High-definition remote sensing image detection method based on improved YOLOv5

Abstract

The invention belongs to the technical field of remote sensing image detection, and particularly relates to a high-definition remote sensing image detection method based on improved YOLOv5. According to the characteristics that a remote sensing image is small in target and low in resolution, a feature function fusion module in an original YOLOv5 model is improved; comprising the steps that a feeling layer of an SPD module expansion model is introduced, so that the model can better reserve fine-grained features, and the representation learning ability of an image is improved; an attention mechanism is added in a feature fusion module of YOLOV5, and a Cotnet module is used to replace a primary up-sampling module in an original model structure, so that the model can more fully utilize rich context information between adjacent keys, thereby obtaining more optimized feature representation; a traditional loss function in YOLOv5 is replaced, and ablation experiment verification is carried out; the accuracy of the improved model is effectively improved compared with that before improvement. The method can be widely applied to the fields of urban planning, land utilization, environment detection and the like.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image detection, and in particular relates to a high-definition remote sensing image detection method. Background Art

[0002] With the continuous development of remote sensing technology, high-resolution remote sensing images have become one of the important means to obtain information about the earth's surface. High-definition remote sensing images play an important role in various application fields, such as land use planning, urban planning, environmental monitoring, and resource management. Therefore, target detection in images has become more and more critical because it can provide valuable information for many practical problems. High-definition remote sensing image recognition is of great significance.

[0003] The target detection method can accurately detect the required target objects and has great application potential in high-definition remote sensing images. It mainly includes traditional target detection methods and deep learning-based target detection methods. Traditional target detection methods usually use manually designed features and machine learning methods to classify and locate targets. Commonly used methods include Haar feature cascade detectors, gradient direction histograms, and feature-based cascade classifiers (such as AdaBoost and SVM). The most representative methods of deep learning-based target detection methods are two important variants in target detection: region-based methods and single-stage methods. Region-based methods such as R-CNN solve the problem of locating specific image target detection and FastR-CNN are used in large-scale target detection methods; single-stage methods such as YOLO methods are used to achieve lightweight target detection and SSD is used to locate targets in video images.

[0004] Due to the high resolution and complexity of HD remote sensing images, as well as the diversity of targets in different scenes, existing deep learning-based object detection methods face many challenges in processing these images: for example, the scale, shape, and appearance of targets vary greatly, and there are factors such as occlusion, illumination changes, and noise in the images, which can lead to a decrease in the accuracy and robustness of traditional methods. Summary of the invention

[0005] The purpose of the present invention is to provide a high-definition remote sensing image detection method based on improved YOLOv5 to solve the problem of decreased accuracy and robustness of the traditional high-definition remote sensing image detection method proposed in the above background technology.

[0006] The high-definition remote sensing image detection method based on improved YOLOv5 provided by the present invention improves the original YOLOv5 high-definition remote sensing image detection model (referred to as YOLOv5, or YOLOv5 model) according to the characteristics of small targets and low resolution of remote sensing images; includes introducing the SPD module to expand the perception layer of the model; adding an attention mechanism to the feature fusion module of YOLOV5; replacing the traditional loss function in YOLOv5, and performing ablation experiment verification; the specific steps are:

[0007] Step 1: Improve the YOLOv5 model: including optimizing the Backbone structure in the YOLOv5 model to retain the fine-grained features of remote sensing images; introducing the SPD module and expanding the perception layer of the YOLOv5 model so that the YOLOv5 model can better retain fine-grained features and improve the image representation learning ability;

[0008] Step 2: Add attention mechanism to the feature fusion module of YOLOV5: Optimize the Neck structure in the YOLOv5 model to integrate the attention mechanism, use the rich contextual information between adjacent keys, and optimize the feature representation;

[0009] Step 3: Replace the loss function in YOLOv5: Based on the detection accuracy optimization loss function IoU (Intersection over Union) in the target detection method, the loss function is used to quantify the difference between the model's prediction and the actual value;

[0010] Step 4: Ablation experiment verification: Experimental verification is performed from the aspects of average precision, precision, recall rate and accuracy to verify the performance of the target detection model.

[0011] Further:

[0012] In step 1, the Backbone structure in the YOLOv5 model is optimized by embedding a CoTC3 module between the Backbone structure and the Neck structure, where the CoTC3 module is used for global context perception and feature interaction;

[0013] The CotC3 module consists of three convolutional layers, namely the multi-scale feature fusion layer, the cross-channel feature transfer layer, and the spatial perception layer. The module extracts features by dividing the input feature map into two parts, processing them separately and then merging them, using the self-attention mechanism. The CotC3 module includes the CotNet module and the C3 module, which work together. The attention mechanism in the CotNet module is responsible for capturing long-distance feature dependencies, while the C3 module focuses on the fusion of local features, thereby achieving an effective combination of global and local features.

[0014] In the improved YOLOv5 model, the Cot module (i.e., CotC3) and the C3 module work together on feature processing to enhance model performance. The C3 module is a standard component in the YOLOv5 architecture. The C3 module is the three convolutional layers of the CotC3 module, which realizes multi-scale feature fusion, cross-channel feature transfer, and spatial perception through three convolutional layers. It divides the input feature map into two parts for convolution operations, and finally merges the results to improve the feature capture ability and interaction efficiency at different scales, focusing mainly on the processing of local features. In contrast, the CotC3 module introduces a self-attention mechanism to better capture long-range feature dependencies and global context information. Specifically, the CotC3 module can extract richer and more contextual feature representations by segmenting the input feature map and using the self-attention mechanism to weight the features of each part. This mechanism enables the CotC3 module to not only retain fine-grained features, but also more effectively capture long-range dependencies in the image. The combination of the two enables the model to maintain high accuracy in local details while understanding image content in a larger range, thereby significantly improving the overall detection effect.

[0015] In step 1, the SPD module expands the receptive layer of the model to capture the spherical features of the remote sensing image;

[0016] Specifically, the SPD module replaces each strided convolution layer and each pooling layer with a depth layer and a non-strided convolution layer. The SPD module downsamples the feature map of the convolutional neural network (Convolutional Neural Network), considers any intermediate feature map X of size S×S×C1, and cuts out a series of sub-feature maps. In the present invention, the size of the feature map is expressed as S×S×C1, where S represents the spatial resolution of the feature map, that is, the height and width of the feature map (assuming that the feature map is square), in pixels. C1 represents the number of channels of the feature map, corresponding to the number of different feature maps output by the layer. Specifically, S×S defines the size of the feature map in the spatial dimension, and C1 represents the number of channels of the feature map in the depth direction. For example, for a feature map of size 64×64×32, its height and width are both 64 pixels, containing 32 different feature maps. This three-dimensional tensor form is a standard representation method for convolutional neural networks (CNNs) when processing image data, and can effectively capture and express multi-scale and multi-channel feature information in images.

[0017] By introducing the SPD module to downsample the CNN feature map, fine-grained features can be better preserved and the image representation learning ability can be improved. The SPD module replaces each strided convolution layer and pooling layer with a deep layer and a non-strided convolution layer, thereby reducing the spatial dimension while maintaining feature richness. This design helps the model to more accurately understand the target location, shape, and contextual relationship in remote sensing images, thereby improving the overall detection performance.

[0018] The series of sub-feature maps cut out are expressed as:

[0019] f0,0=X[0:S:scale,0:S:scale], f1,0=X[1:S:scale,0:S:scale],...,

[0020] fscale-1,0=X[scale-1:S:scale,0:S:scale];

[0021] f0,1=X[0:S:scale,1:S:scale],f1,1,...,

[0022] fscale-1,1=X[scale-1:S:scale,1:S:scale]; . .

[0025] f0,scale-1=X[0:S:scale,scale-1:S:scale],f1,scale-1,...,

[0026] fscale-1,scale-1=X[scale-1:S:scale,scale-1:S:scale]; (1)

[0027] Where scale is the size of any original feature map X, and sub-map f x,y It is formed by the entries X(i, j) with coordinates i and j in the feature map, where i+x and j+y correspond to specific scale factors respectively. The original feature map is downsampled by the scale factor. Each sub-map will downsample X by the scale factor, and these sub-feature maps are connected along the channel dimension to obtain the feature map X0. The spatial dimension of the feature map X0 is reduced by 1 times, and the channel dimension is increased by 1 times. That is, SPD transforms the feature map X(S,S,C1) into an intermediate feature map X'( scale 2C1); where: S represents the spatial resolution of the feature map, that is, the height and width of the feature map (assuming the feature map is square), in pixels. C1 represents the number of channels of the feature map, corresponding to the number of different feature maps output by this layer.

[0028] In step 2, the Neck structure optimization method is as follows: a CotNet module is introduced into the Neck structure of the traditional YOLOv5 model, and a Cot block is used to replace the one-time upsampling module therein, and the CotNet module performs upsampling in a completely new way;

[0029] The steps to build the CotNet module are:

[0030] Given an input feature map X of size H×W×C, define keys, queries, and values ​​as K=X, Q=X, and V=XWv respectively;

[0031] CoT first spatially encodes all adjacent keys in a k×k grid to obtain the context key K 1 ∈R H×W×C , and then the context key K 1 Concatenated with Q, and then through two consecutive 1×1 convolutions to get the attention matrix: A = [K 1 , Q]W θ W δ ;

[0032] The local attention matrix for each spatial position of each head A is based on the query features and the key features with context. By aggregating the context attention matrix A and valueV, the weighted feature map K is obtained. 2 ;

[0033] Given the feature map K 2 The dynamic feature interaction between the inputs is captured and named as dynamic context representation. The final output of the CotNet module is the static context K 1 and dynamic context K 2 fusion of.

[0034] Preferably, Intersection over Union (IoU) is the ratio of the intersection over union between the predicted segmentation result and the actual segmentation mask. IoU is expressed by the following formula:

[0035]

[0036] Where Intersection refers to the area where the predicted segmentation result and the true segmentation mask intersect, while Union is the area of ​​their union area, and IoU is the loss function;

[0037] In step 3, SIOU, an indicator that measures the overlap between the predicted segmentation result and the true segmentation mask in the image segmentation task, matches the IoU, center point distance, and aspect ratio between the true box and the predicted box; SIOU takes into account the angle between the vectors of the expected regression and redefines the angle penalty metric. The calculation formula of SIOU is as follows:

[0038]

[0039] Among them, Δ represents the distance cost, which is the distance between the center points of the real box and the predicted box, and Ω represents the shape cost, which is the aspect ratio of the real box and the predicted box.

[0040] In step 4, the performance of the target detection model is verified by experimental verification in terms of average precision, accuracy, recall rate and accuracy. Specifically:

[0041] The average accuracy is based on the accuracy of the target detection model on different categories, and then these accuracies are averaged to obtain a comprehensive performance evaluation, which is calculated as follows:

[0042]

[0043] In the formula, N is the number of categories and AP is the accuracy;

[0044] The precision is used to evaluate the detection performance of the model for this category. Specifically, the precision value on the PR curve is averaged. The calculation formula is:

[0045]

[0046] Among them, R represents the recall rate, that is, the proportion of samples successfully detected by the model among all samples that actually belong to this category, and the expression is:

[0047]

[0048] In the formula, TP stands for true positive, which means the true category of the sample is positive, and the result predicted by the model is also a positive example, which is a correct prediction; FN stands for false negative: the true category of the sample is a positive example, but the model predicts it as a negative example, which is a wrong prediction;

[0049] P represents the accuracy rate, which is the proportion of samples that actually belong to the category among the samples that the model identifies as the category. The expression is:

[0050]

[0051] In the formula, FP stands for false positive, which means that the true category of the sample is a negative example, but the model predicts it as a positive example, which is an incorrect prediction.

[0052] When calculating the average precision (AP), P and R are two key indicators, and the relationship between them is usually described by the PR curve (Precision-Recall curve). The PR curve is drawn by calculating the precision at different recall thresholds.

[0053] In actual operation, AP is usually calculated using the following method:

[0054] (1) Sort each sample according to the probability predicted by the model.

[0055] (2) Consider each sample as a positive example, calculate the current TP, FP, and FN, and thus obtain the current Precision and Recall.

[0056] (3) Record the maximum Precision value at each recall rate, because Precision is non-decreasing in the PR curve.

[0057] (4) Finally, the recorded precision values ​​are integrated, usually by calculating the area under the PR curve, which is the AP.

[0058] Therefore, the AP in formula (5) can be expressed as formula (8):

[0059] AP=∑(max(Precision at each Recall))ΔR, (8).

[0060] The high-definition remote sensing image detection method based on improved YOLOv5 proposed in the present invention has the following beneficial effects:

[0061] (1) This high-definition remote sensing image detection method improves the feature function fusion module in the original YOLOv5 based on the characteristics of small targets and low resolution of remote sensing images. By introducing the SPD module, the receptive layer of the model is expanded, so that the model can better retain fine-grained features and improve the image representation learning ability.

[0062] (2) An attention mechanism is added to the feature fusion module of YOLOV5, and the Cotnet module is used to replace the single upsampling module in the original model structure, so that the model can make better use of the rich contextual information between adjacent keys, thereby obtaining a more optimized feature representation and improving the detection accuracy;

[0063] (3) In this detection method, the traditional loss function in YOLOv5 is replaced, and ablation experiments are carried out to verify that the accuracy of the improved model is improved by 1.1% compared with that before the improvement. Experiments are carried out on the DOTA and NWPU-VHR10 datasets respectively. The experimental results show that compared with the traditional YOLOv5 model, the accuracy of the improved YOLOv5 model can be significantly improved, and the average accuracy can reach 93.5%. The present invention can be widely used in urban planning, land use, environmental detection and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a structural block diagram of the traditional YOLOv5 high-definition remote sensing image detection model of the present invention.

[0065] Figure 2 This is a structural block diagram of the improved YOLOv5 high-definition remote sensing image detection model of the present invention.

[0066] Figure 3 This is a diagram of the SPD downsampling process of the present invention.

[0067] Figure 4 It is a detailed structural diagram of the Cotnet module of the present invention.

[0068] Figure 5 Graph showing the experimental results of the method of the present invention.

[0069] Figure 6 Result diagram of each training parameter of the method of the present invention.

[0070] Figure 7 This is a recognition effect diagram of the method of the present invention. DETAILED DESCRIPTION

[0071] The present invention is further described below through specific embodiments in conjunction with the accompanying drawings.

[0072] Example:

[0073] Two open source satellite image datasets, DOTA and NWPU-VHR10, were selected for the experiment. In order to increase the robustness of the model to the position of the target object in the image, some data enhancement methods were used for the above two datasets, including random scaling of the images in the dataset, so that the model can process input images of different sizes and has scale invariance; because too large or too small scaling ratio may cause loss or distortion of image information, when selecting the random scaling ratio, it is chosen to enlarge to 80%, 100%, 120% of the original to ensure that the generated image still maintains reasonable recognizability and semantic information; because random flipping can provide horizontal or vertical invariance and reduce the model's sensitivity to image mirroring, the images in the dataset are randomly flipped by randomly selecting whether to flip horizontally or vertically; random rotation allows the model to handle rotation changes in the image and has rotation invariance;

[0074] See also Figure 5 and Figure 6 ,The same data set is used during the experiment to ensure that the model parameters and the input image size remain unchanged. ,The specific experimental results are shown in Table 1: The first row is the original model of YOLOv5, the second row is the improved model of YOLOv5 combined with Cotnet, the third row is the improved model of YOLOv5 combined with SPD, and the fourth row is the model of YOLOv5 using SPD and 3cotC3.

[0075] Table 1: Ablation experiment results

[0076]

[0077] Through ablation experiments, it can be seen that when the CotNet attention mechanism is added to the neck structure of Yolov5, CotNet enables the model to effectively capture more target information by learning the contextual relationship between targets, and the model's map increases from 91.2% to 91.5%;

[0078] When dilated convolution and depthwise separable convolution are introduced into the backbone structure, the model's ability to represent the target can be effectively expanded compared to normal convolution, but the map is reduced, indicating that the SPD module may not match the structure or data characteristics of the current model. When the SPD module and the Cotnet module are added at the same time, the mAP of the model is increased from the original 91.2% to 93.5%, indicating that the SPD module and the Cot module may provide synergy at different levels and task stages of the model. The SPD module may perform better in extracting specific features, while the Cot module may be better at dealing with issues such as scale changes. Their combination can enable the model to take advantage of different aspects, thereby improving the overall performance.

[0079] See also Figure 7 In the experimental comparison, different bounding box loss functions are studied in depth to explore their influence on the network detection accuracy. By comprehensively considering the EIoU, α- and SIOU bounding box loss functions, they are compared with the CIoU bounding box loss function used by YOLOv5 as shown in Table 2: From the experimental results, it can be seen that all loss functions can obtain higher precision scores, which proves the effectiveness of these loss functions. By comparing different loss functions, it is found that the use of SIOU can effectively improve the recognition efficiency of the model for high-definition remote sensing images. The precision is improved by 1.1% compared with the original loss function of YOLOv5, which proves that the modified loss function can cooperate with the YOLOv5 network to improve the target detection accuracy on the high-definition remote sensing image dataset:

[0080] Table 2 Comparison of bounding box loss functions

[0081]

[0082]

[0083] Targets in remote sensing images are usually distributed over a large area. Traditional convolution operations may not be able to capture the long-range relationship between targets. The CotNet module can capture the long-range dependency between targets through the self-attention mechanism, improve the recall rate and positioning accuracy of target detection, and introduce the SPD module, which enables the model to more accurately understand the target position, shape and contextual relationship in remote sensing images, greatly improving the accuracy and robustness of detection.

[0084] The classic Backbone structure of YOLOv5 is optimized, and the SPD module (SpatialPyramidDownsampling) and Cot module (Context-aware Transformer Network module) are introduced, so that the model can better capture the spherical features of remote sensing images, effectively retain fine-grained features, and improve the image representation learning ability. The self-attention mechanism is used to improve the Neck structure of YOLOv5, and the CoT block is used to replace the one-time upsampling module. This module effectively enhances the robustness and generalization ability of the model for different remote sensing scenes and improves the performance of the model in various situations; the loss function of the original method is modified, and the EIOU, CIOU, and SIOU loss functions are optimized. We conduct experiments on different loss functions respectively, and the results show that the best detection accuracy can be obtained by using SIOU.

[0085] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A high-definition remote sensing image detection method based on improved YOLOv5, characterized in that: The specific steps are: Step 1: Improvement of YOLOv5 model: Based on the YOLOv5 high-definition remote sensing image detection model, optimize the Backbone structure in the YOLOv5 model to retain the fine-grained features of the remote sensing image, and introduce the SPD module to expand the perception layer of the model; Step 2: Add attention mechanism to the feature fusion module: Based on the YOLOv5 high-definition remote sensing image detection model, optimize the Neck structure in the YOLOv5 model to integrate the attention mechanism, use the rich context information between adjacent keys, and optimize the feature representation; Step 3: Replace the loss function in YOLOv5: Based on the detection accuracy optimization loss function Intersection over Union in the target detection method, the loss function is used to quantify the difference between the model's prediction and the actual value; Step 4: Ablation experiment verification: Experimental verification is performed from the aspects of average precision, precision, recall rate and accuracy to verify the performance of the target detection model.

2. The high-definition remote sensing image detection method according to claim 1, characterized in that: The optimization method of the Backbone structure in the YOLOv5 model described in step 1 is: embedding the CoTC3 module between the Backbone structure and the Neck structure, and the CoTC3 module is used for global context perception and feature interaction; The CotC3 module consists of three convolutional layers, namely the multi-scale feature fusion layer, the cross-channel feature transfer layer, and the spatial perception layer. The module extracts features by dividing the input feature map into two parts, processing them separately and then merging them, using the self-attention mechanism; Specifically, the CotC3 module includes the CotNet module and the C3 module. The two parts work together. The attention mechanism in the CotNet module is responsible for capturing long-distance feature dependencies, while the C3 module focuses on the fusion of local features, thereby achieving an effective combination of global and local features.

3. The high-definition remote sensing image detection method according to claim 2, characterized in that: In step 1, the SPD module expands the receptive layer of the model to capture the spherical features of the remote sensing image; specifically, the SPD module replaces each strided convolution layer and each pooling layer by using a depth layer and a non-strided convolution layer. The SPD module downsamples the feature map of the convolutional neural network, considering any intermediate feature map X of size S×S×C1, and cuts out a series of sub-feature maps; let the size of the feature map be expressed as S×S×C1, where S represents the spatial resolution of the feature map, that is, the height and width of the feature map, in pixels; C1 represents the number of channels of the feature map, corresponding to the number of different feature maps output by the layer; specifically, S×S defines the size of the feature map in the spatial dimension, and C1 represents the number of channels of the feature map in the depth direction; the cut-out series of sub-feature maps are expressed as: f0,0=X[0:S:scale,0:S:scale], f1,0=X[1:S:scale,0:S:scale],..., fscale-1,0=X[scale-1:S:scale,0:S:scale]; f0,1=X[0:S:scale,1:S:scale],f1,1,..., fscale-1,1=X[scale-1:S:scale,1:S:scale]; . . f0,scale-1=X[0:S:scale,scale-1:S:scale],f1,scale-1,..., fscale-1,scale-1=X[scale-1:S:scale,scale-1:S:scale]; (1) Where scale is the size of any original feature map X, and sub-map f x,y It is formed by the entries X(i,j) with coordinates i and j in the feature map, where i+x and j+y correspond to specific scale factors respectively. The original feature map is downsampled by the scale factor. Each sub-map will downsample X by the scale factor. These sub-feature maps are connected along the channel dimension to obtain the feature map X0. The spatial dimension of the feature map X0 is reduced by 1 times, and the channel dimension is increased by 1 times. That is, SPD transforms the feature map X(S,S,C1) into an intermediate feature map X' Among them: S represents the spatial resolution of the feature map, that is, the height and width of the feature map, in pixels, and C1 represents the number of channels of the feature map, corresponding to the number of different feature maps output by this layer.

4. The high-definition remote sensing image detection method according to claim 3, characterized in that: In step 2, the Neck structure optimization method is as follows: a CotNet module is introduced into the Neck structure of the traditional YOLOv5 model, and a Cot block is used to replace the one-time upsampling module therein, and the CotNet module performs upsampling in a completely new way; The steps to build the CotNet module are: Given an input feature map X of size H×W×C, define keys, queries, and values ​​as K=X, Q=X, and V=XWv respectively; CoT first spatially encodes all adjacent keys in a k×k grid to obtain the context key K 1 ∈R H ×W×C , and then the context key K 1 Concatenated with Q, and then through two consecutive 1×1 convolutions to get the attention matrix: A = [K 1 , Q]W θ W δ ; The local attention matrix for each spatial position of each head A is based on the query features and the key features with context. By aggregating the context attention matrix A and valueV, the weighted feature map K is obtained. 2 ; Given the feature map K 2 The dynamic feature interaction between the inputs is captured and named as dynamic context representation. The final output of the CotNet module is the static context K 1 and dynamic context K 2 fusion of.

5. The high-definition remote sensing image detection method according to claim 4, characterized in that: In step 3, let IoU be the ratio of the intersection and union between the predicted segmentation result and the actual segmentation mask. IoU is expressed by the following formula: In the formula, Intersection refers to the area where the predicted segmentation result and the true segmentation mask intersect, while Union is the area of ​​their union area, and IoU is the loss function.

6. The high-definition remote sensing image detection method according to claim 5, characterized in that: In step 3, SIOU, an indicator that measures the overlap between the predicted segmentation result and the true segmentation mask in the image segmentation task, matches the IoU, center point distance, and aspect ratio between the true box and the predicted box; SIOU takes into account the angle between the vectors of the expected regression and redefines the angle penalty metric. The calculation formula of SIOU is as follows: Among them, Δ represents the distance cost, specifically the distance between the center point of the real box and the predicted box, and Ω represents the shape cost, specifically the aspect ratio of the real box and the predicted box.

7. The high-definition remote sensing image detection method according to claim 6, characterized in that: In step 4, the performance of the target detection model is verified by experimental verification in terms of average precision, accuracy, recall rate and accuracy. Specifically: The average accuracy is based on the accuracy of the target detection model on different categories, and then these accuracies are averaged to obtain a comprehensive performance evaluation, which is calculated as follows: In the formula, N is the number of categories and AP is the accuracy; The precision is used to evaluate the detection performance of the model for this category. Specifically, the precision value on the PR curve is averaged. The calculation formula is: Among them, R represents the recall rate, that is, the proportion of samples successfully detected by the model among all samples that actually belong to this category, and the expression is: In the formula, TP stands for true positive, which means the positive example of the true category of the sample, and the result predicted by the model is also a positive example, which is a correct prediction; FN stands for false negative: the true category of the sample is a positive example, but the model predicts it as a negative example, which is a wrong prediction; P stands for accuracy, which is the proportion of samples that actually belong to this category among the samples identified by the model as this category, and the expression is: In the formula, FP stands for false positive, which means that the true category of the sample is a negative example, but the model predicts it as a positive example, which is an incorrect prediction.

Images