A small target detection method based on high-level and low-level feature fusion

Abstract

The present invention discloses a small target detection method based on high- and low-level feature fusion, comprising: obtaining a target image to be detected as a small target; using a preset target detection neural network based on a multi-scale fusion strategy to detect a small target on the target image; wherein, in the target detection neural network, a high- and low-level feature fusion module is used to realize the fusion of low-level features and high-level features; the high- and low-level feature fusion module is used to pre-process the low-level features and the high-level features respectively, so that the dimensions of the low-level features and the high-level features are the same; the low-level features after pre-processing are first channel compressed and then channel expanded to retain the useful information of the small target and filter out useless information, so as to obtain channel compression and expansion features; the channel compression and expansion features are multiplied by the pre-processed high-level features, and the product result is added to the pre-processed high-level features as the high- and low-level fusion features. The present invention further improves the accuracy of small target detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision target detection, and in particular relates to a small target detection method based on high- and low-level feature fusion. Background Art

[0002] In the field of target detection, small targets usually occupy less than 30 pixels in the image, and the effective information they contain is very limited, making them difficult to detect effectively. Secondly, small targets are easily disturbed by external factors such as the background environment, including exposed soil, forests, hills, and oceans, which further increases the difficulty of small target detection.

[0003] Using a trained neural network for target detection is a commonly used target detection method; however, since small targets have fewer available pixels than regular targets, it is difficult to extract better features, and as the number of network layers increases, the feature information and position information of small targets are gradually lost.

[0004] In response to the difficulty of small target detection, various strategies have been proposed in the existing technology to solve the problem of small target detection, including multi-scale fusion strategy, context information extraction strategy, weak supervision learning with small sample size, etc., which have greatly improved the model performance.

[0005] Among them, for the multi-scale fusion strategy, FPN (Feature Pyramid Networks) is one of the pioneering works, which proposed to construct a feature pyramid for feature fusion; for example, in the YOLO series of networks, from YOLOv3 to the current new version of YOLOv7, multi-scale fusion has appeared.

[0006] Figure 1 The structure diagram of the existing FPN is shown; Figure 1 As can be seen in the figure, the existing FPN realizes multi-scale feature fusion by first upsampling the low-level features and then adding them to the high-level features to form a residual structure, thereby realizing the fusion of high- and low-level features.

[0007] However, although the above multi-scale fusion strategy has performed well when facing small targets, its performance in detection accuracy still needs to be further improved. Summary of the invention

[0008] In order to solve the above problems existing in the prior art, the present invention provides a small target detection method based on high-level and low-level feature fusion.

[0009] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0010] A small target detection method based on high-level and low-level feature fusion, comprising:

[0011] Acquire a target image to be detected as a small target;

[0012] Performing small target detection on the target image using a preset target detection neural network based on a multi-scale fusion strategy;

[0013] Among them, the target detection neural network realizes the fusion of low-level features and high-level features based on a high-low-level feature fusion module; the high-low-level feature fusion module is used to preprocess the low-level features and high-level features respectively so that the dimensions of the low-level features and the high-level features are the same; the preprocessed low-level features are first channel compressed and then channel expanded to retain the useful information of small targets and filter out useless information, so as to obtain channel compression and expansion features; the channel compression and expansion features are multiplied with the preprocessed high-level features, and the product result is added to the preprocessed high-level features as high-low-level fusion features.

[0014] Optionally, the target detection neural network includes:

[0015] YOLOv5, YOLOv7 or ResNet50 that implements high- and low-layer feature fusion based on the high- and low-layer feature fusion module.

[0016] Optionally, the high- and low-layer feature fusion module is applied to the feature pyramid network FPN and / or the path aggregation network PAN of the target detection neural network.

[0017] Optionally, the target detection neural network also includes a feature splicing module;

[0018] The feature splicing module is used to splice the lowest-level low-layer features in the target detection neural network with the highest-level high-low-layer fusion features.

[0019] Optionally, the low-level features are preprocessed by:

[0020]

[0021] in, represents low-level features, Fconv(·) represents the convolution operator, represents the first The learning parameters of the convolution kernels, The number of convolution kernel channels representing low-level features; represents convolution, δ(·) represents SiLU function, and M represents the low-level features after preprocessing.

[0022] Optionally, the channel compression method for low-level features includes:

[0023]

[0024] in, represents the low-level features after preprocessing, ε represents the channel compression factor, Fsqueeze(·) represents the channel compression function, and F represents the low-level features after channel compression.

[0025] Optionally, the channel expansion method for the low-level features after channel compression includes:

[0026]

[0027] in, represents the low-level features after channel compression, l∈[1,C], C represents the number of convolution kernel channels of the target detection neural network, Fexpand(·) represents the channel expansion function, σ(·) represents the sigmoid function, Represents the channel compression and expansion characteristics.

[0028] Optionally, the method of preprocessing the high-level features includes:

[0029] The high-level features are preprocessed using a nearest neighbor interpolation method.

[0030] Optionally, the target image includes: a remote sensing image.

[0031] In the small target detection method based on high- and low-layer feature fusion provided by the present invention, a high- and low-layer feature fusion module is used in the target detection neural network based on a multi-scale fusion strategy to realize the fusion of low-layer features and high-layer features; the high- and low-layer feature fusion module effectively retains the useful information of the small target and filters out useless information by first compressing the channel of the low-layer features and then expanding the channel; then, a spatial channel information interaction method is used to realize the fusion of high- and low-layer features, giving enough attention to the spatial channel, taking into account the relationship between the channels and the correlation between the high- and low-layer feature maps, specifically, multiplying the channel compression and expansion features with the high-layer features, and adding the product result with the high-layer features as the high- and low-layer fusion features. Based on the above operations, small target detection based on the target detection neural network can further improve the detection performance for small targets.

[0032] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the structure of an existing ResNet50 network;

[0034] Figure 2 It is a flow chart of a small target detection method based on high- and low-level feature fusion provided by an embodiment of the present invention;

[0035] Figure 3 yes Figure 2 A schematic diagram of the structure of the high- and low-level feature fusion module used in the target detection neural network in the method shown;

[0036] Figure 4 In the embodiment of the present invention, Figure 1 The ResNet50 shown in the figure uses Figure 3 The structural diagram of the high- and low-level feature fusion module shown;

[0037] Figure 5 The process of fusing three different scale features using high- and low-level feature fusion modules in FPN is shown in FIG.

[0038] Figure 6 It is the label distribution of the data set used in the simulation verification of the embodiment of the present invention. DETAILED DESCRIPTION

[0039] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0040] In order to further improve the accuracy of small target detection, the embodiment of the present invention provides a small target detection method based on high-level and low-level feature fusion, see Figure 2 As shown, the method comprises the following steps:

[0041] S10: Acquire a target image to be subjected to small target detection.

[0042] Here, there may be many specific types of target images, such as images taken by drones at high altitudes, or remote sensing images taken by radars, etc. In such target images, small targets may be pedestrians, vehicles, and buildings on the ground, or ships and buoys on the sea.

[0043] S20: Detect small targets in the target image using a preset target detection neural network based on a multi-scale fusion strategy.

[0044] Among them, the target detection neural network realizes the fusion of low-level features and high-level features based on a high-low-level feature fusion module; see Figure 3 As shown, the high-low feature fusion module is used to preprocess the low-level feature X and the high-level feature Y respectively to make the dimensions of the low-level feature and the high-level feature the same, and obtain the preprocessed low-level feature M and the preprocessed high-level feature The preprocessed low-level feature M is first channel compressed (to obtain F) and then channel expanded to retain the useful information of small targets and filter out useless information to obtain the channel compression and expansion feature The channel companding characteristics With the preprocessed high-level features Multiply and convert the product Add to the preprocessed high-level features Get high-low layer fusion features

[0045] Among them, the channel compression and expansion features are multiplied with the preprocessed high-level features, and the product result is added to the preprocessed high-level features to form a residual structure, so as to realize the spatial channel relationship interaction.

[0046] Exemplarily, the target detection neural network based on the above REBUF to achieve high- and low-level feature fusion may include: YOLOv5, YOLOv7 or ResNet50, but is certainly not limited thereto.

[0047] In order to make the layout of the specification clear, examples are given below to illustrate how to achieve high- and low-level feature fusion based on a high- and low-level feature fusion module in a target detection neural network.

[0048] In the above-mentioned high-level and low-level feature fusion module, the method of preprocessing the low-level features includes:

[0049]

[0050] in, represents low-level features, Fconv(·) represents the convolution operator, represents the first The learning parameters of the convolution kernels, Therefore, s represents the learning parameters of the sth convolution kernel of the target detection neural network, x s Represents the feature value corresponding to the sth convolution kernel channel in the low-level features; The number of convolution kernel channels representing low-level features; represents convolution, δ(·) represents SiLU function, and M represents the low-level features after preprocessing.

[0051] It is understandable that since the output of a single neuron in a neural network is generated by weighted summing of the outputs of all neurons in the previous stage, the interrelationships of the channels are middle.

[0052] Here, by preprocessing the low-level feature X using the Fconv function, the network's sensitivity to the useful information it contains can be improved, and useless information can be suppressed to facilitate subsequent operations.

[0053] Exemplarily, the method of preprocessing the high-level features in the high- and low-level feature fusion module may include: preprocessing the high-level features using a nearest neighbor interpolation method, but is certainly not limited thereto.

[0054] It can be understood that by processing features using the nearest neighbor interpolation method, both upsampling and downsampling of features can be achieved.

[0055] In the above-mentioned high- and low-level feature fusion module, the channel compression method for low-level features includes:

[0056]

[0057] in, represents the low-level features after preprocessing, Therefore, m h represents the feature value of the hth convolution kernel channel in the preprocessed low-level features, v h represents the learning parameters of the hth convolution kernel of the target detection neural network; Fsqueeze(·) represents the channel compression function, ε represents the channel compression factor, δ(·) represents the SiLU function, and F represents the low-level features after channel compression.

[0058] In the above-mentioned high- and low-level feature fusion module, the channel expansion method of the low-level features after channel compression includes:

[0059]

[0060] in, represents the low-level features after channel compression, l∈[1,C], so f l represents the eigenvalue of the lth convolution kernel channel in the low-level features after channel compression, v l represents the learning parameters of the lth convolution kernel of the target detection neural network; Fexpand(·) represents the channel expansion function, C represents the number of convolution kernel channels of the target detection neural network, σ(·) represents the sigmoid function, Indicates the channel compression and expansion characteristics.

[0061] Here, in order to efficiently utilize the underlying semantic information in the low-level features after channel compression, the embodiment of the present invention further expands the channels, with the purpose of more fully capturing the relationship between channels. In addition, in order to ensure that multiple channels are strengthened rather than a single channel, a sigmoid function is used to implement channel expansion during channel expansion.

[0062] It is understandable that in order to retain the useful information of small targets to the greatest extent possible while filtering out useless information such as the background, channel compression and expansion are mainly performed on the channel dimension.

[0063] The following is an example of how to implement high- and low-layer feature fusion based on a high- and low-layer feature fusion module in a target detection neural network.

[0064] For example, Figure 4 A schematic diagram of the structure of ResNet50 based on the high-low layer feature fusion module is shown in FIG. Figure 1 As can be seen from the structure of the existing ResNet50 shown in FIG, in the embodiment of the present invention, a high-low layer feature fusion module can be used in the FPN (feature pyramid network) module of ResNet50 to realize high-low layer feature fusion; specifically, Figure 1 In the existing FPN, the high-level features (Y5, Y4, Y3) are first upsampled (Upsample), and then directly added to the corresponding low-level features (X4, X3, X2), so as to achieve high-level and low-level feature fusion; in the embodiment of the present invention, see Figure 4 As shown in the figure, the un-upsampled high-level features (Y5, Y4, Y3) are respectively input into three high- and low-level feature fusion modules (Merge1, Merge2, Merge3), and the high-level features will be preprocessed (upsampled) inside the high- and low-level feature fusion modules; the low-level features (X4, X3, X2) are respectively input into three high- and low-level feature fusion modules (Merge1, Merge2, Merge3), so that the fusion of Y5 and X4 is realized through Merge1, the fusion of Y4 and X3 is realized through Merge2, and the fusion of Y3 and X2 is realized through Merge3.

[0065] In addition, the high- and low-layer feature fusion module can also be used in the PAN module in ResNet50 to achieve high- and low-layer feature fusion. The usage method is basically the same as that in FPN, except that in FAN, the high-level features are downsampled inside the high- and low-level feature fusion module.

[0066] In one embodiment, the above-mentioned target detection neural network may also include a feature splicing module; the feature splicing module is used to splice the lowest level low-level features in the target detection neural network with the highest level high-low-level fusion features.

[0067] For details, see Figure 4 As shown in the figure, the lowest level feature in the target detection neural network is X4, and the highest level high-low level fusion feature is the high-low level fusion feature obtained by fusing Y3 and X2 output by Merge3, ​​which is denoted as P here. The feature splicing total module can splice X4 with P.

[0068] It can be understood that based on the practice of recursively fusing multi-scale features in FPN / PAN, when the high- and low-layer feature fusion module is applied to FPN / PAN, the multi-scale features input to FPN / PAN can also be fused recursively.

[0069] For example, take FPN that outputs three different scale features as an example. Figure 5 The process of fusing them using REBUF in FPN is shown in Figure 2, which specifically includes:

[0070] (1) Use the convolution operator Fconv to perform low-level features Preprocess and obtain the preprocessed low-level features M∈R C×H×W ; Use the nearest neighbor interpolation method to extract high-level features Perform preprocessing to obtain high-level features after preprocessing

[0071] (2) Use the Fsqueeze function to perform channel compression on the preprocessed low-level features M to obtain the low-level features after channel compression Then use the Fexpand function to expand the channel of F and obtain the channel compression expansion feature

[0072] (3) Compress and expand the channel characteristics With the preprocessed high-level features Multiply them and add the product to Add them together and get in

[0073] The above steps (1) to (3) use a high-low feature fusion module in FPN to realize low-level features. With high-level features fusion of.

[0074] (4) Use the convolution operator Fconv to perform low-level features Preprocessing is performed and the preprocessed Use the Fsqueeze function and the Fexpand function to compress and expand the channel in turn to obtain the channel compression and expansion characteristics At the same time, the nearest neighbor interpolation method is used to Perform preprocessing to obtain high-level features after preprocessing

[0075] (5) Channel compression and expansion features With the preprocessed high-level features Multiply and combine the product with the high-level features Add them together and get in

[0076] The above steps (4) to (5) are implemented by using another high-low layer feature fusion module in FPN.

[0077] Low-level features With high-level features fusion of.

[0078] (6) Low-level features and Perform splicing (Fcat), and get Z∈R C×H×W It not only has the sensitive and useful information contained in the deeper high-level features Y, but also has the sensitive and useful information contained in the deeper high-level features T. Therefore, it can effectively detect small targets based on Z and improve the accuracy of small target detection.

[0079] This step (6) uses a feature concatenation module in FPN to concatenate the lowest-level features High-low layer fusion features with the highest layer Make splicing.

[0080] Similarly, the method of implementing high- and low-layer feature fusion based on REBUF in YOLOv5 or YOLOv7 can refer to the above-mentioned embodiment of implementing high- and low-layer feature fusion based on REBUF in ResNet50, and the embodiments of the present invention will not be described in detail.

[0081] In summary, in YOLOv5, YOLOv7 or ResNet50, the module for implementing high- and low-layer feature fusion based on REBUF may include an FPN module and / or a PAN module.

[0082] It should be noted that the application scenarios of the embodiments of the present invention are relatively broad and are not limited to feature inputs of 2 or 3 scales. They also have good performance even when faced with multi-scale feature inputs.

[0083] In the small target detection method based on the fusion of high- and low-level features provided by the embodiment of the present invention, by first compressing the low-level features and then expanding the channels, the useful information of the small target is effectively retained and the useless information is filtered out; then, the spatial channel information interaction method is used to realize the fusion of high- and low-level features, giving enough attention to the spatial channels, specifically, multiplying the channel compression and expansion features with the high-level features, and adding the product result to the high-level features. Therefore, the target detection neural network used in the embodiment of the present invention takes into account the relationship between channels and the correlation between high- and low-level feature maps, so small target detection based on the target detection neural network can further improve the detection performance for small targets. In particular, the use of the embodiment of the present invention can better solve the problem of small target detection in remote sensing image environments.

[0084] The beneficial effects of the embodiments of the present invention are further illustrated by simulation experiments below.

[0085] Experimental environment

[0086] The experimental platform is PyTorch version 1.10.0, CUDA11.3, GPU (graphics processor) model is RTX3090, hardware memory is 24G, and the compilation language is Python 3.8.

[0087] Experimental Dataset

[0088] The dataset uses the DoTAv1.0 dataset

[21] , which contains aerial images of 2,000 cities and a total of 15 types of targets, with more than 190,000 fully labeled targets, each of which consists of 8 position parameters (x1, y1, x2, y2, x3, y3, x4, y4). In this dataset, five types of smaller targets are selected, including: small vehicle, large vehicle, plane, storage tank, and ship, to form a dataset of small targets. Since the image size in the DoTAv1.0 dataset is large, it is not suitable for direct detection. Therefore, it is cropped to 1024×1024. After cropping, a total of 8,624 images are obtained, and then the data subsets are split according to the ratio of 8:2:2 for training set: validation set: test set. Among them, 5,176 are used for training, and 1,724 are used for validation and testing. The label distribution is as follows Figure 6 As shown, the horizontal axes 0-4 correspond to small vehicle, large vehicle, plane, storage tank, and ship respectively.

[0089] Evaluation indicators

[0090] In order to evaluate the effectiveness of the object detection method, accuracy, recall, average precision (AP), mean average precision (mAP) and model inference time are used to evaluate the algorithm.

[0091] The accuracy is the proportion of correct samples in the total test, and the calculation formula is:

[0092]

[0093] Among them, true positive (TP) is the number of accurately predicted positive samples, and false positive (FP) is the number of negative samples that are mistakenly detected as positive samples. Recall represents the proportion of accurately predicted positive samples, which is defined as:

[0094]

[0095] Here, false negatives (FN) is the number of positive samples that are predicted as negative samples.

[0096] AP is the area enclosed by the accuracy curve and the recall rate on the x-axis, which is expressed as:

[0097]

[0098] Among them, y is the recall curve under different intersection ratio thresholds, and P(y) is the accuracy corresponding to the recall rate y.

[0099] mAP refers to the average value of all AP categories, and the calculation formula is:

[0100]

[0101] Where N is the number of categories, and N=5 in the experiment.

[0102] Experimental procedures

[0103] Based on YOLOv5 and YOLOv7, the small target detection performance with and without REBUF is compared. The experimental results on the test dataset are shown in the following table:

[0104]

[0105] Among them, Yolov5 and Yolov7 represent the existing Yolov5 and Yolov7 respectively without using the high- and low-layer feature fusion module and the feature splicing total module; Yolov5-REBUF and Yolov7-REBUF represent Yolov5 and Yolov7 respectively using the high- and low-layer feature fusion module and the feature splicing total module.

[0106] As can be seen from the table above, after using the high- and low-layer feature fusion module and the feature concatenation module, the mAP on YOLOv5 increased from 90.2% to 90.5%, and the inference time increased from 4.7ms to 5.0ms for processing 32 batches of images, and the increase in time consumption is almost negligible. After using the high- and low-layer feature fusion module and the feature concatenation module, the mAP on YOLOv7 increased from 90.7% to 91.2%, and the inference speed increased from 4.4ms to 4.5ms for processing 32 batches of images, and the increase in time consumption is almost negligible.

[0107] The method provided in the embodiment of the present invention can be applied to electronic devices. Specifically, the electronic device can be: a desktop computer, a portable computer, an intelligent mobile terminal, a server, etc. This is not limited here, and any electronic device that can implement the present invention belongs to the protection scope of the present invention.

[0108] The present invention also provides a computer-readable storage medium in which a computer program is stored, and when the computer program is executed by a processor, any of the above-mentioned small target detection methods based on high- and low-level feature fusion or the steps included in the above-mentioned REBUF is implemented.

[0109] Optionally, the computer-readable storage medium may be a non-volatile memory (NVM), such as at least one disk memory.

[0110] Optionally, the computer-readable memory may also be at least one storage device located away from the aforementioned processor.

[0111] In another embodiment of the present invention, a computer program product comprising instructions is also provided. When the computer is run on a computer, the computer executes any of the above-mentioned small target detection methods based on high- and low-level feature fusion or the steps included in the above-mentioned REBUF.

[0112] It should be noted that, for the storage medium / computer program product embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0113] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0114] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification.

[0115] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art may understand and implement other variations of the disclosed embodiments by viewing the drawings and the disclosed content.

[0116] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A small target detection method based on high- and low-level feature fusion, characterized in that: include: Acquire a target image to be detected as a small target; Performing small target detection on the target image using a preset target detection neural network based on a multi-scale fusion strategy; Among them, the target detection neural network realizes the fusion of low-level features and high-level features based on a high-low-level feature fusion module; the high-low-level feature fusion module is used to preprocess the low-level features and high-level features respectively so that the dimensions of the low-level features and the high-level features are the same; the low-level features after preprocessing are first channel compressed and then channel expanded to retain the useful information of small targets and filter out useless information, so as to obtain channel compression and expansion features; the channel compression and expansion features are multiplied with the preprocessed high-level features, and the product result is added to the preprocessed high-level features as the high-low-level fusion features; Methods for preprocessing low-level features include: in, represents low-level features, Fconv(·) represents the convolution operator, represents the first The learning parameters of the convolution kernels, The number of convolution kernel channels representing low-level features; represents convolution, δ(·) represents SiLU function, and M represents the low-level features after preprocessing; Ways to compress channels of low-level features include: in, represents the low-level features after preprocessing, ε represents the channel compression factor, Fsqueeze(·) represents the channel compression function, and F represents the low-level features after channel compression; The methods for expanding the channels of low-level features after channel compression include: in, represents the low-level features after channel compression, l∈[1,C], C represents the number of convolution kernel channels of the target detection neural network, Fexpand(·) represents the channel expansion function, σ(·) represents the sigmoid function, Represents the channel compression and expansion characteristics.

2. The small target detection method based on high- and low-level feature fusion according to claim 1 is characterized in that: The target detection neural network comprises: YOLOv5, YOLOv7 or ResNet50 that implements high- and low-layer feature fusion based on the high- and low-layer feature fusion module.

3. The small target detection method based on high- and low-level feature fusion according to claim 2 is characterized in that: The high- and low-layer feature fusion module is applied to the feature pyramid network FPN and / or the path aggregation network PAN of the target detection neural network.

4. The method according to claim 1, characterized in that: The target detection neural network also includes a feature splicing module; The feature splicing module is used to splice the lowest layer low-level features in the target detection neural network with the highest layer high-low layer fusion features; The target detection neural network includes a pyramid network FPN; in the pyramid network FPN, the low-level features of each layer and the high-level features of the upper layer are spliced ​​by a high-low layer feature fusion module to form a high-low layer fusion feature.

5. The method according to claim 1, characterized in that The method of preprocessing the high-level features includes: The high-level features are preprocessed using the nearest neighbor interpolation method.

6. The method according to claim 1, characterized in that The target image includes: a remote sensing image.

Images