Rotating target detection method based on corner point positioning and target contour point set

Abstract

The present invention discloses a rotation target detection method based on corner point positioning and target contour point set, comprising the following steps: Step 1: input remote sensing images, use CornerNet model to read corner points, and obtain target frame; Step 2: perform feature extraction on remote sensing images through Reppoint deformable convolution to obtain target contour point set; Step 3: obtain the central axis of the target contour point set; Step 4: determine the rotation angle θ based on the central axis; Step 5: rotate the target frame to determine the rotated target: rotate the target frame accordingly by the angle θ, and then translate and scale each boundary line of the target frame respectively, and the figure defined by the multiple boundary lines is the rotated target. The present invention has a simple structure and a reasonable design. The present invention uses the CornerNet network structure and the central axis of the target contour point set to respectively perform target positioning, calculates the rotation angle based on the central axis, rotates the target frame of the CornerNet network structure, and rotates the horizontal target frame of the CornerNet network structure by the corresponding angle to complete the rotation target detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a rotation target detection method based on corner point positioning and target contour point sets. Background Art

[0002] Images are a crucial source of human understanding of the world, conveying richer, more tangible, and more specific information than any other form. With the continuous development of society and the accelerating pace of urbanization, urban order is becoming increasingly chaotic. Object detection, a key branch of computer vision, is being widely applied in fields such as industrial inspection, road traffic, and aerospace. For example, cameras can capture, process, and store real-time information about people on the road, thereby reducing the intensity of criminal investigations and the consumption of human capital. Therefore, this technology has important practical significance.

[0003] Object detection can be generally categorized into horizontal detection and rotation detection based on the orientation of the target bounding box. Specifically, horizontal bounding box detection is generally more suitable for general natural scene images. However, scenarios such as remote sensing images, face recognition, and license plate recognition typically require more precise positioning, which necessitates an effective rotational object detection model.

[0004] The CornerNet model, also known as a corner object detection model, is a new single-stage object detection method that eliminates anchor boxes. Each object is detected as a pair of keypoints—the top-left and bottom-right corners of its bounding box. Using a single convolutional network, it predicts a heatmap for the top-left corner and a heatmap for the bottom-right corner of all instances of the same object category, along with an embedding vector for each detected corner point. (This serves to group pairs of corner points belonging to the same object, training the network to predict similar embedding vectors for both corner points belonging to the same object.) CornerNet simplifies the network output, eliminating the need for designing anchor boxes.

[0005] RepPoints is a set of target contour points that can adaptively surround the target and contain the semantic features of the local area. The training of RepPoints is driven by both target positioning and target classification, which can constrain RepPoints to tightly surround the target and guide the detector to correctly classify the target. This adaptive representation method is differentiable and can be used continuously in multiple stages of the detector, and does not require additional anchors to generate a large number of initial boxes.

[0006] During image rotation, objects that have been rotated at a certain angle need to be detected to determine whether they have reached the specified angle. In certain scenarios, such as remote sensing image detection, objects in images often exhibit rotation angles. Using conventional object detection methods to detect objects in these situations will result in low accuracy and may contain a large amount of useless redundant information.

[0007] In view of this, it is necessary to propose a rotation target detection method based on the CornerNet model and RepPoints to solve the above problems. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the deficiencies in the above-mentioned prior art and provide a method for detecting rotated targets based on corner point positioning and target contour point set. The method has a simple structure and a reasonable design. The present invention adopts the CornerNet network structure and the central axis of the target contour point set to respectively locate the target, calculates the rotation angle according to the central axis, rotates the target frame of the CornerNet network structure by a corresponding angle, and completes the detection of rotated targets.

[0009] To solve the above technical problems, the present invention adopts a technical solution: a rotating target detection method based on corner point positioning and target contour point set, characterized in that it includes the following steps:

[0010] Step 1: Input remote sensing images, use the CornerNet model to read corner points, obtain target boxes, and perform preprocessing;

[0011] Step 2: Extract features from the remote sensing image through Reppoint deformable convolution to obtain the target contour point set;

[0012] Step 3: Get the central axis of the target contour point set;

[0013] Step 4: Determine the rotation angle based on the central axis ;

[0014] Step 5: Rotate the target frame accordingly Angle, and then translate and scale each boundary line of the target frame until it hits any point in the target contour point set, and then stops. The figure defined by multiple boundary lines is the rotated target.

[0015] The above-mentioned rotating target detection method based on corner point positioning and target contour point set is characterized in that the specific method of step five is:

[0016] Step 501: Establish the equation of each boundary line based on the corner points of the target box , where i represents the number of boundary lines, represents the slope of the i-th boundary line equation, represents the intercept of the i-th boundary line equation, ;

[0017] Step 502: Determine whether any point in the target contour point set satisfies the boundary line equation If it is satisfied, the boundary line equation at this time is determined to be the modified boundary line equation; if it is not satisfied, the modified boundary line equation is constructed ,Change , until any point in the target contour point set satisfies the modified boundary line equation , Represents the translation and scaling parameters of the i-th modified boundary line equation.

[0018] The above-mentioned rotation target detection method based on corner point positioning and target contour point set is characterized in that: the "changing The specific method is: in the equations of two parallel boundary lines, if ,but gradually decreases, Gradually increase, .

[0019] The above-mentioned rotating target detection method based on corner point positioning and target contour point set is characterized in that: the specific method of step three is: establishing the central axis equation of the target contour point set , determine the parameters k and b so that the sum of the distances from the target contour point set to the central axis is minimized.

[0020] The above-mentioned rotation target detection method based on corner point positioning and target contour point set is characterized in that: the specific method of step 4 is: the rotation angle , k represents the slope of the central axis.

[0021] The above-mentioned rotating target detection method based on corner point positioning and target contour point set is characterized in that: if When the angle is positive, the target frame is rotated clockwise angle; if When the angle is negative, the target frame is rotated counterclockwise angle.

[0022] The above-mentioned rotating target detection method based on corner point positioning and target contour point set is characterized in that the specific method of step 1 is:

[0023] Step 101: A certain amount of remote sensing images with different shooting environments, different lighting conditions, and different tilt angles are obtained as training samples;

[0024] Step 102: manually mark the upper left corner and lower right corner of the training sample, where the combination of the upper left corner and the lower right corner constitutes the target box;

[0025] Step 103: Building an image recognition network based on the CornerNet network structure;

[0026] Step 104: Send the training samples to the image recognition network for machine learning. After the learning is completed, a trained image recognition network is obtained.

[0027] Step 105: Use the trained image recognition network to predict the remote sensing image to obtain the target box, upper left corner point and lower right corner point of the remote sensing image.

[0028] The above-mentioned rotating target detection method based on corner point positioning and target contour point set is characterized in that: the step 1 is specifically as follows:

[0029] The remote sensing image is fed into a feature extraction module to obtain a feature map with deep features;

[0030] The feature map is then sent to the corner pooling layer to obtain the corner feature map, and then the corner feature map is sent to the prediction network to obtain the corner points and target boxes.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] 1. The present invention has a simple structure, reasonable design, and is easy to implement and operate.

[0033] 2. The present invention builds an image recognition network based on the CornerNet network structure, obtains the corner points of the remote sensing image through the trained image recognition network, and then uses the minimum circumscribed rectangle of the corner points as the target frame, avoiding the defect of huge computational complexity when predicting the target frame in the existing technology.

[0034] 3. The present invention adopts Reppoint deformable convolution to obtain the target contour point set of the target, and uses a set of point sets to represent the contour of an object, indicating the shape, posture and semantically important local areas, providing finer-grained positioning and more convenient classification.

[0035] 4. The present invention uses the central axis of the CornerNet network structure and the target contour point set to respectively locate the target, calculates the rotation angle according to the central axis, rotates the target frame of the CornerNet network structure, and rotates the horizontal target frame of the CornerNet network structure by a corresponding angle to complete the rotation target detection.

[0036] 5. When the horizontal target frame of the CornerNet network structure is rotated by a corresponding angle, the four boundary lines simultaneously move toward the center of the graphic defined by the boundary lines, eliminating redundant information and achieving good use effect.

[0037] In summary, the present invention has a simple structure and a reasonable design. The present invention uses the CornerNet network structure and the central axis of the target contour point set to respectively locate the target, calculates the rotation angle according to the central axis, rotates the target frame of the CornerNet network structure, and rotates the horizontal target frame of the CornerNet network structure by a corresponding angle to complete the rotation target detection.

[0038] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Flow chart of the method of the present invention.

[0040] Figure 2 This is a flow chart of the method for rotating a target frame and determining a rotated target according to the present invention.

[0041] Figure 3 Schematic diagram of the boundary line in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments of the present invention.

[0043] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0045] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can, for example, be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0046] For ease of description, spatially relative terms such as "above", "above", "on the upper surface of", "above", etc. may be used herein to describe the spatial positional relationship of a device or feature to other devices or features as shown in the figures. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figures. For example, if the device in the drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" can include both "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatially relative descriptions used here are interpreted accordingly.

[0047] The rotating target detection method based on corner point positioning and target contour point set provided in the embodiment of the present application can be applied to terminal devices such as mobile phones, tablet computers, laptops, ultra-mobile personal computers (UMPCs), netbooks, and personal digital assistants (PDAs). The embodiment of the present application does not impose any restrictions on the specific type of the system.

[0048] like Figure 1 As shown, according to an embodiment of the present invention, an embodiment of a rotation target detection method based on corner point positioning and target contour point set is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0049] Step 1: Input the remote sensing image, use the CornerNet model to read the corner points, and obtain the target frame. The specific method is as follows:

[0050] Step 101: A certain number of remote sensing images with different shooting site environments, different lighting conditions, and different tilt angles are obtained as training samples; and the remote sensing images are pre-processed by band synthesis, image fusion, and image segmentation.

[0051] Remote sensing images have very high image resolution. Directly inputting raw images is unrealistic, whether in training or actual detection. This will place a heavy burden on the equipment and significantly affect the training speed. Therefore, in one possible embodiment, the remote sensing images should be preprocessed by band synthesis, image fusion, and image segmentation before being input into the model for training and prediction.

[0052] In one possible embodiment, training set data and test set data are obtained from the public remote sensing image dataset DOTA dataset, the ratio of training set to test set image data is 3:2, the training set and test set image file formats are ".png" file format, and the data categories include 15 categories: airplanes, ships, oil storage tanks, baseball fields, tennis courts, basketball courts, track and field fields, ports, bridges, large vehicles, small vehicles, helicopter airports, roundabouts, football fields and swimming pools.

[0053] In one possible embodiment, the training set data and the test set data are obtained from the public remote sensing image dataset Pascal VOC 2012 detection.

[0054] Step 102: Manually mark the upper left corner and lower right corner of the training sample, where the combination of the upper left corner and the lower right corner constitutes the target frame; the upper left corner and the lower right corner can confirm the width w and height h of the target frame, thereby determining the target frame.

[0055] Step 103: Build an image recognition network based on the CornerNet network structure.

[0056] CornerNet's network structure mainly includes: borrowing the hourglass network as the backbone structure for feature extraction; the network output content adopts a two-branch output, top-left corners and bottom-right corners, each branch contains its own corner pooling, and three-branch output. CornerNet simplifies the network output and eliminates the need to design anchor boxes.

[0057] Specifically, the remote sensing image is fed into the hourglass network to obtain a feature map with deep features, without changing the data size but only the data depth. The feature map is then fed into the two branches of top-left corners and bottom-right corners, and a corner feature map is obtained through the corner pooling layer. The corner feature map is then fed into the prediction network to obtain the corner points and target boxes.

[0058] The loss function of the CornerNet network structure is , represents the loss of the heat map, represents the loss of similar grouped vectors, represents the loss of heterogeneous grouping vectors, Indicates the loss of offset.

[0059]

[0060] and are the predicted probability and true probability of channel c at point (i, j), C, H, and W are the number of channels and height and width respectively. N represents the number of objects in the image.

[0061]

[0062] e tk With e bk are the grouping vectors of the upper left corner and the lower right corner respectively, e k Indicates e tk and e bk The mean of .

[0063]

[0064] N represents the number of objects in the image, and n is the downsampling ratio.

[0065] Step 104: Send the training samples to the image recognition network for machine learning. After the learning is completed, a trained image recognition network is obtained.

[0066] Step 105: Use the trained image recognition network to predict the remote sensing image to obtain the target box, upper left corner point and lower right corner point of the remote sensing image.

[0067] Based on the CornerNet network structure, an image recognition network is constructed. The corner points of the remote sensing image are obtained through the trained image recognition network, and then the minimum enclosing rectangle of the corner points is used as the target frame, avoiding the defect of huge computational complexity when predicting the target frame in existing technologies.

[0068] Step 2: Extract features from the remote sensing image through Reppoint deformable convolution to obtain the target contour point set;

[0069] Deformable convolution consists of two steps: 1) sampling the input feature map x with a regular grid R (3×3); 2) summing the weighted sampled values.

[0070] The grid R defines the size and expansion of the receptive field. For example, a 3×3 convolution kernel with an expansion size of 1 is defined. , for each position p0 on the output feature map y, we have , where P n The positions in R are enumerated.

[0071] In deformable convolution, a regular grid R is shifted by {∆p n |n=1,…,N} increases, where N=|R|. The equation becomes .

[0072] Usually the number of point sets is the square of a certain number. In this embodiment, the number of sampling points is 9. Classification and regression are performed by learning the offsets of the sample points in the x and y directions and combining them with pixel features.

[0073] RepPoint deformable convolution changes the traditional representation of objects in object detection from rectangular boxes to a set of points that represent the outline of an object. RepPoints learning is driven by both target localization and target identification losses. Deformable convolution and RepPoints are combined to learn offsets for the deformable convolution using supervisory signals from localization and classification. The offsets can be used to locate the object, while semantic information can be used to identify the object. RepPoints adaptively learns to position itself above the object, limiting its spatial extent and representing its shape, pose, and semantically important local regions. This approach provides finer-grained localization and facilitates classification.

[0074] Step 3: Get the central axis of the target contour point set , determine the parameters k and b so that the sum of the distances from the target contour point set to the central axis is minimized; the parameter k represents the slope, and the parameter b represents the intercept.

[0075] The nine target contour point sets determined by Reppoint and the central axis obtained by linear regression can describe the main rotation direction of the object to the greatest extent, and can effectively avoid the difficulties in determining the rotation direction caused by the complex shape and irregular contour of the target.

[0076] In a two-dimensional figure, the central axis is the line formed by the centers of the largest circle that has two or more tangent points to the figure's edges. Using the central axis to locate the target and calculate the rotation angle based on it is a universal solution with excellent performance, as all target contour point sets have a corresponding central axis.

[0077] In specific implementation, for any central axis, the sum of its distances to all points in the target contour point set is calculated, and the central axis with the smallest sum of distances from the target contour point set to the central axis is the central axis required to determine the rotation angle.

[0078] Step 4: Determine the rotation angle according to the central axis. ;

[0079] Step 5: Rotate the target frame accordingly angle, if When the angle is positive, the target frame is rotated clockwise angle; if When the angle is negative, the target frame is rotated counterclockwise angle.

[0080] Step 5: Rotate the target frame accordingly After the rotation is completed, each boundary line of the target frame is translated and scaled separately. When the boundary line touches any point in the target contour point set, the translation and scaling is stopped. The figure defined by multiple boundary lines is the rotated target.

[0081] The present invention determines the rotation angle of the remote sensing image through the central axis, and determines the position of the rotating target of the remote sensing image through a graphic defined by multiple boundary lines, retains the direction information of the rotating target, avoids the defect of losing the rotation angle when rotating the remote sensing image in the prior art, and at the same time does not reduce the robustness of target positioning, thereby improving the detection accuracy of the rotating target.

[0082] like Figure 2 As shown, each boundary line of the target frame is translated and scaled separately. The specific method is:

[0083] Step 501: Establish the equation of each boundary line based on the corner points of the target box , where i represents the number of boundary lines, represents the slope of the i-th boundary line equation, represents the intercept of the i-th boundary line equation.

[0084] It should be noted that among the four boundary lines, two are parallel, and two of them are parallel to the central axis. are parallel, so the slope of these two boundary lines is k, and the slope of the remaining two boundary lines is So the slope is known, for the linear equation For example, it must pass through one of the two corner points. According to the corner point that each boundary line passes through, solve an unknown quantity in the boundary line equation. , we can get the equations of the four boundary lines.

[0085] Step 502: Determine whether any point in the target contour point set satisfies the boundary line equation If it is satisfied, the boundary line equation at this time is determined to be the modified boundary line equation; if it is not satisfied, the modified boundary line equation is constructed , the correction method is to change , until any point in the target contour point set satisfies the modified boundary line equation , Represents the translation and scaling parameters of the i-th modified boundary line equation.

[0086] It stops until it hits any point in the target contour point set, that is, if any point in the target contour point set satisfies the boundary line equation of translation and scaling , that is, it is considered that the boundary line hits the point where the target contour points are concentrated, and the figure defined by multiple boundary lines is the rotated target.

[0087] like Figure 3 As shown, the equations of the four boundary lines are It should be noted that Boundary Line 1 and Boundary Line 2 are parallel to each other, but the intercept of Boundary Line 1 is 1, and the intercept of Boundary Line 2 is -1. Boundary Line 1's intercept is greater than that of Boundary Line 2, so Boundary Line 1 is located in the upper part of the figure defined by the boundary lines. Similarly, Boundary Line 3 is located in the upper part, and Boundary Lines 2 and 4 are located in the lower part.

[0088] Therefore, during the translation and scaling process, the boundary lines 2 and 4 Take positive value, and Gradually increase, so that boundary line 2 and boundary line 4 shrink toward the center of the figure defined by the boundary line; boundary line 1 and boundary line 3 Takes a negative value, and Gradually decrease, so that boundary line 1 and boundary line 3 shrink toward the center of the figure defined by the boundary lines, so that the four boundary lines move toward the center of the figure defined by the boundary lines at the same time, eliminating redundant information.

[0089] like Figure 3 As shown, the boundary line , boundary line , boundary line and boundary lines To correct the boundary line equation, it can be seen from the figure that compared with boundary line 1, boundary line 2, boundary line 3, and boundary line 4, boundary line , boundary line , boundary line and boundary lines At the same time, it moves toward the center of the figure defined by the boundary line, shrinking the range of the figure defined by the boundary line. The dotted line frame in the figure is the detected rotated target. In this way, the advantages of CornerNet and RepPoint point set can be combined to convert the horizontal frame into a rotated frame while ensuring accuracy and completing the detection of rotated targets.

[0090] In summary, the embodiment of the present application provides a method for detecting a rotating target based on corner point positioning and target contour point set, which adopts the CornerNet model to read the corner points, and then uses the minimum circumscribed rectangle of the corner points as the target frame, avoiding the defect of large amount of calculation when predicting the target frame in the existing technology; the remote sensing image is feature extracted by Reppoint deformable convolution to obtain the target contour point set, and the Reppoint deformable convolution provides finer-grained positioning and more convenient classification; the central axis is determined according to the target contour point set, the rotation angle is calculated according to the central axis, and the target frame of the CornerNet network structure is rotated by the corresponding angle to complete the detection of the rotating target, and the four boundary lines are simultaneously moved toward the center of the figure defined by the boundary lines, eliminating redundant information, and the use effect is good.

[0091] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0092] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0093] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A rotating target detection method based on corner point positioning and target contour point set, characterized by: The following steps are involved: Step 1: Input remote sensing images, use the CornerNet model to read corner points, obtain target frames, and perform preprocessing; Step 2: Extract features of remote sensing images through Reppoint deformable convolution to obtain the target contour point set; Step 3: Get the central axis of the target contour point set; Step 4: Determine the rotation angle based on the central axis ; Step 5: Rotate the target frame and determine the rotated target: Rotate the target frame accordingly Angle, and then translate and scale each boundary line of the target frame until it hits any point in the target contour point set, and then stops. The figure defined by multiple boundary lines is the rotated target; The specific method of step five is: Step 501: Establish the equation of each boundary line based on the corner points of the target box , where i represents the number of boundary lines, represents the slope of the i-th boundary line equation, represents the intercept of the i-th boundary line equation, ; Step 502: Determine whether any point in the target contour point set satisfies the boundary line equation , if it is satisfied, then the boundary line equation at this time is determined to be the modified boundary line equation; If not satisfied, construct the modified boundary line equation ,Change , until any point in the target contour point set satisfies the modified boundary line equation , represents the translation and scaling parameter of the i-th modified boundary line equation; The "change The specific method is: in the equations of two parallel boundary lines, if ,but gradually decreases, Gradually increase, ; The specific method of step three is: establish the central axis equation of the target contour point set , determine the parameters k and b so that the sum of the distances from the target contour point set to the central axis is minimized; The specific method of step 4 is: rotation angle , k represents the slope of the central axis; like When the angle is positive, the target frame is rotated clockwise angle; if When the angle is negative, the target frame is rotated counterclockwise angle.

2. The rotating target detection method based on corner point positioning and target contour point set according to claim 1, characterized in that: The specific method of step one is: Step 101: A certain amount of remote sensing images with different shooting environments, different lighting conditions, and different tilt angles are obtained as training samples; Step 102: manually mark the upper left corner and lower right corner of the training sample, where the combination of the upper left corner and the lower right corner constitutes the target box; Step 103: Building an image recognition network based on the CornerNet network structure; Step 104: Send the training samples to the image recognition network for machine learning. After the learning is completed, a trained image recognition network is obtained. Step 105: Use the trained image recognition network to predict the remote sensing image to obtain the target frame and four corner points of the remote sensing image.

3. The rotating target detection method based on corner point positioning and target contour point set according to claim 1 or 2, characterized in that: The step 1 is specifically as follows: The remote sensing image is fed into a feature extraction module to obtain a feature map with deep features; The feature map is then sent to the corner pooling layer to obtain the corner feature map, and then the corner feature map is sent to the prediction network to obtain the corner points and target boxes.

Images