Data enhancement method and device for large-format sparse sample remote sensing image target detection

Abstract

The embodiment of the application provides a data enhancement method and device for large-format sparse sample remote sensing image target detection, and relates to the technical fields of image processing and target detection. The method completely detects a target as an expansion object of data enhancement, ensures that more negative samples are input into network training while positive samples are increased, and thus more accurate target detection results are obtained. Therefore, the recognition of the background is improved, the false detection rate is reduced, the detection precision is improved, and the mixed area on the image is utilized to generate a feature map similar to an original target map, so that the detection precision is improved, the method is easy to implement, a large amount of calculation is not required, and the robustness of the algorithm itself is not affected.

Description

Technical Field

[0001] This invention relates to the fields of image processing and target detection technology, and in particular to a data augmentation method and apparatus for target detection in large-format sparse sample remote sensing images. Background Technology

[0002] In recent years, with the development of high-resolution optical satellite imaging technology and artificial intelligence, modern remote sensing technology has enabled multi-band, high-speed, and dynamic observation. Research in fields such as ordinary optical remote sensing images, hyperspectral images, and synthetic aperture radar images is receiving widespread attention and is being applied in agriculture, marine and forestry fields.

[0003] Object detection, a classic research area in image processing, is also facing new challenges and breakthroughs. Some scholars have applied deep learning to object detection in remote sensing images and achieved some success. However, how to better implement processing methods suitable for remote sensing images has become a new challenge.

[0004] As the resolution of remote sensing images continues to increase, visible light remote sensing images are characterized by large image size and high resolution. Especially in the field of ship target detection, the large image size and sparse target distribution often necessitate scale segmentation before network training. Due to the sparse distribution of ships on the sea surface, most segmented images lack targets. This results in a significant disparity between the number of positive and negative samples in the training set, with much of the sea and land background, clouds, and fog unusable as negative samples. Consequently, the trained network often exhibits numerous false positives and false negatives during target detection, leading to high false positive and false negative rates in the practical application of visible light remote sensing image target detection networks.

[0005] Data augmentation is a method to expand the size of the data sample, which can enhance the generalization ability of the model. Common data augmentation methods include random cropping, image flipping, and random scaling. However, these methods cannot effectively solve the problems encountered in practical remote sensing target detection tasks, such as low detection accuracy and overfitting due to small and sparse targets and few samples. Therefore, enriching the data and increasing its diversity are particularly important in today's rapidly developing remote sensing technology. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to propose a data augmentation method and apparatus for target detection in large-format sparse sample remote sensing images, in order to at least partially solve the shortcomings of the prior art.

[0007] The first aspect of this invention provides a data augmentation method for target detection in large-format sparse sample remote sensing images, comprising: acquiring a dataset of images to be detected; classifying the dataset of images to be detected to obtain a first image dataset containing the target image and a second image dataset without the target image; training a network using the first image dataset to obtain a first detection network; performing verification inference on the image dataset without the target image using the first detection network to obtain a first false positive image dataset, wherein the first false positive image dataset contains the target image; cropping the target image in the first image dataset and randomly pasting the cropped target image into the false positive images in the first false positive image dataset to obtain a second false positive image dataset; and combining the first image dataset and the second false positive image dataset... The first detection network is trained using the image dataset to obtain the second detection network. The second detection network is then used to perform validation inference on the image dataset to be detected, resulting in a third false positive image dataset, which contains the target image. The target image in the third false positive image dataset is cropped, and the cropped target image is randomly pasted into images in the first image dataset to obtain an enhanced target image dataset. The target image dataset, the enhanced target image dataset, and the first false positive image dataset are combined to form the final image dataset. The final image dataset is then trained using the second detection network to obtain the final target detection network. Finally, the remote sensing image to be detected is input into the target detection network, and the detection results of the target image are output.

[0008] Furthermore, the dataset of images to be detected includes multiple remote sensing images of ships taken by satellites under different imaging conditions, sea conditions, and open ocean surfaces.

[0009] Furthermore, before classifying the image dataset to be detected, the process includes: acquiring a remote sensing image with a preset frame; segmenting the remote sensing image according to the preset frame with a preset size, preset overlapping area, and preset edge processing method to obtain multiple segmented images; and labeling the segmented images containing the target to be detected to obtain an image dataset with labeling information.

[0010] The preset size is 1024×1024, the preset overlap area is 200, and the preset edge processing method includes filling the edges of images with 0 pixels if they are missing pixels.

[0011] Furthermore, the target image dataset to be tested is used for network training to obtain the first detection network, including: using a rotated target detector to train the network on the target image dataset to be tested.

[0012] Furthermore, the target images in the target image dataset are cropped, and the cropped target images are randomly pasted into the images in the false positive image dataset to obtain the second false positive image dataset. This includes: traversing the annotation information of the target image dataset, converting the coordinates of the rotated rectangles in the annotation information into the coordinates of the bounding horizontal boxes, cropping the target images in the target image dataset using horizontal boxes, and randomly selecting false positive images from the first false positive image dataset for pasting to form the second false positive image dataset.

[0013] Furthermore, the target image in the third false positive image dataset is cropped, and the cropped image is selected at a certain density position. The selected image is then randomly pasted into the image in the target image dataset to obtain an enhanced target image dataset. This process includes: cropping the horizontal bounding box of the target image in the third false positive image dataset and the pixels within a preset range around the horizontal bounding box; selecting the cropped image at a certain density position and randomly pasting the selected image into the image in the target image dataset to obtain an enhanced target image dataset. If the target image is occluded by more than a preset occlusion percentage during pasting, the target image is placed on the top layer.

[0014] A second aspect of the present invention provides a data augmentation apparatus for target detection in large-format sparse sample remote sensing images, comprising:

[0015] The image preprocessing module acquires the image dataset to be detected, classifies it to obtain a first image dataset containing the target image and a second image dataset without the target image; the first network training module trains the network using the first image dataset to obtain a first detection network; the first verification and inference module uses the first detection network to perform verification and inference on the image dataset without the target image to obtain a first false positive image dataset, wherein the first false positive image dataset contains the target image; the first data augmentation module crops the target image in the first image dataset and randomly pastes the cropped target image into the false positive images in the first false positive image dataset to obtain a second false positive image dataset; the second network training module inputs the first image dataset and the second false positive image dataset into the first detection network. The system is divided into three modules: a second detection network and a third false positive image dataset. The second detection network is trained to obtain a second detection network. A second verification and inference module is used to perform verification and inference on the image dataset to be detected, resulting in a third false positive image dataset containing the target image. A second data augmentation module is used to crop the target image from the third false positive image dataset and randomly paste the cropped target image into the images in the first image dataset, resulting in an augmented target image dataset. A third network training module is used to combine the target image dataset, the augmented target image dataset, and the first false positive image dataset into a final image dataset. This final image dataset is then input into the second detection network for training, resulting in the final target detection network. A target detection module is used to input the remote sensing image to be detected into the target detection network and output the detection result of the target image.

[0016] A third aspect of the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the above-described non-intrusive load detection method based on harmonic wavelet decomposition.

[0017] As can be seen from the above technical solutions, the data augmentation method and apparatus for target detection in large-format sparse sample remote sensing images provided by the present invention have the following beneficial effects:

[0018] (1) The data augmentation method for target detection in large-format sparse sample remote sensing images provided by the present invention takes the complete detection target as the extended object of data augmentation, ensuring that more negative samples are input into the network training while increasing positive samples, thereby increasing the diversity of samples, improving background recognition, reducing false detection rate, and improving detection accuracy.

[0019] (2) The data augmentation method for target detection in large-format sparse sample remote sensing images provided by this invention generates a feature map similar to the original target map by utilizing mixed regions on the image, thereby improving detection accuracy. It is easy to implement, does not incur significant computational overhead, and does not affect the robustness of the algorithm itself. Attached Figure Description

[0020] Figure 1a A flowchart illustrating a data augmentation method for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention is shown.

[0021] Figure 1b Another flowchart illustrating a data augmentation method for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention is shown.

[0022] Figure 2 The illustration shows a schematic diagram of a data augmentation method for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention.

[0023] Figure 3 The diagram illustrates an image segmentation flowchart before classifying the image dataset to be detected, according to an embodiment of the present invention.

[0024] Figure 4 The flowchart illustrating the first data augmentation of a first image dataset according to an embodiment of the present invention is shown schematically.

[0025] Figure 5 The flowchart illustrating a second data augmentation of a first image dataset according to an embodiment of the present invention is shown schematically.

[0026] Figure 6 The diagram illustrates the structure of a data augmentation apparatus for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention.

[0027] Figure 7 A block diagram schematically illustrates an electronic device suitable for implementing a data augmentation method for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0030] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0031] As the resolution of remote sensing images continues to increase, visible light remote sensing images are characterized by large image size and high resolution. However, networks trained on these images often experience numerous false positives and false negatives when performing target detection, resulting in high false positive and false negative rates in practical applications. Therefore, this invention aims to propose a data augmentation method for target detection in large-format, sparse-sample remote sensing images. This method uses complete target data as the extension object for data augmentation, ensuring that more positive samples are added while simultaneously inputting more negative samples into the network training. This increases sample diversity, improves background recognition, reduces false positive rates, and improves detection accuracy.

[0032] Figure 1a A flowchart illustrating a data augmentation method for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention is shown.

[0033] Figure 1b Another flowchart illustrating a data augmentation method for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention is shown.

[0034] Figure 2 The illustration shows a schematic diagram of a data augmentation method for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention.

[0035] like Figure 1a , Figure 1b as well as Figure 2 As shown, the data augmentation method for target detection in large-format sparse sample remote sensing images according to this embodiment may include operations S110 to S190.

[0036] In operation S110, the image dataset to be detected is obtained, and the image dataset to be detected is classified to obtain a first image dataset A containing the target to be detected and a second image dataset B without the target to be detected.

[0037] The aforementioned dataset of images to be detected may, for example, include multiple remote sensing images of ships taken by satellites under different imaging conditions, sea states, and open ocean surfaces. For instance, the dataset may include the following formats or requirements: 25 training images, approximately 300 ship samples, 23 test images, a ground resolution of 5 meters, image sizes of approximately 20000×20000 pixels, images in PNG format, and annotation files in TXT format, with ships annotated using rotated bounding boxes.

[0038] Figure 3 The diagram illustrates an image segmentation flowchart before classifying the image dataset to be detected, according to an embodiment of the present invention.

[0039] like Figure 3 As shown in the embodiment of the present invention, before classifying the image dataset to be detected in the above-mentioned operation S110, it may also include operations S1101 to S1103.

[0040] In operation S1101, a remote sensing image is acquired, which has a preset image frame.

[0041] In operation S1102, the remote sensing image is segmented according to the preset image size, preset overlapping area and preset edge processing method to obtain multiple segmented images.

[0042] For example, the preset size can be 1024×1024, the preset overlap area can be 200, and the preset edge processing method includes filling the image edges with 0 for missing pixels.

[0043] In operation S1103, the segmented images containing the target to be tested in multiple segmented images are labeled to obtain a dataset of images to be detected with labeled information.

[0044] In operation S120, the first image dataset A is used to train the network, resulting in the first detection network N1.

[0045] Specifically, the RoiTransformer can be used to train the network on the dataset of images of the target object to be tested.

[0046] In operation S130, the first detection network N1 is used to perform verification inference on the image dataset B without a target to be tested, and a first false detection image dataset C is obtained, wherein the first false detection image dataset C contains the target to be tested.

[0047] In operation S140, the target to be tested in the first image dataset A is cropped, and the cropped target image is randomly pasted into the false detection images in the first false detection image dataset C to obtain the second false detection image dataset C1.

[0048] Figure 4 The flowchart illustrating the first data augmentation of a first image dataset according to an embodiment of the present invention is shown schematically.

[0049] like Figure 4 As shown, the operation S140 above, which involves cropping the target image in the first image dataset A and randomly pasting the cropped target image into the false detection images in the first false detection image dataset C, may include steps S1401 to S1402.

[0050] In operation S1401, the annotation files in the first image dataset A are traversed, and the coordinates of the rotated boxes in the annotations are converted into the coordinates of the outer horizontal box.

[0051] In operation S1402, the images in the first image dataset A are cropped out using horizontal frames and pasted into the randomly selected false detection images in the first false detection image dataset to form the second false detection image dataset C2.

[0052] In operation S150, the first image dataset A and the second false detection image dataset C1 are input into the first detection network for training, resulting in the second detection network N2.

[0053] In operation S160, the second detection network N2 is used to perform verification inference on the image dataset to be detected, resulting in a third false detection image dataset C3, which contains the target to be tested.

[0054] In operation S170, the target image to be tested in the third false detection image dataset C3 is cropped, and the cropped target image is randomly pasted into the image in the first image dataset A to obtain the enhanced target image dataset A1.

[0055] Figure 5 The flowchart illustrating a second data augmentation of a first image dataset according to an embodiment of the present invention is shown schematically.

[0056] like Figure 5 As shown, the operation S170 above, which involves cropping the target image in the first image dataset A and randomly pasting the cropped target image into the false detection images in the first false detection image dataset C, may include steps S1701 to S1702.

[0057] In operation S1701, the horizontal bounding box of the target to be tested and the pixels within a preset range around the horizontal bounding box in the third false detection image dataset C3 are cropped together.

[0058] In operation S1702, the cropped target image is randomly pasted into the images of the first image dataset A, so that the cropped target image is distributed as evenly as possible in the first image dataset A, resulting in the enhanced target image dataset A1. If the target is occluded by more than a preset occlusion percentage when pasting, the pasting is canceled.

[0059] For example, the preset occlusion percentage can be 10%.

[0060] In operation S180, the target image dataset A, the enhanced target image dataset A1, and the first false detection image dataset C1 are combined to form the final image dataset. The final image dataset R is then input into the second detection network N2 for training to obtain the final target detection network N.

[0061] In operation S190, the remote sensing image to be detected is input into the target detection network N, and the detection result of the target is output.

[0062] Through the above embodiments, the present invention uses the target object to be detected as an extension object for data augmentation, increasing positive samples while simultaneously inputting more negative samples into the network training, thus increasing sample diversity. Furthermore, the present invention improves the recognition of background in remote sensing images, reduces the false detection rate of target detection, and improves the accuracy of target detection.

[0063] To verify the effectiveness of the method proposed in this invention, experiments are conducted to further illustrate its effectiveness:

[0064] This method uses a ship remote sensing image dataset as its foundation. Before using the data augmentation method proposed in this invention, there were only 172 effective training images (1024×1024 in size). After using the data augmentation method proposed in this invention, the number of effective training images increased to 690 (1024×1024 in size), which is a 4-fold increase in effective information compared to before.

[0065] Evaluation Metric: The detection accuracy evaluation metric used in this simulation experiment is the F1 score. The higher the score, the better the detection performance. The F1 score is defined as follows: F1 = 2PR / (P+R), where: R (Recall) = Number of correctly detected targets / (Number of correctly detected targets + Number of missed targets), P (Precision) = Number of correctly detected targets / (Number of correctly detected targets + Number of falsely detected targets).

[0066] Ships are detected using the original algorithm model and the target detection network algorithm model in this invention, and are marked in the form of rotated bounding boxes. In the index calculation, the intersection-over-union (IoU) ratio between the predicted target box and the ground truth bounding box is first calculated. If the IoU value is ≥0.5, the prediction is correct (if there are n predicted boxes that intersect with the target box, and there are detection boxes with an IoU value greater than the threshold, then it is considered that 1 target of that type is correctly predicted, and (n-1) are incorrectly predicted; if there are no detection boxes with an IoU value greater than the threshold, then there are n incorrectly predicted). Then the precision P and recall R of ship detection are calculated to obtain the F1 score.

[0067] Experimental Results: Table 1 compares the experimental results of the augmented dataset with the original dataset. The first column represents the effect of the unaugmented dataset, and the second column represents the effect of the augmented dataset. As can be seen, the data augmentation significantly improved both recall and precision, with precision increasing by 46.68% and recall by 15.63%. The overall score F1 = 78.89, an improvement of 38.99 points.

[0068] Table 1

[0069]

[0070] Based on the methods disclosed above, the present invention also provides a data augmentation method and apparatus for target detection in large-format sparse sample remote sensing images, which will be described below in conjunction with... Figure 6 The device is described in detail.

[0071] Figure 6 The diagram illustrates the structure of a data augmentation apparatus for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention.

[0072] like Figure 6 As shown, the data augmentation method apparatus 600 for target detection in large-format sparse sample remote sensing images according to this embodiment includes an image preprocessing module 610, a first network training module 620, a first verification and inference module 630, a first data augmentation module 640, a second network training module 650, a second verification and inference module 660, a second data augmentation module 670, a third network training module 680, and a target detection module 690.

[0073] The image preprocessing module is used to obtain the image dataset to be detected, classify the image dataset to be detected, and obtain a first image dataset containing the target to be detected and a second image dataset without the target to be detected.

[0074] The first network training module is used to train the network using the first image dataset to obtain the first detection network.

[0075] The first verification and inference module is used to perform verification and inference on the image dataset without a target using the first detection network to obtain the first false detection image dataset, wherein the first false detection image dataset contains the target.

[0076] The first data augmentation module is used to crop the target to be tested in the first image dataset and randomly paste the cropped target image into the false detection images in the first false detection image dataset to obtain the second false detection image dataset.

[0077] The second network training module inputs the first image dataset and the second false detection image dataset into the first detection network for training, thereby obtaining the second detection network.

[0078] The second verification and reasoning module is used to perform verification and reasoning on the image dataset to be detected using the second detection network to obtain the third false detection image dataset, wherein the third false detection image dataset contains the target to be tested;

[0079] The second data augmentation module is used to crop the target image in the third false detection image dataset, and randomly paste the cropped target image into the image in the first image dataset to obtain the augmented target image dataset.

[0080] The third network training module is used to combine the target image dataset to be tested, the enhanced target image dataset to be tested, and the first false detection image dataset to form the final image dataset. The final image dataset is then input into the second detection network for training to obtain the final target detection network.

[0081] The target detection module is used to input the remote sensing image to be detected into the target detection network and output the detection results of the target.

[0082] It should be noted that the embodiments of the device section are similar to those of the method section, and the technical effects achieved are also similar. For specific details, please refer to the above-mentioned method embodiment section, which will not be repeated here.

[0083] According to embodiments of the present invention, any plurality of the image preprocessing module 610, the first network training module 620, the first verification and inference module 630, the first data augmentation module 640, the second network training module 650, the second verification and inference module 660, the second data augmentation module 670, the third network training module 680, and the object detection module 690 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of the present invention, at least one of the image preprocessing module 610, the first network training module 620, the first verification inference module 630, the first data augmentation module 640, the second network training module 650, the second verification inference module 660, the second data augmentation module 670, the third network training module 680, and the target detection module 690 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or any other reasonable means of integrating or packaging circuits, or implemented in hardware or firmware, or in any one of the three implementation methods of software, hardware, and firmware, or in a suitable combination of any of them. Alternatively, at least one of the image preprocessing module 610, the first network training module 620, the first verification and inference module 630, the first data augmentation module 640, the second network training module 650, the second verification and inference module 660, the second data augmentation module 670, the third network training module 680, and the object detection module 690 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0084] Figure 7 A block diagram schematically illustrates an electronic device suitable for implementing a data augmentation method for target detection in large-format sparse sample remote sensing images according to an embodiment of the present invention.

[0085] like Figure 7As shown, an electronic device 700 according to an embodiment of the present invention includes a processor 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage portion 707 into a random access memory (RAM) 703. The processor 701 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 701 may also include onboard memory for caching purposes. The processor 701 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0086] RAM 703 stores various programs and data required for the operation of electronic device 700. Processor 701, ROM 702, and RAM 703 are interconnected via bus 704. Processor 701 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 702 and / or RAM 703. It should be noted that the programs may also be stored in one or more memories other than ROM 702 and RAM 703. Processor 701 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in said one or more memories.

[0087] According to an embodiment of the present invention, the electronic device 700 may further include an input / output (I / O) interface 707, and an input / output (I / O) interface 705 is also connected to the bus 704. The electronic device 700 may also include one or more of the following components connected to the I / O interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 707 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the I / O interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 710 as needed so that computer programs read from it can be installed into the storage section 707 as needed.

[0088] The accompanying drawings show some block diagrams and / or flowcharts. It should be understood that some blocks or combinations thereof in the block diagrams and / or flowcharts can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when executed by the processor, these instructions can create means for implementing the functions / operations described in these block diagrams and / or flowcharts.

[0089] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements.

[0090] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are 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 data enhancement method for large-format sparse sample remote sensing image target detection, characterized in that, The method comprises the following steps: S110, obtaining a to-be-detected picture data set, classifying the to-be-detected picture data set to obtain a first picture data set containing a to-be-detected target and a second picture data set without the to-be-detected target; S120, training a network using the first picture data set to obtain a first detection network; S130, verifying and reasoning the first detection network using the second picture data set without the to-be-detected target to obtain a first false detection picture data set, wherein the first false detection picture data set contains the to-be-detected target; S140, cropping the to-be-detected target in the first picture data set, and randomly pasting the cropped to-be-detected target picture into a false detection picture of the first false detection picture data set to obtain a second false detection picture data set; S150, inputting the first picture data set and the second false detection picture data set into the first detection network for training to obtain a second detection network; S160, verifying and reasoning the second detection network using the to-be-detected picture data set to obtain a third false detection picture data set, wherein the third false detection picture data set contains the to-be-detected target; S170, cropping the to-be-detected target in the third false detection picture data set, and randomly pasting the cropped to-be-detected target picture into a picture of the first picture data set to obtain an enhanced to-be-detected target picture data set; S180, inputting the to-be-detected target picture data set, the enhanced to-be-detected target picture data set, and the first false detection picture data set into the second detection network for training to obtain a final target detection network; S190, inputting a to-be-detected remote sensing image into the target detection network to output a detection result of the to-be-detected target.

2. The data enhancement method for target detection of a large-format sparse sample remote sensing image according to claim 1, wherein the to-be-detected picture data set comprises a plurality of ship remote sensing images captured by a satellite under different imaging conditions, different sea conditions, and open ocean conditions. Before the classification of the to-be-detected picture data set, the method further comprises:

3. The data enhancement method for large-format sparse sample remote sensing image target detection according to claim 1, characterized in that, obtaining a remote sensing image, wherein the remote sensing image has a preset format; segmenting the remote sensing image into a plurality of segmented images according to a preset size, a preset overlapping area, and a preset edge processing mode; labeling segmented images containing a to-be-detected target in the plurality of segmented images to obtain the to-be-detected picture data set with labeling information. The preset size is 1024x1024, the preset overlapping area is 200, and the preset edge processing mode comprises filling a picture edge lacking pixels with 0.

4. The data enhancement method for large-format sparse sample remote sensing image target detection according to claim 3, characterized in that, The network training of the to-be-detected target picture data set to obtain the first detection network comprises:

5. The data enhancement method for large-format sparse sample remote sensing image target detection according to claim 1, characterized in that, training the network of the to-be-detected target picture data set using a rotating target detector. The cropping of the to-be-detected target in the to-be-detected target picture data set and the random pasting of the cropped to-be-detected target picture into the picture of the false detection picture data set to obtain the second false detection picture data set comprises:

6. The data enhancement method for large-format sparse sample remote sensing image target detection according to claim 1, characterized in that, ​ Traverse the annotation information of the to-be-tested target picture data set, convert the coordinates of the rotated rectangular frame in the annotation information into the coordinates of the circumscribed horizontal frame, cut the to-be-tested target picture of the to-be-tested target picture data set in the form of a horizontal frame, and randomly select a mis-detection picture in the first mis-detection picture data set to perform pasting, to form a second mis-detection picture data set.

7. The data enhancement method for large-format sparse sample remote sensing image target detection according to claim 1, characterized in that, The to-be-tested target in the third mis-detection picture data set is cut, and the cut to-be-tested target picture is randomly pasted into the picture of the first picture data set to obtain an enhanced to-be-tested target picture data set, including: The to-be-tested target in the third mis-detection picture data set is cut, and the cut to-be-tested target picture is randomly pasted into the picture of the first picture data set to obtain an enhanced to-be-tested target picture data set, including: The to-be-tested target in the third mis-detection picture data set is cut, and the cut to-be-tested target picture is randomly pasted into the picture of the first picture data set to obtain an enhanced to-be-tested target picture data set, including:

8. A data enhancement device for large-format sparse sample remote sensing image target detection, characterized in that, A picture preprocessing module is configured to obtain a to-be-detected picture data set, classify the to-be-detected picture data set, and obtain a first picture data set containing a to-be-tested target and a second picture data set containing no to-be-tested target; A first network training module is configured to use the first picture data set to perform network training, and obtain a first detection network; A first verification and reasoning module is configured to use the first detection network to perform verification and reasoning on the no to-be-tested target picture data set, and obtain a first mis-detection picture data set containing a to-be-tested target; A first data enhancement module is configured to cut the to-be-tested target in the first picture data set, and randomly paste the cut to-be-tested target picture into the mis-detection picture of the first mis-detection picture data set, to obtain a second mis-detection picture data set; A second network training module is configured to input the first picture data set and the second mis-detection picture data set into the first detection network for training, to obtain a second detection network; A second verification and reasoning module is configured to use the second detection network to perform verification and reasoning on the to-be-detected picture data set, to obtain a third mis-detection picture data set containing a to-be-tested target; A second data enhancement module is configured to cut the to-be-tested target in the third mis-detection picture data set, and randomly paste the cut to-be-tested target picture into the picture of the first picture data set, to obtain an enhanced to-be-tested target picture data set; A third network training module is configured to group a to-be-tested target picture data set, an enhanced to-be-tested target picture data set, and a first mis-detection picture data set into a final picture data set, input the final picture data set into the second detection network for training, and obtain a final target detection network; A target detection module is configured to input a to-be-detected remote sensing image into the target detection network, and output a detection result of a to-be-tested target. One or more processors; 9. An electronic device, comprising: a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the method according to any one of claims 1-7. One or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the method according to any one of claims 1-7.

Images