A remote sensing image target detection method and system based on an improved YOLOV8 algorithm

Abstract

The present invention discloses a remote sensing image target detection method and system based on an improved YOLOv8 algorithm, which relates to the technical fields of deep learning and computer vision. The method comprises the following steps: receiving a remote sensing target data set, preprocessing the remote sensing target data set to obtain a remote sensing target processing data set, dividing the remote sensing target processing data set into a training set, a verification set and a test set, wherein the remote sensing target data set is a sampled image obtained by frame sampling of a remote sensing video; inputting the training set and the verification set into a pre-established YOLOv8 original model for training and verification and improvement to obtain an improved YOLOv8 model, wherein the YOLOv8 original model requires parameter setting during training; and inputting the test set into the improved YOLOv8 model for remote sensing image target detection to obtain a remote sensing image target detection result.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and computer vision technology, and specifically to a remote sensing image target detection method and system based on an improved YOLOV8 algorithm. Background Art

[0002] Object detection technology is a key research area in computer vision and has garnered widespread attention in recent years. It has been applied in a variety of fields, including autonomous driving, security surveillance, drones, and medical imaging. In these areas, object detection can automatically identify and track objects in images or videos, bringing convenience to people's daily lives and production.

[0003] Current object detection algorithms for remote sensing imagery still have some drawbacks. These include: remote sensing images contain many small objects, such as vehicles and equipment, for which current mainstream algorithms have high rates of missed and false detections; objects in remote sensing images are multidirectional and have large aspect ratios, requiring models to adapt to object rotation and specific sizes; and objects in remote sensing images can appear different due to variations in lighting conditions and viewing angles. Current object detection algorithms have limited adaptability to changes in lighting and viewing angles, making them prone to false or missed detections.

[0004] To address the shortcomings of object detection algorithms in remote sensing images, Liu Tao et al. proposed an improved YOLOv5 algorithm, introducing a channel attention mechanism to enhance feature capture and fusion capabilities, adding a fine-grained detection layer that incorporates shallow semantic information to improve detection of small objects, and using copy-paste for data augmentation. Zhang Zheng et al. proposed an improved SFP-DETR algorithm, combining dilated convolution with a single-level feature pyramid structure. They constructed a new bounding box regression loss to achieve anchor-free rotation object detection and added a weight constraint to the decoder's cross-attention to restrict global attention calculations to a local scope. Yuan Yuman et al. proposed an improved FCOS model, which fuses and optimizes multi-scale features by scaling them and extracting contextual information from high-level features using an appropriate receptive field. These algorithms have improved the performance of various object detectors to some extent, but detection accuracy still needs improvement. Summary of the Invention

[0005] In order to solve the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide a remote sensing image target detection method and system based on an improved YOLOV8 algorithm, which can solve the problems of low accuracy and recall rate of remote sensing targets.

[0006] The purpose of the present invention can be achieved by the following technical solution: a remote sensing image target detection method based on an improved YOLOV8 algorithm, the method comprising the following steps:

[0007] Receiving a remote sensing target data set, preprocessing the remote sensing target data set to obtain a remote sensing target processed data set, and dividing the remote sensing target processed data set into a training set, a validation set, and a test set, wherein the remote sensing target data set is a sampled image obtained by frame sampling of a remote sensing video;

[0008] Input the training set and the validation set into a pre-established YOLOv8 original model for training, validation, and improvement to obtain an improved YOLOv8 model, wherein the YOLOv8 original model requires parameter setting during training;

[0009] The test set is input into the improved YOLOv8 model for remote sensing image target detection to obtain the remote sensing image target detection results.

[0010] Preferably, the preprocessing of the remote sensing target data set is to use the Mosaic data enhancement algorithm to randomly scale, randomly crop and randomly arrange the sampled images.

[0011] Preferably, the improvement of the YOLOv8 original model includes: using an improved bounding box regression loss function with a dynamic focusing mechanism, using a residual attention network that combines a residual structure with global attention, and using deformable convolution and deformable RoI pooling to change the model structure of the YOLOv8 original model.

[0012] Preferably, the improved bounding box regression loss function with a dynamic focusing mechanism adopts a non-monotonic focusing mechanism, constructs a dynamic gradient gain coefficient, and uses the outlier degree instead of the intersection-over-union ratio to evaluate the quality of the anchor box, thereby obtaining a bounding box loss function that combines the non-monotonic focusing mechanism with geometric factors. The formula is as follows:

[0013] L WIoU =γR WIoU L IoU ,

[0014]

[0015] Among them, L WIoU is the bounding box regression loss function of WIoU, L IoU is the initial bounding box regression loss function, IoU is the intersection-over-union ratio between the real box and the bounding box, x, y are the coordinates of the center point of the predicted box, x gt ,y gt is the center point coordinate of the predicted box of the real box, W g and H g is the width and height of the minimum rectangular box formed by the predicted box and the true box, w, h are the width and height of the predicted box respectively. δ, α are hyper parameters, γ is the gradient gain, β is the outlier degree, is the dynamic average intersection-combination ratio with momentum m. To prevent the generation of gradients that hinder convergence, the L in β is IoU 、R WIoU W in g and H g Separate from the gradient calculation (superscript * indicates this operation).

[0016] Preferably, in the residual attention network that combines the residual structure with global attention, the process of the global attention mechanism is expressed as:

[0017]

[0018] y=w1K1 T +b1,

[0019]

[0020] Among them, F1 is the input feature map, F2 is the output feature map of the channel attention submodule, sigmoid and ReLu are activation functions, w1, w2 and b1, b2 are the initial weights and bias terms of the multi-layer perceptron respectively, y is the output of the multi-layer perceptron, K1, K2 are the eigenvalues ​​of the image, T represents transposition, ConvBN is the normalization layer, M C is the channel attention function, F3 is the global attention output feature map, M S is the spatial attention function;

[0021] After introducing the residual structure, the resulting residual attention network is as follows:

[0022]

[0023] Where F4 is the output feature map of residual attention.

[0024] Preferably, the position of each pixel point of the convolution kernel in the deformable convolution is determined by an offset, so that the sampling grid is no longer restricted to a regular distribution and can adapt to objects of different shapes. The deformable convolution expands the regular sampling grid R by adding an offset. For any position p, the output feature map Y can be expressed as

[0025]

[0026] Among them, p represents any position of the input feature map, K represents the number of convolution kernels, and w k and p k Represent the weight and offset of the k-th position, Δp k and Δm k denote the learnable offset and modulation scalar at the kth position, Δm kThe weights of the convolution sampling points on the input feature map are calculated to remove irrelevant context information, and the weights of irrelevant sampling points are directly learned to be 0;

[0027] Due to the introduction of the offset, the 9 sampling positions are no longer regular, and the offset Δp n It is usually a fraction, so it is implemented using bilinear interpolation, which can be expressed as

[0028]

[0029] G(q,p)=max(0,1-|q x -p x |)·max(0,1-|q y -p y |),

[0030] Among them, p represents any fractional position, q enumerates all integral spatial positions in the feature map X, and p x 、p y ,q x ,q y are the horizontal and vertical coordinates of p and q respectively, G(·,·) is the bilinear interpolation kernel, and MAX(·,·) is the maximum value function.

[0031] Preferably, the deformable RoI pooling introduces an offset to enhance the expressiveness of feature representation. RoI pooling defines a rectangular box on the feature map after the last convolution as the region of interest RoI. For a given input feature map, the RoI pooling layer divides the RoI into k×k histograms and outputs k×k output feature maps. The (i, j)th histogram is defined as bin(i, j) (0≤i, j≤k), and the corresponding output feature map y(i, j) is

[0032]

[0033] Where x(·) is the input feature map, p0 is the coordinate of the upper left corner of interest, p is the pixel in the enumeration histogram bin (i, j), and n ij is the number of pixels in histogram bin (i, j);

[0034] Add offset Δp to RoI pooling in deformable RoI pooling ij , generating a new pixel position, the corresponding output feature map y(i,j) is expressed as follows:

[0035]

[0036] Preferably, the evaluation indicators of the improved YOLOv8 model performance include: precision P, recall R, average precision AP and detection speed FPS, wherein the precision P is defined as the correct detection rate of all detected targets, which can be expressed as

[0037]

[0038] Recall rate R—defined as the detection accuracy rate among all positive samples, which can be expressed as

[0039]

[0040] Average precision AP—defined as the mean of the precision under different recall rates, which can be expressed as

[0041]

[0042] Among them, TP is the number of positive samples predicted as positive; FP is the number of negative samples predicted as positive; FN is the number of positive samples predicted as negative; TN is the number of negative samples predicted as negative;

[0043] FPS is defined as the number of images detected per second by the improved YOLOv8 model.

[0044] In a second aspect, in order to achieve the above-mentioned purpose, the present invention discloses a remote sensing image target detection system based on an improved YOLOV8 algorithm, comprising:

[0045] Data preprocessing module: used to receive the remote sensing target data set, preprocess the remote sensing target data set to obtain the remote sensing target processing data set, and divide the remote sensing target processing data set into a training set, a validation set, and a test set, wherein the remote sensing target data set is a sampled image obtained by frame sampling of remote sensing video;

[0046] Model improvement module: used to input the training set and validation set into the pre-established YOLOv8 original model for training, validation and improvement to obtain an improved YOLOv8 model, wherein the YOLOv8 original model requires parameter setting during training;

[0047] Target detection module: used to input the test set into the improved YOLOv8 model to perform remote sensing image target detection and obtain remote sensing image target detection results.

[0048] In another aspect of the present invention, in order to achieve the above object, a device is disclosed, comprising:

[0049] one or more processors;

[0050] a memory for storing one or more programs;

[0051] When one or more of the programs are executed by one or more of the processors, the one or more processors implement the remote sensing image target detection method of the improved YOLOV8 algorithm as described above.

[0052] Beneficial effects of the present invention:

[0053] The present invention can accurately and efficiently detect remote sensing image targets and improve the accuracy of detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0055] Figure 1 It is a schematic flow chart of the method of the present invention;

[0056] Figure 2 It is a workflow diagram of the present invention;

[0057] Figure 3 1 is a schematic diagram of a model of an improved YOLOv8 algorithm according to an embodiment of the present invention;

[0058] Figure 4 is a global attention diagram according to an embodiment of the present invention;

[0059] Figure 5 Schematic diagram of residual attention according to an embodiment of the present invention;

[0060] Figure 6 Schematic diagram of deformable convolution principle according to an embodiment of the present invention;

[0061] Figure 7 Schematic diagram of deformable RoI pooling according to an embodiment of the present invention;

[0062] Figure 8 This is a schematic diagram of visual comparison of target detection results of remote sensing images using four methods according to an embodiment of the present invention;

[0063] Figure 9 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0065] like Figure 1 As shown, a remote sensing image target detection method based on an improved YOLOV8 algorithm includes the following steps:

[0066] Receiving a remote sensing target data set, preprocessing the remote sensing target data set to obtain a remote sensing target processed data set, and dividing the remote sensing target processed data set into a training set, a validation set, and a test set, wherein the remote sensing target data set is a sampled image obtained by frame sampling of a remote sensing video;

[0067] Input the training set and the validation set into a pre-established YOLOv8 original model for training, validation, and improvement to obtain an improved YOLOv8 model, wherein the YOLOv8 original model requires parameter setting during training;

[0068] The test set is input into the improved YOLOv8 model for remote sensing image target detection to obtain the remote sensing image target detection results.

[0069] Training phase:

[0070] The DOTA dataset is used for preprocessing remote sensing target datasets. The DOTA dataset is a large-scale image dataset for object detection in aerial images, which can be used to discover and evaluate objects in aerial images. The DOTA dataset contains 2,806 images, 188,282 instances, and 15 categories. The dataset is split into 21,046 images with a resolution of 1024×1024, and then divided into training and test sets with a ratio of 7:3.

[0071] To build a remote sensing target detection network model, this paper proposes an improved YOLOv8 algorithm, as shown in the reference Figure 3 shown.

[0072] The non-monotonic focusing mechanism combined with geometric factors can be expressed as the bounding box loss function:

[0073] L WIoU =γR WIoU L IoU ,

[0074]

[0075] Among them, x, y are the coordinates of the center point of the prediction box, xgt ,y gt is the center point coordinate of the predicted box of the real box, W g and H g Is the width and height of the minimum rectangular box formed by the predicted box and the real box, w, h are the width and height of the predicted box respectively. δ, α are hyperparameters (δ is 3 and α is 2 during training), γ is the gradient gain, and β is the outlier degree. is the dynamic average intersection-combination ratio with momentum m.

[0076] Given that training data inevitably contains low-quality examples, blindly strengthening bounding box regression on low-quality samples will reduce the model's generalization performance. When the pre-selected box has a high degree of overlap with the target box, a good loss function should weaken the penalty for geometric factors, enabling the model to achieve better generalization. WIoU adopts a non-monotonic focusing mechanism. By constructing a dynamic gradient gain coefficient and using outlier scores instead of intersection over union to evaluate the quality of anchor boxes, it provides a sensible gradient gain allocation strategy. A low outlier score indicates high-quality pre-selected boxes, so only a lower gradient gain needs to be allocated, focusing the loss regression on pre-selected boxes of average quality.

[0077] In remote sensing image detection, due to the wide variety of object shapes and sizes, the aspect ratios of bounding boxes often vary significantly. If the model only focuses on regressing the center coordinates while ignoring the influence of the bounding box aspect ratio, it may make it difficult to accurately detect slender or compressed objects. Therefore, to further achieve a reasonable gradient distribution, while minimizing the influence of geometric factors on gradient distribution, an aspect ratio penalty term is added, allowing the model to accurately regress the aspect ratio while paying appropriate attention to it.

[0078] As reference Figure 4 As shown in the figure, the global attention mechanism inherits the structure of the multi-head attention mechanism and improves the two sub-modules of channel and space. It uses the attention weights between the three dimensions of channel, spatial width and spatial height to improve efficiency and amplify cross-dimensional interactions.

[0079] The process of the global attention mechanism can be expressed as

[0080]

[0081] y=w1K1 T +b1,

[0082]

[0083] Among them, F1 is the input feature map, F2 is the output feature map of the channel attention submodule, sigmoid and ReLu are activation functions, w1, w2 and b1, b2 are the initial weights and bias terms of the multi-layer perceptron respectively, y is the output of the multi-layer perceptron, K1, K2 are the eigenvalues ​​of the image, T represents transposition, ConvBN is the normalization layer, M C is the channel attention function, F3 is the global attention output feature map, M S is the spatial attention function.

[0084] Given an input feature map F1 with a dimensional order of C×W×H, the channel attention module first transforms the feature map's dimensional order to preserve channel information, resulting in a dimensional order of W×H×C. Then, through two layers of multi-layer perceptrons, nonlinearity is introduced, enabling the model to learn adaptive weights and nonlinear transformations to capture complex relationships between features, enhance spatial dependencies between dimensional channels, and employ activation functions between layers to prevent gradient vanishing and explosion. Finally, the three-dimensional order is restored to C×W×H and multiplied with the original input feature map F1 to produce the channel attention output feature map F2. After receiving the feature map F2, the spatial attention module extracts and fuses spatial semantic information using two convolutions with kernels of size 7×7 and padding of 3, enhancing the model's focus on spatial information. To further preserve spatial information, pooling is eliminated and the activation function is directly applied to the attention output feature map F2 before multiplication to produce the final global attention output feature map F3.

[0085] Considering that as the network depth increases, the training error also increases, resulting in network degradation and other problems, the residual structure is integrated into the global attention mechanism and the residual attention module is designed to alleviate this problem. Figure 5 As shown in Figure 2, the new output feature map is obtained by fusing the input feature map with the channel attention output feature map and the global attention input feature map. The introduction of the residual structure enables the model to extract information at different levels, enhances the feature fusion capability, alleviates the network degradation problem, makes it easier for the model to learn the mapping relationship between input and output, and prevents the problem of gradient disappearance and diffusion. The process of the residual attention module is expressed as

[0086]

[0087] Where F4 is the output feature map of residual attention.

[0088] Remote sensing images have complex backgrounds, and the detected targets occupy a small area of ​​the image, resulting in a high proportion of redundant features during the model iteration process. The introduction of the residual attention mechanism reduces the weight of background information and enhances channels containing large amounts of feature information. This allows the model to fit important semantic information and focus more closely on the target area. It also prevents gradient explosion and diffusion, preserves the effective information in the original input feature map, and allows the model to capture more complex features.

[0089] In traditional convolution operations, the convolution kernel has a fixed pixel position, and the same convolution kernel is applied to every position in the input image. However, in reality, images at different positions may have different deformations. Conventional convolution cannot adapt to the irregular layout and non-rigid deformation of remote sensing images, which may lead to inaccurate feature extraction of remote sensing targets.

[0090] Taking 3*3 convolution as an example, in deformable convolution, the position of each pixel of the convolution kernel is determined by an offset, so that the sampling grid is no longer restricted to a regular distribution and can adapt to objects of different shapes. Deformable convolution expands the regular sampling grid R by adding an offset. In order to make the offset of the sampling point accurately cover the detection target, a modulation variable is introduced to enhance the ability of deformable convolution to manipulate space, allowing the model to learn the offset of the sampling point and the weight of each sampling point, reducing the interference of irrelevant factors. Then, for any position p, the output feature map Y can be expressed as

[0091]

[0092] Among them, K represents the number of convolution kernels, w k and p k Represent the weight and offset of the k-th position, Δp k and Δm k Denote the learnable offset and modulation scalar of the kth position respectively. Δm k The weights of the convolution sampling points on the input feature map are calculated to remove irrelevant contextual information, and the weights of irrelevant sampling points are directly learned to be 0.

[0093] Due to the introduction of the offset, the 9 sampling positions are no longer regular, and the offset Δp n It is usually a fraction, so it is implemented using bilinear interpolation, which can be expressed as

[0094]

[0095] G(q,p)=g(q x ,p x )·g(q y ,p y ),

[0096] g(q y ,py )=max(0,1-|ab|),

[0097] where p denotes an arbitrary (fractional) position, q enumerates all integral spatial positions in the feature map X, and G(·,·) is the bilinear interpolation kernel.

[0098] As reference Figure 6 As shown in the figure, the deformable convolution first preprocesses the input feature map to generate offsets and modulations. Regularly distributed pixels are then offset in different directions to obtain irregularly distributed pixels. The input image is then sampled based on the generated new pixels to obtain the sampled feature map. Finally, the sampled feature map is element-wise multiplied by the convolution kernel and summed to obtain the final convolution result.

[0099] The pooling layer is a hierarchical structure commonly used in convolutional neural networks. It is used to reduce the dimension of the input data, thereby reducing model parameters and computational complexity, while also being able to extract key features from the input data. In traditional pooling methods, the feature values ​​in a local area are usually statistically summarized to obtain the feature representation of the area. Deformable RoI pooling introduces an offset to enhance the expressiveness of feature representation. RoI pooling defines a rectangular box on the feature map after the last convolution as the region of interest (RoI). For a given input feature map, the RoI pooling layer divides the RoI into k×k histograms and outputs k×k output feature maps. Define the (i, j)th histogram as bin(i, j) (0≤i, j≤k), then the corresponding output feature map y(i, j) is

[0100]

[0101] Where x(·) is the input feature map, p0 is the coordinate of the upper left corner of interest, p is the pixel in the enumeration histogram bin (i, j), and n ij is the number of pixels in histogram bin (i, j).

[0102] Similar to deformable convolution, in deformable RoI pooling an offset Δp is added to the RoI pooling. ij , generating a new pixel position, the corresponding output feature map y(i,j) is expressed as follows:

[0103]

[0104] As reference Figure 7As shown, first, for the input feature map, deformable RoI pooling generates the corresponding region of interest. Secondly, on the feature map, for the position of each region of interest, an additional deformable convolution layer is used to predict the offset of each position in the region of interest, which will be used to adjust the sampling position. The coordinates of the region of interest and the predicted offset are then used to divide the region of interest into a series of sub-regions. For each sub-region, the sampling position is adjusted according to the offset, and the feature value of the corresponding position is obtained from the feature map. Then, for each sub-region, an adaptive maximum pooling operation is performed on the feature value obtained by sampling it to obtain an output of a fixed size. Finally, the pooling results of all sub-regions are spliced ​​together in a certain order to form the final deformable RoI pooling result.

[0105] The evaluation indicators of the improved YOLOv8 model performance include: precision P, recall R, average precision AP and detection speed FPS, where precision P is defined as the correct detection rate of all detected targets, which can be expressed as

[0106]

[0107] Recall rate R—defined as the detection accuracy rate among all positive samples, which can be expressed as

[0108]

[0109] Average precision AP—defined as the mean of the precision under different recall rates, which can be expressed as

[0110]

[0111] Among them, TP is the number of positive samples predicted as positive; FP is the number of negative samples predicted as positive; FN is the number of positive samples predicted as negative; TN is the number of negative samples predicted as negative;

[0112] In practical object detection applications, network models must not only be highly accurate but also perform real-time detection. The network's detection speed is measured in FPS, which is defined as the number of images detected per second by the modified YOLOv8 model.

[0113] To verify the performance of each module, a series of ablation experiments were designed using the DOTA dataset with the YOLOv8 original model as the benchmark method for comparative verification. Precision, Recall, and mAP@0.5 were used as quantitative evaluation indicators. The experimental results are shown in Table 1.

[0114] Table 1 Ablation experiment

[0115]

[0116]

[0117] From the analysis of Table 1, it can be seen that after the network uses the non-monotonic focused bounding box regression loss function, the average accuracy of each type is not significantly improved. After adding the geometric factor, the average accuracy of each type is improved by 0.4%, the precision is improved by 0.3%, and the recall rate and FPS remain basically unchanged; after the global attention mechanism is introduced into the model, the average accuracy of each type is improved by 0.5%. After the residual global attention mechanism is introduced, the average accuracy of each type is improved by 0.8%, the recall rate is improved by 0.6%, and the FPS decreases slightly; after the C2f module is changed to the C2f-DCN module, the FPS decreases significantly. The reason for this is that the introduction of the After deformed convolution, the convolution complexity of the C2f_DCN module is much higher than that of the C2f module using ordinary convolution, and the computational complexity increases significantly, which has a significant impact on the network detection speed. However, the average accuracy of all categories increases by 1.2%, the precision increases by 0.6%, and the recall rate increases by 0.4%. After the network incorporates the bounding box regression loss function that combines a non-monotonic focusing mechanism with geometric factors, the residual global attention mechanism, and the C2f_DCN module, the FPS decreases, but the accuracy increases by 0.7%, the recall rate increases by 0.7%, and the average accuracy of all categories increases by 2.3%, with a significant overall performance improvement.

[0118] To verify the detection performance of the improved YOLOv8 algorithm model, FPS and mAP@0.5 are used as quantitative indicators to quantitatively analyze the target detection results of the YOLOv8 algorithm, YOLOv5 algorithm and YOLOv3 algorithm on the DOTA dataset. Different models are used to detect the verification part of the dataset, and the comparison of the result indicators is shown in Table 2.

[0119] Table 2 Comparison of performance of different network models

[0120] Table2 Performance comparison of different network models

[0121] Analysis of Table 2 shows that the improved YOLOv8 achieves significant improvements in both PFS and mean average precision compared to the Faster R-CNN and SSD algorithms. This is attributed to the fact that the R-CNN algorithm, unlike other algorithms, does not perform multi-scale fusion of high-level features in its feature extraction network, resulting in poor robustness for detecting objects of varying scales. The SSD algorithm uses multiple bounding boxes to detect objects of varying sizes, but this can lead to poor detection performance when objects have inconsistent aspect ratios. Compared to YOLOv3, YOLOv5, and YOLOv8, the improved YOLOv8 achieves improvements of 7.1%, 4.2%, and 1.9% in mean average precision, respectively. This is attributed to the relatively old Darknet architecture used in YOLOv3, which is less scalable and flexible than the latest deep learning architectures. Because YOLOv5 uses an anchor-free approach during training, further optimization is needed for small object detection. YOLOv8 utilizes a more gradient-rich architecture, significantly improving performance and achieving an FPS of 98, a significant advantage over other models. However, it lacks the ability to model and generalize deformable objects. The improved YOLOv8 algorithm model achieves an average precision of 85.3% across all categories, outperforming other algorithms. However, its FPS is significantly lower than YOLOv8, attributed to the introduction of a more complex convolutional approach and backbone architecture. Overall, however, its performance is superior to the other compared algorithms.

[0122] In addition, the improved YOLOv8 algorithm will be used in DOTA with YOLOv3, YOLOv5, and YOLOv8 algorithms.

[0123]

[0124] The detection effects of some images in the dataset are compared intuitively, such as reference Figure 8As shown in the figure, (a) to (c) are the original images; (d) to (f) are the detection results of YOLOv3; (g) to (i) are the detection results of YOLOv5; (j) to (i) are the detection results of YOLOv8; and (m) to (o) are the detection results of improved YOLOv8. As shown in Figures (d) to (f), YOLOv3 has a high miss detection rate for small and densely packed objects. This is because YOLOv3's detection network only predicts objects based on feature maps of three different scales, which makes it unable to accurately process image semantic information. As shown in Figures (g) to (i), YOLOv5's confidence score has increased significantly, but its miss detection rate remains high. This is because although YOLOv5 uses a multi-scale detection head and an adaptive training strategy, it has certain limitations when dealing with objects in extreme situations (such as extremely small or extremely large objects). As shown in Figures (j) to (l), the YOLOv8 algorithm significantly reduces its missed detection rate for small and densely packed targets. This is due to the fact that YOLOv8 replaces the C3 module with a C2f module, which has a richer gradient flow, and uses a decoupling head instead of a coupling head, significantly improving its small target detection capabilities. As shown in Figures (m) to (o), the improved YOLOv8 algorithm has improved its overall confidence compared to other algorithms and significantly reduced its missed detection rate, especially for small and irregular targets. Its overall performance is superior to other detection algorithms.

[0125] On the other hand, Figure 9 As shown, the embodiment of the present invention further discloses a remote sensing image target detection system based on an improved YOLOV8 algorithm, including:

[0126] Data preprocessing module: used to receive the remote sensing target data set, preprocess the remote sensing target data set to obtain the remote sensing target processing data set, and divide the remote sensing target processing data set into a training set, a validation set, and a test set, wherein the remote sensing target data set is a sampled image obtained by frame sampling of remote sensing video;

[0127] Model improvement module: used to input the training set and validation set into the pre-established YOLOv8 original model for training, validation and improvement to obtain an improved YOLOv8 model, wherein the YOLOv8 original model requires parameter setting during training;

[0128] Target detection module: used to input the test set into the improved YOLOv8 model to perform remote sensing image target detection and obtain remote sensing image target detection results.

[0129] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0130] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium having a computer program stored thereon, which executes the above method when executed by a processor. The storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component.

[0131] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0132] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present disclosure. Various changes and improvements may be made to the present disclosure without departing from the spirit and scope of the present disclosure, and such changes and improvements shall fall within the scope of the present disclosure.

Claims

1. A remote sensing image target detection method based on an improved YOLOV8 algorithm, characterized in that: The method comprises the following steps: Receiving a remote sensing target data set, preprocessing the remote sensing target data set to obtain a remote sensing target processed data set, and dividing the remote sensing target processed data set into a training set, a validation set, and a test set, wherein the remote sensing target data set is a sampled image obtained by frame sampling of a remote sensing video; Input the training set and the validation set into a pre-established YOLOv8 original model for training, validation, and improvement to obtain an improved YOLOv8 model, wherein the YOLOv8 original model requires parameter setting during training; The test set is input into the improved YOLOv8 model to perform remote sensing image target detection and obtain the remote sensing image target detection results; The improved YOLOv8 model includes: The residual attention network combines the residual structure with global attention. The process of the global attention mechanism is expressed as: <h2 style=";text-align:left;direction:ltr">[K1·ReLU(w2y+b2)<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> ], y=w1K1 T +b1, [ConvBN(ConvBNRELU(K2))], Among them, F1 is the input feature map, F2 is the output feature map of the channel attention submodule, sigmoid and ReLu are activation functions, w1, w2 and b1, b2 are the initial weights and bias terms of the multi-layer perceptron respectively, y is the output of the multi-layer perceptron, K1, K2 are the eigenvalues ​​of the image, T represents transposition, ConvBN is the normalization layer, M C is the channel attention function, F3 is the global attention output feature map, M S is the spatial attention function; After introducing the residual structure, the resulting residual attention network is as follows: Among them, F4 is the output feature map of residual attention; Deformable convolution, the position of each pixel of the convolution kernel in the deformable convolution is determined by an offset, so that the sampling grid is no longer restricted to a regular distribution and can adapt to objects of different shapes. The deformable convolution adds an offset to the regular sampling grid R to expand it. For any position p, the output feature map Y can be expressed as Among them, p represents any position of the input feature map, K represents the number of convolution kernels, and w k and p k Represent the weight and offset of the k-th position, Δp k and Δm k denote the learnable offset and modulation scalar at the kth position, Δm k The weights of the convolution sampling points on the input feature map are calculated to remove irrelevant context information, and the weights of irrelevant sampling points are directly learned to be 0; Due to the introduction of the offset, the 9 sampling positions are no longer regular, and the offset Δp n It is usually a fraction, so it is implemented using bilinear interpolation, which can be expressed as G(q,p)=max(0,1-|q x -p x |)·max(0,1-|q y -p y |), Among them, p represents any fractional position, q enumerates all integral spatial positions in the feature map X, and p x 、p y ,q x ,q y are the horizontal and vertical coordinates of p and q respectively, G(·,·) is the bilinear interpolation kernel, and MAX(·,·) is the maximum value function; Deformable RoI pooling, Deformable RoI pooling introduces an offset to enhance the expressiveness of feature representation. RoI pooling defines a rectangular box on the feature map after the last convolution as the region of interest RoI. For a given input feature map, the RoI pooling layer divides the RoI into k×k histograms and outputs k×k output feature maps. The (i, j)th histogram is defined as bin(i, j) (0≤i, j≤k), and the corresponding output feature map y(i, j) is Where x(·) is the input feature map, p0 is the coordinate of the upper left corner of interest, p is the pixel in the enumeration histogram bin (i, j), and n ij is the number of pixels in histogram bin (i, j); Add offset Δp to RoI pooling in deformable RoI pooling ij , generating a new pixel position, the corresponding output feature map y(i,j) is expressed as follows: It also includes the use of an improved bounding box regression loss function with a dynamic focusing mechanism.

2. The remote sensing image target detection method based on the improved YOLOV8 algorithm according to claim 1 is characterized in that: The preprocessing of the remote sensing target data set is to use the Mosaic data enhancement algorithm to randomly scale, randomly crop and randomly arrange the sampled images.

3. The remote sensing image target detection method based on the improved YOLOV8 algorithm according to claim 1 is characterized in that: The improved bounding box regression loss function with dynamic focusing mechanism adopts a non-monotonic focusing mechanism. By constructing a dynamic gradient gain coefficient and using outlier degree instead of intersection-over-union ratio to evaluate the quality of the anchor box, a bounding box loss function combining the non-monotonic focusing mechanism with geometric factors is obtained. The formula is as follows: IT WIoU =γR WIoU IT IoU , L IoU =1-IoU Among them, L WIoU is the bounding box regression loss function of WIoU, L IoU is the initial bounding box regression loss function, IoU is the intersection-over-union ratio between the real box and the bounding box, x, y are the coordinates of the center point of the predicted box, x gt ,y gt is the center point coordinate of the predicted box of the real box, W g and H g It is the width and height of the minimum rectangular box formed by the predicted box and the real box, w, h are the width and height of the predicted box respectively, δ, α are hyper parameters, γ is the gradient gain, β is the outlier degree, is the dynamic average intersection-combination ratio with momentum m. The superscript * indicates that the L in β is replaced by IoU 、R WIoU W in g and H g Separate from gradient calculation.

4. The remote sensing image target detection method based on the improved YOLOV8 algorithm according to claim 1, characterized in that: The evaluation indicators of the improved YOLOv8 model performance include: precision P, recall R, average precision AP and detection speed FPS, where precision P is defined as the correct detection rate of all detected targets, which can be expressed as Recall rate R—defined as the detection accuracy rate among all positive samples, which can be expressed as Average precision AP—defined as the mean of the precision under different recall rates, which can be expressed as Among them, TP is the number of positive samples predicted as positive; FP is the number of negative samples predicted as positive; FN is the number of positive samples predicted as negative; TN is the number of negative samples predicted as negative; FPS is defined as the number of images detected per second by the improved YOLOv8 model.

5. A remote sensing image target detection system using an improved YOLO V8 algorithm, which adopts the remote sensing image target detection method using an improved YOLO V8 algorithm according to any one of claims 1 to 4, characterized in that: include: Data preprocessing module: used to receive the remote sensing target data set, preprocess the remote sensing target data set to obtain the remote sensing target processing data set, and divide the remote sensing target processing data set into a training set, a validation set, and a test set, wherein the remote sensing target data set is a sampled image obtained by frame sampling of remote sensing video; Model improvement module: used to input the training set and validation set into the pre-established YOLOv8 original model for training, validation and improvement to obtain an improved YOLOv8 model, wherein the YOLOv8 original model requires parameter setting during training; Target detection module: used to input the test set into the improved YOLOv8 model to perform remote sensing image target detection and obtain remote sensing image target detection results.

6. A device, characterized in that include: one or more processors; a memory for storing one or more programs; When one or more of the programs are executed by one or more of the processors, the one or more processors implement a remote sensing image target detection method based on an improved YOLOV8 algorithm as described in any one of claims 1 to 4.

Images