairs. and “essence” In this embodiment, the features of the first layer are not fed into the interaction decoupling module. This is mainly because the semantic information contained in this layer is too little, and more essential information cannot be extracted from it.

[0047] In each interaction decoupling module, the features of the two input images interact to generate the state part first, then generate a set of weights based on the importance of the state part in the current scene, and then calculate the essential part through a simple operation process. In this way, the backbone features are decoupled into the state part and the essential part. Finally, all the essential parts And the backbone features of the first layer is sent to the decoder to generate the final predicted change map.

[0048] In this embodiment, the interaction decoupling module decouples the complete feature pairs at each layer into essential pairs and state pairs. The state portion is generated using global information, and the essential portion is further divided based on the importance of the generated state portion. The input of each interaction decoupling module is a feature pair from one of the backbone layers.

[0049] The interaction decoupling module has two sub-branches, each of which is responsible for one of the features in the pairwise feature set.

[0050] Given a pair of image features Where N is the number of feature channels, H and W are the spatial resolutions of the features. The interactive decoupling module first combines the two sub-branches and Connect them in a cascade manner, and then reduce the number of channels to N again through a 1×1 convolution layer to generate the state part Since the state component should focus on invariant regions, and information from both phases of the two images is important, this embodiment combines these two features to provide complete information for each sub-branch. Different weights are assigned to each convolutional layer, allowing each sub-branch to carefully label non-important regions of different states. This ensures that the obtained state component is accurate and facilitates the subsequent acquisition of essential features.

[0051] Next, the essential part is calculated by eliminating the state part from the backbone feature. However, directly subtracting a state part from the complete feature may not be a good idea. This is because there are many different scenes in low-altitude drone imagery, and the differences between all input image pairs are more or less different. For some input pairs, the invariant regions emphasized by the state part are almost the same, while for other input pairs, these regions may be very different. For this reason, although the state part can locate the basically invariant regions, the contribution of the state part to the overall semantic variation is different. Therefore, in the current case, the importance of the state part must be carefully considered to correctly obtain the essential part. Based on this motivation, a method is designed to calculate the channel and pixel importance weights of each state part, and remove it from the backbone feature with the help of this importance weight.

[0052] Specifically, in each sub-branch, the complete backbone features and the generated state features are concatenated in a cascaded manner, and then adjusted by a convolutional layer to reach the same number of channels as the state part, i.e., N channels. In this step, the important channels of the state part will be further emphasized, which will result in a greater impact of these channels in the final weights. After that, two consecutive 3×3 convolutional layers will concentrate the information into M (M < N) channels and then restore it to N, so that the information of all channels can be thoroughly concentrated. By default, M is set to N / 2. When the channels are concentrated, the accessed channel content will be more general, which makes it easier for the model to analyze the spatial importance of all pixels, thus facilitating the improvement of the effectiveness of the final importance weights. Finally, after two consecutive 3×3 convolutional layers and then passing through the Sigmoid function, the values will be normalized to the interval [0, 1], and the importance weights of the channels and pixels will be output. In this process, the contribution of the state part to the differences between the backbone features of different channels and different spatial positions can be thoroughly analyzed and incorporated into the weights.

[0053] With the state part and the corresponding importance weights, the essential part features can be obtained in the following way

[0054] where the superscript j ∈ {1, 2} represents different sub-branches, and both subtraction and multiplication are pixel-wise. conv represents a 1×1 convolutional layer, and concat represents a concatenation operation.

[0055] To further enhance the features of each essential part, the method of this embodiment adds a deformable convolutional layer. In the default configuration, only the fifth-layer essential feature pair uses this deformable convolutional layer because it has the most advanced, most complete, and highest-level semantics and shoulders the responsibility of comprehensively guiding the prediction.

[0056] Since the change detection task aims to locate the changed regions of two images, the method of this embodiment calculates the differences between all essential pairs D i , i ∈ {2, 3, 4, 5} in the second stage to the fifth stage, as well as the differences between the backbone features D1 in the first stage, i.e.:

[0057] The decoding module fuses the change features obtained at each layer layer by layer. Starting from D5, it is first upsampled by a factor of 2 to align it with D4 and cascaded with D4. The connection is passed through a 3×3 convolution layer to fuse the features of the two different semantic levels and reduce the aliasing effect caused by different resolutions. Then, the output data is upsampled by a factor of 2 and connected to D3, and a 3×3 convolution layer is used to fuse the two feature maps. Then, the output data is upsampled by a factor of 2 and connected to D2, and a 3×3 convolution layer is used to fuse the two feature maps. Then, the output data is upsampled by a factor of 2 and connected to D1, and a 3×3 convolution layer is used to fuse the two feature maps to obtain the feature map after all features are fused. Finally, a 1×1 convolution layer is used to obtain the final change prediction map.

[0058] It should be noted that except for the last 1×1 convolution layer in the decoding module, all convolution layers are followed by a batch normalization layer to help convergence, and a Leaky ReLU function is used as the activation function.

[0059] Although the model proposed in this embodiment achieves a relatively good performance level when trained on the current dataset, it is believed that it never truly understands the concept of "change" and instead only attempts to find differences with certain specific features, which is considered to be a form of overfitting to some extent. Between a pair of input images, all regions have some differences in most cases, but the network never knows that they were taken in the same place. It can roughly locate the areas of change because the differences in these areas are somewhat different from the differences in areas with no essential changes. Obviously, this is not the best strategy for training the model, so it is necessary to let the network grasp the concept of "change" to fully tap the capabilities of deep neural networks.

[0060] Therefore, this embodiment proposes a solution: adding pairs of images with identical backgrounds and different foregrounds to assist in training and constantly remind the network. The identical backgrounds remind the network that the images depict the same location, while the different foreground objects cause essential changes in certain areas. This allows the deep network to focus on the nature of these changes.

[0061] Specifically, background images are first collected from the internet as a set of t0 images. Then, foreground regions are cropped from images in the salient object detection dataset DUTS and attached to the collected background images as t1 images, paired with the original background images. Because salient object detection is a task aimed at detecting salient regions, and these salient regions are often the same as foreground regions, the pixel-level annotations in the salient object detection dataset can be directly used for cropping, and the ground truth map GT of the salient object detection is exactly the GT of the synthesized virtual dataset. In this way, the goal of expanding the change detection training data can be achieved with very little additional annotation time. The dataset generated in this way may not be completely logical to humans, but it can provide the network with sufficiently distinct change semantics, forcing the network to understand the meaning of the change.

[0062] The loss function of this embodiment has two parts. The first part is the cross-entropy loss between the predicted value and the ground truth. In pixel classification tasks, the number of pixels in the unchanged area is significantly greater than that in the changed area, which leads to class imbalance. Therefore, this embodiment uses a weighted cross-entropy loss, where the weight of each class is the proportion of pixels belonging to the other class.

[0063] In addition, this embodiment adds a contrast loss for essential feature pairs at the fourth and fifth layers. This can push the essential features of the unchanged regions closer together, while pulling the features of the essentially changed regions further apart, making the changed regions easier for the subsequent decoder to construct, thereby improving the quality of the final prediction. This contrast loss Cont can be calculated as follows:

[0064] Among them, D x,y Represents the difference between the pixels at (x, y) in the intrinsic pair. For simplicity, the subscript i representing the feature level is removed here. GT x,y is the ground truth value of the pixel, 1 indicates a change, 0 indicates no change, and m is the margin. When the Euclidean distance between the pixel pairs is greater than m, they are considered sufficiently dissimilar and the loss becomes 0.

[0065] The complete loss function can be expressed as follows: L = CE(pred,GT) + Cont(D,GT) (5)

[0066] Where CE stands for weighted cross entropy and pred is the predicted change graph.

[0067] Figure 3 shows the detection effect of the method of this embodiment. It can be seen that the method of this embodiment performs well and is significantly better than other methods. In particular, the method of this embodiment can better suppress the expression of non-changing areas and find areas where essential changes have occurred. For example, the fallen leaves in the first picture are caused by the season, which is not the change that this embodiment wants to find, so the method of this embodiment suppresses them all very well. In addition, the method of this embodiment also performs well in positioning and details. For example, for the third and fourth rows in Figure 3.

[0068] Example 2

[0069] The purpose of this embodiment is to provide a low-altitude UAV image change detection system, including:

[0070] Acquisition module: acquires the time-phase image pair of the area to be detected;

[0071] Change detection module: The acquired time-phase image pairs are input into the trained change detection module to obtain the change detection results of the area to be detected;

[0072] The change detection module includes a feature extraction module, an interactive decoupling module and a decoding module; the feature extraction module is used to extract backbone features from the phase image pairs respectively to obtain multi-level backbone feature pairs;

[0073] The interactive decoupling module is used to extract state features from the backbone feature pairs to obtain state feature pairs. Based on the importance of the state features in the time phase image, the state feature weights are obtained. Based on the state feature pairs, the backbone feature pairs, and the state feature weights corresponding to the state feature pairs, the essential feature pairs are obtained.

[0074] The decoding module is used to cascade and fuse different essential feature pairs to obtain the change detection results of the area to be detected.

[0075] Example 3

[0076] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the program.

[0077] Example 4

[0078] The purpose of this embodiment is to provide a computer-readable storage medium.

[0079] A computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the above method.

[0080] The steps involved in the apparatuses of Examples 2, 3, and 4 above correspond to those of Method Example 1. For detailed implementations, please refer to the relevant description of Example 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any method of the present invention.

[0081] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0082] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for detecting changes in low-altitude drone images, characterized in that: include: Acquire a time-phase image pair of the area to be detected, input the acquired time-phase image pair into a trained change detection module, and obtain a change detection result of the area to be detected; The change detection module includes a feature extraction module, an interactive decoupling module and a decoding module; the feature extraction module is used to extract backbone features from the phase image pairs respectively to obtain multi-level backbone feature pairs; The interactive decoupling module is used to extract state features from the backbone feature pairs respectively to obtain state feature pairs, and obtain state feature weights based on the importance of the state features in the phase image; and obtain essential feature pairs according to the state feature pairs, the backbone feature pairs, and the state feature weights corresponding to the state feature pairs respectively; The decoding module is used to cascade and fuse different essential feature pairs to obtain the change detection result of the area to be detected; The state features are extracted from the backbone feature pairs respectively to obtain the state feature pairs. Based on the importance of the state features in the phase image, the state feature weights are obtained, which are specifically: The backbone feature pairs are cascaded, and the cascaded backbone feature pairs are respectively passed through the convolution layer to reduce the number of channels to obtain the state feature pairs; According to the importance of the state feature in the phase image, the number of channels of the state feature is adjusted to obtain the state feature weight; According to the importance of the state feature in the phase image, the number of channels of the state feature is adjusted to obtain the state feature weight, which specifically includes: Cascade each state in the state-feature pair with the corresponding backbone feature; The cascade result is passed through a 3×3 convolutional layer to adjust the number of channels to be the same as the number of channels of the state feature pair; The cascaded result after adjusting the number of channels is passed through two 3×3 convolutional layers in sequence to make the number of channels the same as the number of channels of the backbone feature pair, and then normalized by an activation function to obtain the state feature weight; According to the state feature pair, the backbone feature pair, and the state feature weights corresponding to the state feature pair, an essential feature pair is obtained, specifically: Performing a pixel-level multiplication operation on the state feature weight and the corresponding state feature to obtain a first result; Performing a pixel-level subtraction operation on the first result and the backbone feature corresponding to the state feature weight to obtain a second result; The second result is cascaded with the backbone feature corresponding to the state feature weight and then passed through a 1×1 convolution layer to obtain the essential feature.

2. A low-altitude UAV image change detection method as claimed in claim 1, characterized in that: The decoding module is used to cascade fuse different essential feature pairs to obtain the change detection result of the area to be detected. Specifically, the essential feature pairs are upsampled and then fused with adjacent essential feature pairs layer by layer, and the final fused result is passed through a 1×1 convolution layer to obtain the change detection result of the area to be detected.

3. A low-altitude UAV image change detection method as claimed in claim 1, characterized in that: The feature extraction module is a twin convolutional neural network composed of two VGG16 networks, which is used to extract multi-level backbone feature pairs from the phase image pairs.

4. A low-altitude UAV image change detection method as claimed in claim 1, characterized in that: The training of the change detection module includes: Get the background image and the foreground area cut out from the image of the salient object detection dataset; Pasting the acquired foreground area into the background image and forming an image pair with the background image; The image pairs and the phase image pair data set together constitute a training set, and the training set is used to train the change detection module.

5. A low-altitude UAV image change detection system, characterized in that: include: Acquisition module: acquires the time-phase image pair of the area to be detected; Change detection module: input the acquired time phase image pair into the trained remote sensing change detection module to obtain the change detection result of the area to be detected; The change detection module includes a feature extraction module, an interactive decoupling module and a decoding module; the feature extraction module is used to extract backbone features from the phase image pairs respectively to obtain multi-level backbone feature pairs; The interactive decoupling module is used to extract state features from the backbone feature pairs respectively to obtain state feature pairs, and obtain state feature weights based on the importance of the state features in the phase image; and obtain essential feature pairs according to the state feature pairs, the backbone feature pairs, and the state feature weights corresponding to the state feature pairs respectively; The decoding module is used to cascade and fuse different essential feature pairs to obtain the change detection result of the area to be detected; The state features are extracted from the backbone feature pairs respectively to obtain the state feature pairs. Based on the importance of the state features in the phase image, the state feature weights are obtained, which are specifically: The backbone feature pairs are cascaded, and the cascaded backbone feature pairs are respectively passed through the convolution layer to reduce the number of channels to obtain the state feature pairs; According to the importance of the state feature in the phase image, the state feature is divided into the channel number Adjustment of , and obtain the state feature weight; According to the importance of the state feature in the phase image, the number of channels of the state feature is adjusted to obtain the state feature weight, which specifically includes: Cascade each state in the state-feature pair with the corresponding backbone feature; The cascade result is passed through a 3×3 convolutional layer to adjust the number of channels to be the same as the number of channels of the state feature pair; The cascaded result after adjusting the number of channels is passed through two 3×3 convolutional layers in sequence to make the number of channels the same as the number of channels of the backbone feature pair, and then normalized by an activation function to obtain the state feature weight; According to the state feature pair, the backbone feature pair, and the state feature weights corresponding to the state feature pair, an essential feature pair is obtained, specifically: Performing a pixel-level multiplication operation on the state feature weight and the corresponding state feature to obtain a first result; Performing a pixel-level subtraction operation on the first result and the backbone feature corresponding to the state feature weight to obtain a second result; The second result is cascaded with the backbone feature corresponding to the state feature weight and then passed through a 1×1 convolution layer to obtain the essential feature.

6. A computer device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate through the bus. When the machine-readable instructions are executed by the processor, a low-altitude unmanned aerial vehicle image change detection method as described in any one of claims 1 to 4 is executed.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for detecting changes in low-altitude UAV images as described in any one of claims 1 to 4 is executed.

Images