Remote sensing image small target identification method based on improved YOLOv8 algorithm

Abstract

The invention discloses a remote sensing image small target identification method based on improved YOLOv8, and the method comprises the steps: obtaining a remote sensing image data set, carrying out the preprocessing, and generating a preprocessed data set; an improved YOLOv8 model is constructed, the improved YOLOv8 model comprises a backbone network, a neck network and a head network, the head network comprises a small target detection layer, the improved YOLOv8 model adopts an EIoU loss function to measure the difference between a real frame and a prediction frame, a regression weighted non-maximum suppression method is adopted to perform weighted averaging on the position of the prediction frame, and the position of the prediction frame is corrected; training an improved YOLOv8 model based on the preprocessed data set to obtain a trained model; and performing small target identification detection on the remote sensing image based on the trained model, and outputting a detection result. The invention further discloses a remote sensing image small target recognition device based on the improved YOLOv8, corresponding equipment and a storage medium. According to the invention, missing detection and false detection of small targets in remote sensing images can be effectively reduced, and the detection performance is improved.

Description

Technical Field

[0001] The present application relates to the field of target detection technology, and more specifically, to a method for recognizing small targets in remote sensing images based on an improved YOLOv8 algorithm. Background Art

[0002] With the rapid development of remote sensing technology, the resolution and data volume of remote sensing images have significantly increased, enabling their widespread application in a variety of fields, including security and environmental protection. However, the detection and identification of small targets in remote sensing images has always been a technical challenge in this field. Small targets are typically small in size, contain limited information, and are easily interfered with in complex backgrounds, making traditional target detection algorithms difficult to meet the needs of practical applications. For example, in some fields, small targets in remote sensing images may include enemy equipment and facilities, and their accurate identification is crucial for intelligence analysis and decision support. In the security field, small target detection can be used in scenarios such as border monitoring and maritime surveillance to help promptly detect potential threats. Therefore, how to efficiently and accurately detect and identify small targets in remote sensing imagery has become a hot research topic.

[0003] In recent years, the development of deep learning technology has brought new advancements to small target detection in remote sensing imagery. Deep learning-based object detection algorithms, such as the YOLO series, have been widely used in various fields due to their high efficiency and accuracy. However, these algorithms still have shortcomings when dealing with small targets, such as difficulties in feature extraction and strong background interference. Existing technologies still need to be further optimized in terms of the accuracy and efficiency of small target detection, especially in applications with complex remote sensing imagery. Therefore, developing an efficient, accurate, and robust method for small target detection in remote sensing imagery has important practical significance and broad application prospects. Summary of the Invention

[0004] In response to at least one defect or improvement need in the prior art, the present invention provides a method, apparatus, device and storage medium for identifying small targets in remote sensing images based on an improved YOLOv8 algorithm, which can solve at least one of the problems existing in the above-mentioned background technology.

[0005] To achieve the above objectives, according to a first aspect of the present invention, a method for identifying small targets in remote sensing images based on an improved YOLOv8 algorithm is provided, the method comprising:

[0006] Acquiring a remote sensing image data set, preprocessing the remote sensing image data to generate a preprocessed data set;

[0007] An improved YOLOv8 model is constructed, comprising a backbone network, a neck network, and a head network. The head network includes a small object detection layer. The improved YOLOv8 model uses the EIoU loss function to measure the difference between the true bounding box and the predicted bounding box, and uses a regression-weighted non-maximum suppression method to perform weighted averaging on the predicted bounding box positions to correct the predicted bounding box positions.

[0008] Train the improved YOLOv8 model based on the preprocessed dataset to obtain the trained model;

[0009] Based on the trained model, small target recognition and detection are performed on remote sensing images, and the detection results are output.

[0010] Furthermore, in the remote sensing image small target recognition method based on improved YOLOv8, the head network includes a small target detection layer, specifically including:

[0011] A 4x down-sampling detection head is added to the YOLOv8 model, and the minimum resolution of the 4x down-sampling detection head is 4×4 pixels.

[0012] Furthermore, the above-mentioned remote sensing image small target recognition method based on the improved YOLOv8 model uses the EIoU loss function to measure the difference between the true box and the predicted box, specifically including:

[0013] For each predicted box and true box, calculate their intersection-over-union ratio, which is the ratio of the overlapping area of ​​the predicted box and the true box to their union area;

[0014] Calculate the distance between the center point of the predicted box and the real box, and calculate the distance difference between the predicted box and the real box in width and height respectively;

[0015] The distance between the center point of the predicted box and the real box and the distance difference between the predicted box and the real box in width and height are weighted and combined with the intersection over union ratio to obtain the EIoU loss function.

[0016] Furthermore, the above-mentioned remote sensing image small target recognition method based on improved YOLOv8 calculates the distance difference between the predicted box and the real box in width and height respectively, specifically including:

[0017] The upper left corner coordinate point position of the predicted box and the real box are obtained respectively, and the lower right corner coordinate point position of the predicted box and the real box are obtained respectively. Based on the coordinate positions of the corresponding points of the predicted box and the real box, the distance difference in width and height between the predicted box and the real box is obtained.

[0018] Furthermore, the above-mentioned remote sensing image small target recognition method based on the improved YOLOv8 uses the regression weighted non-maximum suppression method to perform weighted averaging on the predicted box position and correct the predicted box position, specifically including:

[0019] Sort all prediction boxes by confidence from high to low, select the prediction box with the highest confidence as the current optimal box, and retain it in the final result;

[0020] Calculate the intersection-over-union (IoU) ratio between the current optimal box and the remaining prediction boxes, and filter out the prediction box clusters whose IoU ratio is higher than the set IoU threshold;

[0021] Using the confidence of the prediction box as the weight, the weighted average of the prediction box positions in the filtered prediction box cluster is performed to correct the position of the current optimal box.

[0022] Furthermore, in the above-mentioned remote sensing image small target recognition method based on improved YOLOv8, the preprocessing of the remote sensing image data specifically includes:

[0023] Normalize the pixel values ​​of remote sensing images to the range of [0,1];

[0024] Use data augmentation techniques to stitch multiple images together;

[0025] Convert the original annotation information into the TXT format required by YOLOv8.

[0026] Furthermore, the above-mentioned remote sensing image small target recognition method based on improved YOLOv8 uses precision, recall, F1, and multi-category average precision as evaluation criteria to assess model accuracy during the training process.

[0027] According to a second aspect of the present invention, a remote sensing image small target recognition device based on improved YOLOv8 is also provided, which includes:

[0028] A preprocessing module is used to obtain a remote sensing image data set, preprocess the remote sensing image data, and generate a preprocessed data set;

[0029] A model construction module is used to build an improved YOLOv8 model, including a backbone network, a neck network, and a head network. The head network includes a small object detection layer. The improved YOLOv8 model uses the EIoU loss function to measure the difference between the true box and the predicted box, and uses the regression weighted non-maximum suppression method to perform weighted averaging on the predicted box position to correct the predicted box position.

[0030] The model training module is used to train the improved YOLOv8 model based on the preprocessed dataset and obtain the trained model;

[0031] The detection module is used to identify and detect small targets in remote sensing images based on the trained model and output the detection results.

[0032] According to a third aspect of the present invention, a remote sensing image small target recognition device based on improved YOLOv8 is also provided, which includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit performs the steps of any one of the above methods.

[0033] According to a fourth aspect of the present invention, a storage medium is also provided, which stores a computer program that can be executed by a remote sensing image small target recognition device based on an improved YOLOv8. When the computer program runs on the remote sensing image small target recognition device based on an improved YOLOv8, the remote sensing image small target recognition device based on an improved YOLOv8 performs the steps of any of the above methods.

[0034] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0035] The remote sensing image small target recognition method based on improved YOLOv8 provided by the present invention can more effectively capture the feature information of small targets through data preprocessing, model improvement, optimized loss function and weighted non-maximum suppression method, thereby improving the accuracy of small target detection, significantly improving the detection performance of small targets in remote sensing images, and effectively reducing missed detection and false detection phenomena, providing reliable technical support for remote sensing image analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 A flowchart of a method for identifying small targets in remote sensing images based on an improved YOLOv8 is provided in an embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0039] The terms "first," "second," "third," and the like in the specification and claims of this application and the accompanying drawings are used to distinguish between different objects, not to describe a particular order. Furthermore, the terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements, but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0040] Figure 1 A flowchart of a method for identifying small targets in remote sensing images based on improved YOLOv8 is provided in the embodiment of the present application. Figure 1 As shown, the embodiment of the present application provides a remote sensing image small target recognition method based on improved YOLOv8, including the following steps:

[0041] S1 obtains a remote sensing image data set, preprocesses the remote sensing image data, and generates a preprocessed data set;

[0042] S2 builds an improved YOLOv8 model, including a backbone network, a neck network, and a head network, wherein the head network includes a small object detection layer. The improved YOLOv8 model uses the EIoU loss function to measure the difference between the true box and the predicted box, and uses the regression weighted non-maximum suppression method to perform weighted averaging on the predicted box position to correct the predicted box position;

[0043] S3 trains the improved YOLOv8 model based on the preprocessed dataset to obtain the trained model;

[0044] S4 performs small target recognition and detection on remote sensing images based on the trained model and outputs the detection results.

[0045] Specifically, the network structure of the existing YOLOv8 remote sensing image recognition algorithm is mainly composed of a backbone (Backbone), a neck (Neck) and a head (Head).

[0046] The Backbone module uses the CSPDarknet-53 architecture, which consists of 53 convolutional layers, including some residual blocks. These residual blocks help the network better capture image details and context. The C2f module is used to fuse feature maps of different scales, thereby extracting rich feature information and maintaining a rich gradient flow.

[0047] For the Neck module, YOLOv8 removes the post-upsampling convolutions based on the PAN architecture, achieving lightweight performance without sacrificing performance. Traditional FPN uses a top-down approach to transmit deep semantic information, which can result in the loss of some target positioning information. PAN-FPN adds PAN to FPN, enhancing path information by complementing shallow and deep information.

[0048] The YOLOv8 detection head module uses a decoupled head structure, employing two independent branches for object classification and prediction regression. The head module then uses a 1x1 convolutional layer to complete classification and localization tasks. YOLOv8 also uses anchor-free detection, which enables faster model convergence and improved accuracy.

[0049] The improved YOLOv8 remote sensing image small target recognition method provided in the embodiment of the present application first obtains a remote sensing image dataset containing small targets from platforms such as Google Earth. The dataset contains multiple pictures, covering a variety of small targets of different categories. The dataset is divided into a training set, a validation set, and a test set according to a certain ratio. For example, in one embodiment, the dataset contains 852 pictures, covering 4 different categories of small targets, and the dataset is divided into a training set (689 pictures), a validation set (77 pictures), and a test set (86 pictures) according to a ratio of 8:1:1. The remote sensing image is normalized, and the pixel values ​​are normalized to the range of [0,1] to reduce the impact of numerical range differences on model training. In addition, the Mosaic data enhancement technology is used to splice multiple pictures together to enhance the model's adaptability to different scenes. At the same time, the image is randomly cropped, flipped, rotated, and other operations are performed to further enrich the training samples. Finally, the annotation information in the original XML format is converted into the TXT format required by YOLOv8 to generate a preprocessed dataset.

[0050] An improved YOLOv8 model was constructed. Its backbone network adopts the CSPDarknet-53 architecture, consisting of 53 convolutional layers and residual blocks for image feature extraction. The neck network uses a PAN-FPN structure, fusing feature maps of different scales through upsampling and downsampling to enhance the model's detection capabilities for multi-scale objects. The head network adds a detection head with 4x downsampling, enabling it to detect small objects as small as 4×4 pixels. It also adopts a decoupled head structure, with two independent branches for object classification and prediction regression. The improved YOLOv8 model uses the EIoU loss function to measure the difference between the ground-truth and predicted bounding boxes. By dynamically adjusting weights, the model focuses more on small objects during training, accelerating network convergence and improving small object detection accuracy. Furthermore, a regression-weighted non-maximum suppression method is used to perform weighted averaging of predicted box positions to correct them. By calculating the IoU matrix between predicted boxes, clusters of predicted boxes near each optimal box are selected. The predicted box scores are used as weights to perform a weighted averaging of the predicted box coordinates, thereby improving the localization accuracy of the predicted boxes.

[0051] The improved YOLOv8 model was iteratively trained using the preprocessed training and validation sets. Hyperparameters such as anchor box size, optimizer, learning rate, and number of iterations were set. During training, model performance was optimized using the EIoU loss function and regression-weighted non-maximum suppression to generate the optimal model weight file. Model performance was evaluated using the test set, with metrics such as precision, recall, F1 score, and multi-class average precision output to verify model effectiveness.

[0052] Use the trained model to detect small objects in remote sensing imagery. Preprocess the input remote sensing imagery and convert it to the model's required input format. Use the improved YOLOv8 model for object detection, outputting the detection bounding box, category, and confidence score. Visualize the detection results by drawing the detection bounding box, category label, and confidence score.

[0053] This application provides a remote sensing image small target recognition method based on improved YOLOv8. Through data preprocessing, model improvement, optimized loss function and weighted non-maximum suppression method, it can more effectively capture the feature information of small targets, improve the accuracy of small target detection, and significantly improve the detection performance of small targets in remote sensing images. It can effectively reduce missed detection and false detection phenomena, and provide reliable technical support for remote sensing image analysis.

[0054] Optionally, in the remote sensing image small target recognition method based on improved YOLOv8 provided in an embodiment of the present application, the head network includes a small target detection layer, specifically including:

[0055] A 4x down-sampling detection head is added to the YOLOv8 model, and the minimum resolution of the 4x down-sampling detection head is 4×4 pixels.

[0056] Specifically, in YOLOv8, the feature map used for target detection can perceive the minimum range of the original image as 8×8, which makes the model easily miss small targets with a width and height of less than 8 pixels. To improve this problem, this application adds a 4x downsampling detection head based on the YOLOv8 model, which can be named YOLOv8s-P2. The minimum target resolution that YOLOv8s-P2 can detect is 4×4. It is more powerful than the original network and can predict targets of different scales on 4x, 8x, 16x, and 32x downsampled feature maps, greatly improving the algorithm's multi-scale target detection performance.

[0057] Since the improved model adds a small-scale detection head, the feature fusion method on the Neck side has also changed accordingly, but the overall structure still follows the FPN+PAN structure.

[0058] The Neck network first performs upsampling fusion on the input feature map, then downsampling fusion, and then predicts the fused feature map separately. The specific process is as follows:

[0059] Upsampling and fusion. To achieve fine-grained feature detection, the 20*20*512 feature map input from the Backbone network is first upsampled and fused with a 40*40*512 feature map to obtain a 40*40*512 feature map in the Neck network. The 40*40*512 feature map is then upsampled and fused with an 80*80*256 feature map to obtain a finer-grained 80*80*768 feature map. The 80*80*768 feature map is further upsampled to a 160*160*256 feature map, which is then fused with a 160*160*128 feature map. Small object detection is performed on this fused large-scale feature map, and the 160*160*128 detection layer outputs the detection results.

[0060] Downsampling and fusion. The 160*160*128 detection layer feature map is first downsampled and fused with the 80*80*256 feature map at the same scale. The 80*80*256 detection layer then outputs the detection results for small objects. The network continues downsampling and is then fused with 40*40*512 and 40*40*256 feature maps, respectively. Medium and large objects are detected on the resulting 16x and 32x downsampled feature maps. Finally, the 40*40*512 and 20*20*1024 detection layers output the prediction results.

[0061] Optionally, the embodiment of the present application provides a remote sensing image small target recognition method based on the improved YOLOv8, wherein the improved YOLOv8 model uses the EIoU loss function to measure the difference between the real box and the predicted box, specifically including:

[0062] For each predicted box and true box, calculate their intersection-over-union ratio, which is the ratio of the overlapping area of ​​the predicted box and the true box to their union area;

[0063] Calculate the distance between the center point of the predicted box and the real box, and calculate the distance difference between the predicted box and the real box in width and height respectively;

[0064] The distance between the center point of the predicted box and the real box and the distance difference between the predicted box and the real box in width and height are weighted and combined with the intersection over union ratio to obtain the EIoU loss function.

[0065] Optionally, the remote sensing image small target recognition method based on improved YOLOv8 provided in the embodiment of the present application calculates the distance difference between the predicted box and the real box in width and height respectively, specifically including:

[0066] The upper left corner coordinate point position of the predicted box and the real box are obtained respectively, and the lower right corner coordinate point position of the predicted box and the real box are obtained respectively. Based on the coordinate positions of the corresponding points of the predicted box and the real box, the distance difference in width and height between the predicted box and the real box is obtained.

[0067] Specifically, to address the defects of the CIoU loss function in existing remote sensing image target recognition methods in dealing with issues such as border size differences and sample imbalance, the EIoU loss function and weight coefficient are introduced for optimization and adjustment.

[0068] Since the CIoU loss function cannot distinguish bounding boxes with the same center and the same aspect ratio but different sizes, the EIoU loss function is introduced to modify the original aspect ratio and regress the width and height values ​​into the upper left corner position information and the lower right corner position information based on the EIoU loss function.

[0069] The calculation formula of CIoU loss function is as follows:

[0070]

[0071] Where IoU is the intersection-over-union ratio between the predicted box and the real box, (b,b gt) is the center point of the predicted box and the ground-truth box, ρ is the Euclidean distance between them, c is the minimum diagonal length of the bounding box between the predicted box and the ground-truth box, α is a parameter used to balance the ratio, and v is a parameter used to describe the consistency of the aspect ratio between the predicted box and the ground-truth box. The CIoU loss function is relatively robust, can better cope with objects of different shapes and sizes, and can also reduce the risk of overfitting.

[0072] The calculation formula of the EIoU loss function is as follows:

[0073]

[0074] Where IoU is the intersection-over-union ratio between the predicted box and the real box, (b,b gt ) is the center point of the predicted box and the real box, ρ is the Euclidean distance between the two, and c is the minimum diagonal length of the external box between the predicted box and the real box; (w,w gt ) is the width of the predicted box and the real box, c w is the normalized parameter of the width difference between the predicted box and the real box; (h,h gt ) is the height of the predicted box and the real box, c h is the normalized parameter for the height difference between the predicted box and the true box.

[0075] The calculation formula of the improved loss function is as follows:

[0076]

[0077] in, Where IoU is the intersection-over-union ratio between the predicted box and the real box, (b,b gt ) is the center point of the predicted box and the real box, ρ is the Euclidean distance between the two, and c is the minimum diagonal length of the external box between the predicted box and the real box; (xmin, xmin gt ) and (ymin,ymin gt ) is the coordinate point of the upper left corner of the predicted box and the real box, c wmin and c hmin is the normalized parameter of the width and height of the upper left corner of the predicted box and the real box; (xmax, xmax gt ) and (ymax,ymax gt ) is the coordinate point of the lower right corner of the predicted box and the real box, c hmax and c wmax W is the normalized parameter of the width and height of the lower right corner of the predicted box and the real box. c 、W min 、W maxIt is the normalized weight of the center point, upper left corner and lower right corner of the predicted box and the real box.

[0078] Optionally, the remote sensing image small target recognition method based on improved YOLOv8 provided in an embodiment of the present application uses the regression weighted non-maximum suppression method to perform weighted averaging on the predicted box position to correct the predicted box position, specifically including:

[0079] Sort all prediction boxes by confidence from high to low, select the prediction box with the highest confidence as the current optimal box, and retain it in the final result;

[0080] Calculate the intersection-over-union (IoU) ratio between the current optimal box and the remaining prediction boxes, and filter out the prediction box clusters whose IoU ratio is higher than the set IoU threshold;

[0081] Using the confidence of the prediction box as the weight, the weighted average of the prediction box positions in the filtered prediction box cluster is performed to correct the position of the current optimal box.

[0082] Specifically, the traditional non-maximum suppression method does not require redundant calculations. Although its screening method of directly deleting all suboptimal boxes is more efficient, it cannot deal with the problem of position offset of the optimal box. The regression weighted non-maximum suppression method utilizes the information of all predicted boxes and can use the information of the remaining boxes for fine-tuning to improve the prediction box positioning accuracy of the model. First, the non-maximum suppression method is used to filter out the predicted box with the highest score near each target as the optimal box, and the subscript set i of the optimal box is returned. Assume that there are M optimal boxes in the set i; then, by calculating the IoU of the two boxes, the prediction box cluster near each optimal box is filtered out. The specific algorithm implementation is to calculate the IoU of each optimal box with all the predicted boxes respectively, and filter out the prediction box cluster with an intersection over union ratio higher than the set intersection over union ratio threshold. If the IoU is greater than the threshold, it is set to True, and if it is less than the threshold, it is set to False, and finally an M*N IoU matrix is ​​obtained; finally, the IoU matrix is ​​multiplied by the scores of all predicted boxes as weights, and the weights are weighted averaged with the position coordinates to obtain the final output prediction box. The set intersection over union ratio threshold can be set according to actual conditions, and this application does not make specific restrictions here.

[0083] Optionally, in the remote sensing image small target recognition method based on improved YOLOv8 provided in an embodiment of the present application, the preprocessing of the remote sensing image data specifically includes:

[0084] Normalize the pixel values ​​of remote sensing images to the range of [0,1];

[0085] Use data augmentation techniques to stitch multiple images together;

[0086] Convert the original annotation information into the TXT format required by YOLOv8.

[0087] Specifically, remote sensing image data preprocessing includes but is not limited to the following steps:

[0088] Normalize the pixel values ​​of remote sensing images to the range [0, 1]. This is done by dividing each pixel value by the maximum possible pixel value, ensuring that all pixel values ​​are in the same numerical range.

[0089] Mosaic data augmentation technology is used to stitch multiple images together. Mosaic augmentation randomly selects four images and stitches them together into a single large image, adjusting the coordinates of the annotation boxes to accommodate the new image size. This augmentation method can simulate more complex background environments, enhance the model's adaptability to different scenarios, and improve its generalization performance.

[0090] Convert the original annotation information from XML to the TXT format required by YOLOv8. The annotation information typically includes the coordinates of the target box (such as the center point coordinates, width, and height) and the category label. The converted TXT annotation file can be directly read and parsed by the YOLOv8 model, facilitating the model's use of this annotation information during training.

[0091] Optionally, the remote sensing image small target recognition method based on improved YOLOv8 provided in the embodiment of the present application uses precision, recall, F1, and multi-category average precision as evaluation criteria to assess model accuracy during training.

[0092] Specifically, Mosaic data augmentation is used throughout the training process to enrich training samples and enhance the generalization performance of the model. Model training evaluation criteria can use precision, recall, F1, and multi-class average precision (mAP) as objective evaluation criteria. The calculation formulas for precision and recall are as follows:

[0093]

[0094] TP: The label value is True and the model prediction is Positive;

[0095] FN: The label value is False and the model prediction is Negative;

[0096] FP: The label value is False and the model prediction is Positive;

[0097] TN: The label value is True, and the model prediction is Negative.

[0098] AP is calculated based on the true label and predicted probability of each category. The area formed by the corresponding Precision and Recall values ​​is called Average Precision (AP); mAP is the average AP value of all categories, and its value range is [0,1]. The higher the values ​​of the above two indicators, the better the detection performance of the algorithm.

[0099] The present application also provides a device for identifying small targets in remote sensing images based on an improved YOLOv8, including:

[0100] A preprocessing module is used to obtain a remote sensing image data set, preprocess the remote sensing image data, and generate a preprocessed data set;

[0101] A model construction module is used to build an improved YOLOv8 model, including a backbone network, a neck network, and a head network. The head network includes a small object detection layer. The improved YOLOv8 model uses the EIoU loss function to measure the difference between the true box and the predicted box, and uses the regression weighted non-maximum suppression method to perform weighted averaging on the predicted box position to correct the predicted box position.

[0102] The model training module is used to train the improved YOLOv8 model based on the preprocessed dataset and obtain the trained model;

[0103] The detection module is used to identify and detect small targets in remote sensing images based on the trained model and output the detection results.

[0104] The present application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above method. The computer-readable storage medium may include, but is not limited to, any type of disk, including a floppy disk, an optical disk, a DVD, a CD-ROM, a microdrive, a magneto-optical disk, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory device, a magnetic card or an optical card, a nanosystem (including a molecular memory IC), or any type of medium or device suitable for storing instructions and / or data.

[0105] It should be noted that for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0106] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0107] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of the device or unit can be electrical or other forms.

[0108] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0109] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0110] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a memory, including a number of instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned memory includes: various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0111] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable memory, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0112] The above is only an exemplary embodiment of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the specification and practicing the disclosure herein, those skilled in the art will easily think of the implementation scheme of the present disclosure. This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art that are not recorded in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

[0113] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0114] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for recognizing small targets in remote sensing images based on improved YOLOv8, characterized in that: The following steps are involved: Acquiring a remote sensing image data set, preprocessing the remote sensing image data to generate a preprocessed data set; An improved YOLOv8 model is constructed, comprising a backbone network, a neck network, and a head network. The head network includes a small object detection layer. The improved YOLOv8 model uses the EIoU loss function to measure the difference between the true bounding box and the predicted bounding box, and uses a regression-weighted non-maximum suppression method to perform weighted averaging on the predicted bounding box positions to correct the predicted bounding box positions. Train the improved YOLOv8 model based on the preprocessed dataset to obtain the trained model; Based on the trained model, small target recognition and detection are performed on remote sensing images, and the detection results are output.

2. The remote sensing image small target recognition method based on improved YOLOv8 according to claim 1, characterized in that: The head network includes a small target detection layer, specifically including: A 4x down-sampling detection head is added to the YOLOv8 model, and the minimum resolution of the 4x down-sampling detection head is 4×4 pixels.

3. The remote sensing image small target recognition method based on improved YOLOv8 according to claim 1, characterized in that: The improved YOLOv8 model uses the EIoU loss function to measure the difference between the true box and the predicted box, specifically including: For each predicted box and true box, calculate their intersection-over-union ratio, which is the ratio of the overlapping area of ​​the predicted box and the true box to their union area; Calculate the distance between the center point of the predicted box and the real box, and calculate the distance difference between the predicted box and the real box in width and height respectively; The distance between the center point of the predicted box and the real box and the distance difference between the predicted box and the real box in width and height are weighted and combined with the intersection over union ratio to obtain the EIoU loss function.

4. The remote sensing image small target recognition method based on improved YOLOv8 according to claim 3, characterized in that: Calculate the distance difference between the predicted box and the real box in width and height respectively, including: The upper left corner coordinate point position of the predicted box and the real box are obtained respectively, and the lower right corner coordinate point position of the predicted box and the real box are obtained respectively. Based on the coordinate positions of the corresponding points of the predicted box and the real box, the distance difference in width and height between the predicted box and the real box is obtained.

5. The remote sensing image small target recognition method based on improved YOLOv8 according to claim 1, characterized in that: The method of using regression weighted non-maximum suppression to perform weighted averaging on the predicted frame position and correct the predicted frame position specifically includes: Sort all prediction boxes by confidence from high to low, select the prediction box with the highest confidence as the current optimal box, and retain it in the final result; Calculate the intersection-over-union (IoU) ratio between the current optimal box and the remaining prediction boxes, and filter out the prediction box clusters whose IoU ratio is higher than the set IoU threshold; Using the confidence of the prediction box as the weight, the weighted average of the prediction box positions in the filtered prediction box cluster is performed to correct the position of the current optimal box.

6. The remote sensing image small target recognition method based on improved YOLOv8 according to claim 1, characterized in that: The preprocessing of the remote sensing image data specifically includes: Normalize the pixel values ​​of remote sensing images to the range of [0,1]; Use data augmentation techniques to stitch multiple images together; Convert the original annotation information into the TXT format required by YOLOv8.

7. The remote sensing image small target recognition method based on improved YOLOv8 according to claim 1, characterized in that: During the training process, precision, recall, F1, and multi-category average accuracy are used as evaluation criteria to assess model accuracy.

8. A remote sensing image small target recognition device based on improved YOLOv8, characterized in that: include: A preprocessing module is used to obtain a remote sensing image data set, preprocess the remote sensing image data, and generate a preprocessed data set; A model construction module is used to build an improved YOLOv8 model, including a backbone network, a neck network, and a head network. The head network includes a small object detection layer. The improved YOLOv8 model uses the EIoU loss function to measure the difference between the true box and the predicted box, and uses the regression weighted non-maximum suppression method to perform weighted averaging on the predicted box position to correct the predicted box position. The model training module is used to train the improved YOLOv8 model based on the preprocessed dataset and obtain the trained model; The detection module is used to identify and detect small targets in remote sensing images based on the trained model and output the detection results.

9. A remote sensing image small target recognition device based on improved YOLOv8, characterized in that: The method comprises at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit executes the steps of the method according to any one of claims 1 to 7.

10. A storage medium, characterized in that: It stores a computer program that can be executed by a remote sensing image small target recognition device based on improved YOLOv8. When the computer program runs on the remote sensing image small target recognition device based on improved YOLOv8, the remote sensing image small target recognition device based on improved YOLOv8 executes the steps of the method described in any one of claims 1 to 7.

Images