Object detection method based on boundary continuity-aware Gaussian encoder

Abstract

The present invention relates to the field of directed target detection, and in particular to a target detection method based on a boundary continuity-aware Gaussian encoder. The method comprises: obtaining two-dimensional image data to be subjected to directed target detection; inputting the two-dimensional image data into a constructed directed target detector; wherein the directed target detector includes a boundary continuity-aware Gaussian encoder module; obtaining a periodically continuous two-dimensional Gaussian distribution based on the two-dimensional image data and the boundary continuity-aware Gaussian encoder module, and obtaining a two-dimensional directed detection frame based on the two-dimensional Gaussian distribution. The present invention utilizes a periodically continuous two-dimensional Gaussian distribution to describe a directed two-dimensional rectangular region, thereby eliminating boundary discontinuities and improving detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of directed target detection in the field of computer vision, and in particular to a target detection method based on a boundary continuity-aware Gaussian encoder. Background Art

[0002] Object detection is a fundamental research topic in computer vision. As an emerging branch of object detection, oriented object detection aims to extend the capabilities of traditional horizontal detectors to handle instances with rotated orientations. Common applications of oriented detectors include remote sensing image understanding and scene text detection. Aerial remote sensing images are typically captured from significant altitudes to focus on ground objects. Compared to traditional horizontal detectors, objects in these rotated object detection scenarios are often arbitrarily oriented, densely arranged, and exhibit large aspect ratios. Therefore, precise detection and localization accuracy are essential detector output requirements.

[0003] Typical solutions include: directed target detection using traditional non-Gaussian representation and directed target detection using Gaussian representation.

[0004] The most commonly used representation in target detection is the five-parameter OBB (Oriented Bounding Box), which uses L1 loss as the regression loss for the five parameters. Perform regression, where Representing the center coordinates, width, height, and angle of the box, respectively. Early methods, such as RRPN (Rotation Region Proposal Network), predicted the angle as a simple learning parameter and directly performed angle regression. Only by specifying the representation can the OBB be accurately converted into a precisely oriented rectangular region. The various shapes in this process also hinder the loss calculation when dealing with two different forms of predicted bounding boxes and ground truth boxes.

[0005] Recently, CSL (Circular Smooth Label) employs a classification approach to estimate the degree of rotation of a target rectangular region by classifying angle values ​​into multiple angle categories. Furthermore, some researchers have proposed using the eight-parameter QBB (Quadrilateral Bounding Box) to represent the coordinates of the four corner points of a quadrilateral bounding box, thereby obtaining more refined boundaries and corners. In addition to the BBox representation, anchor-free methods employ flexible point sets to represent oriented objects. This approach is inspired by the horizontal object detection method RepPoint, which proposes learning a set of points that can represent objects of varying scales, shapes, and orientations. OriRepPoints further uses various quality assessment metrics to obtain higher-quality point sets. In addition, there are many other methods for representing a given oriented rectangular region in an image. GlidingVertex encodes a rectangular region as four vertex sliding offsets on the regression branch of the prediction head, using a vertex sliding method on a horizontal bounding box to accurately locate the object.

[0006] In addition, there are other complex representation methods, such as polar coordinates that use polar coordinates to bypass the angular cycle, and middle lines that predict the target box as a pair of middle lines. However, due to the increasing complexity of encoder design, these different representation methods are difficult to continue to use in subsequent research, which makes it challenging to integrate them with existing mainstream detection frameworks. This difficulty has led to the current state-of-the-art oriented target detectors still relying on Five-parameter OBB encoder function.

[0007] The current directional target detection method mainly adheres to the horizontal target detection strategy, using a rotating envelope to represent the rotating rectangular area. The most commonly used representation method is the oriented envelope. However, when using the OBB with smooth L1 loss to regress the angle value (such as ), the visual distance between the ground truth box (GT) angle value 89° and the predicted angle value 1° is small, which is acceptable in vision and practical applications, but the calculated distance and loss value from 1° to 89° will become very large. This mismatch error caused by the representation is called the boundary discontinuity problem.

[0008] Recently, several methods have been proposed to address the above problems. Classification-based methods treat angle values ​​as 256 classes, converting the angle estimation problem into a classification task. However, this method has difficulty in accurately locating finer values ​​within the angle class. For example, in the [1.4°, 2.8°] class, the model cannot further accurately estimate the rotation angle. In addition, PointSet utilizes several individual points (usually 9 points) to estimate the boundary of the oriented object instead of directly predicting the angle value, and the prediction error of a single point may lead to misalignment of the positioning of the entire box. The above representations encounter inherent problems due to the nature of their definition, as they all use discontinuous angle definitions to describe the degree of rotation.

[0009] To address the boundary discontinuity problem, an intuitive solution involves calculating the actual intersection over union (SkewIoU) to measure the distance between the predicted box and the ground truth box. However, the SkewIoU score is sensitive to the deviation of object position and is not differentiable, making it unsuitable for neural network regression.

[0010] In order to break the paradigm of the existing regular regression framework, GWD (Gaussian Wasserstein DistanceLoss, standardized Gaussian Wasserstein distance) introduces a matrix transformation to map OBB to a Gaussian distribution to simulate SkewIoU. This indicates a high probability of boundary discontinuities. While GWD mitigates this by designing an angle-independent bounding box position loss, it does not fundamentally address the issue. The predictions output by the detection head still exhibit boundary discontinuities. Consequently, the output optimization is not geometrically aligned with the quadratic transform loss function, reducing the effectiveness of the subsequently designed GWD loss function. Summary of the Invention

[0011] In order to solve the problem of existing oriented target detectors selecting angle values To describe the degree of rotation of the envelope, the angle value has a cycle period of 0°-180°. This cycle causes this representation method to have boundary discontinuity, which affects the positioning accuracy of the detector. The embodiment of the present invention provides a target detection method and device based on a boundary continuity-aware Gaussian encoder. The technical solution is as follows:

[0012] In one aspect, a method for object detection based on a boundary continuity-aware Gaussian encoder is provided. The method is implemented by an object detection device and includes:

[0013] S1. Acquire two-dimensional image data for directed target detection.

[0014] S2. Input the two-dimensional image data into the constructed directed target detector; wherein the directed target detector includes a boundary continuity-aware Gaussian encoder module.

[0015] S3. According to the two-dimensional image data and the boundary continuity perception Gaussian encoder module, a periodically continuous two-dimensional Gaussian distribution is obtained, and a two-dimensional directional detection frame is obtained according to the two-dimensional Gaussian distribution.

[0016] Optionally, the construction process of the directed object detector in S2 includes:

[0017] S21. Obtain two-dimensional image sample data.

[0018] S22. Obtain a candidate frame in the two-dimensional image sample data through a candidate frame generation network module.

[0019] S23. Input the candidate box and the two-dimensional image sample data into the boundary continuity-aware Gaussian encoder module to obtain a directed envelope box, and convert the directed envelope box into a two-dimensional Gaussian distribution.

[0020] S24. Input the two-dimensional Gaussian distribution into the decoder module to obtain the decoder module output.

[0021] S25. Design a loss function for the directed target detector, optimize the directed target detector according to the loss function, and obtain a constructed directed target detector.

[0022] Optionally, the encoding process of the boundary continuity-aware Gaussian encoder module in S23 is as shown in the following formula (1):

[0023] (1)

[0024] Subdividing formula (1), we get the following formula (2):

[0025] (2)

[0026] Where, represents the normalized two-dimensional Gaussian distribution of the feature space of the ground truth, represents the mean of the normalized two-dimensional Gaussian distribution of the feature space of the ground truth, The variance of the normalized two-dimensional Gaussian distribution representing the feature space of the ground truth, The set of vectors representing the ground truth values, represents the mean of the two-dimensional Gaussian distribution of the actual scale space of the ground truth, The variance of the two-dimensional Gaussian distribution representing the actual scale space of the ground truth, Represents the encoder encoding function that converts the two-dimensional Gaussian distribution of the actual scale space of the ground truth into the two-dimensional Gaussian distribution of the unitized feature space, represents the two-dimensional Gaussian distribution of the actual scale space, represents the candidate anchor box, represents the horizontal coordinate corresponding component in the ground truth vector group, Indicates the central horizontal coordinate of the directed envelope. represents the central horizontal coordinate of the ground truth directed envelope, represents the central horizontal coordinate of the candidate anchor box, The subscript anchor represents the anchor box. Indicates the width of the candidate anchor box, represents the ordinate corresponding component in the ground truth vector group, Indicates the central ordinate of the directed envelope. represents the central ordinate of the ground truth directed envelope, represents the central ordinate of the candidate anchor box, Indicates the height value of the candidate anchor box, represents the scale adjustment parameter, represents the scaling mapping function, Represents the three components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space, represents the scale-invariant adaptive factor component, Indicates the scale, represents the angle-dependent component, Represents the angle of the directed envelope, Indicates the first adjustable parameter, Represents the respective index numbers of the components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space.

[0027] Optionally, the decoding process of the decoder module in S24 is as shown in the following formula (3):

[0028] (3)

[0029] Subdividing formula (3), we get the following formula (4):

[0030] (4)

[0031] Where, represents the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the variance of the predicted two-dimensional Gaussian distribution of the actual scale space, Indicates the predicted subscript, Represents the decoding conversion function that transforms the predicted unit vectorized two-dimensional Gaussian distribution combined with the candidate anchor box into a two-dimensional Gaussian distribution in the actual scale space. Represents the predicted two-dimensional Gaussian distribution of the feature space unit, represents the candidate anchor box, Represents the abscissa component of the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the horizontal coordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unit, Indicates the actual scale width value of the candidate anchor box, Indicates the actual scale horizontal coordinate value of the candidate anchor box, Represents the abscissa component of the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the ordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unitization, Indicates the actual scale height value of the candidate anchor box, Indicates the actual scale ordinate value of the candidate anchor box, Represents the scaling mapping function The inverse function of Represents the output vector group of the prediction head.

[0032] Optionally, the construction process of the directed object detector further includes:

[0033] Convert the two-dimensional Gaussian distribution into a directed envelope as shown in the following equation (5):

[0034] (5)

[0035] Where, Represents the angle of the predicted directional envelope, Represents the predicted value, The first component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The second component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The third component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, Indicates the width corresponding to the equivalent OBB, , , Indicates the height corresponding to the equivalent OBB.

[0036] Optionally, the loss function of the oriented target detector in S25 is as shown in the following formula (6):

[0037] (6)

[0038] Where, represents the loss function based on KL divergence, Represents the equivalent distance between the predicted Gaussian distribution based on KL divergence and the ground truth Gaussian distribution, represents the ground truth Gaussian distribution, represents the predicted Gaussian distribution, represents the loss function based on Bhattacharyya distance, Represents the equivalent distance between the predicted Gaussian distribution based on Bhattacharyya distance and the ground truth Gaussian distribution, represents the loss function based on Wasserstein distance, Represents the equivalent distance between the predicted Gaussian distribution and the ground truth Gaussian distribution based on the Wasserstein distance.

[0039] On the other hand, a target detection device based on a boundary continuity-aware Gaussian encoder is provided, which is applied to a target detection method based on a boundary continuity-aware Gaussian encoder, and the device includes:

[0040] The acquisition module is used to acquire two-dimensional image data to be used for oriented target detection.

[0041] An input module is used to input two-dimensional image data into a constructed oriented target detector; wherein the oriented target detector includes a boundary continuity-aware Gaussian encoder module.

[0042] The output module is used to perceive the Gaussian encoder module based on the two-dimensional image data and the boundary continuity, obtain a periodically continuous two-dimensional Gaussian distribution, and obtain a two-dimensional directed detection frame based on the two-dimensional Gaussian distribution.

[0043] Optionally, the input module is further configured to:

[0044] S21. Obtain two-dimensional image sample data.

[0045] S22. Obtain a candidate frame in the two-dimensional image sample data through a candidate frame generation network module.

[0046] S23. Input the candidate box and the two-dimensional image sample data into the boundary continuity-aware Gaussian encoder module to obtain a directed envelope box, and convert the directed envelope box into a two-dimensional Gaussian distribution.

[0047] S24. Input the two-dimensional Gaussian distribution into the decoder module to obtain the decoder module output.

[0048] S25. Design a loss function for the directed target detector, optimize the directed target detector according to the loss function, and obtain a constructed directed target detector.

[0049] Optionally, the encoding process of the boundary continuity-aware Gaussian encoder module is as shown in the following equation (1):

[0050] (1)

[0051] Subdividing formula (1), we get the following formula (2):

[0052] (2)

[0053] Where, represents the normalized two-dimensional Gaussian distribution of the feature space of the ground truth, represents the mean of the normalized two-dimensional Gaussian distribution of the feature space of the ground truth, The variance of the normalized two-dimensional Gaussian distribution representing the feature space of the ground truth, The set of vectors representing the ground truth values, represents the mean of the two-dimensional Gaussian distribution of the actual scale space of the ground truth, The variance of the two-dimensional Gaussian distribution representing the actual scale space of the ground truth, Represents the encoder encoding function that converts the two-dimensional Gaussian distribution of the actual scale space of the ground truth into the two-dimensional Gaussian distribution of the unitized feature space, represents the two-dimensional Gaussian distribution of the actual scale space, represents the candidate anchor box, represents the horizontal coordinate corresponding component in the ground truth vector group, Indicates the central horizontal coordinate of the directed envelope. represents the central horizontal coordinate of the ground truth directed envelope, represents the central horizontal coordinate of the candidate anchor box, The subscript anchor represents the anchor box. Indicates the width of the candidate anchor box, represents the ordinate corresponding component in the ground truth vector group, Indicates the central ordinate of the directed envelope. represents the central ordinate of the ground truth directed envelope, represents the central ordinate of the candidate anchor box, Indicates the height value of the candidate anchor box, represents the adjustable parameter for scaling, represents the scaling mapping function, Represents the three components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space, represents the scale-invariant adaptive factor component, Indicates the scale, represents the angle-dependent component, Represents the angle of the directed envelope, Indicates the first adjustable parameter, Represents the respective index numbers of the components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space.

[0054] Optionally, the decoding process of the decoder module is as shown in the following formula (3):

[0055] (3)

[0056] Subdividing formula (3), we get the following formula (4):

[0057] (4)

[0058] Where, represents the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the variance of the predicted two-dimensional Gaussian distribution of the actual scale space, Indicates the predicted subscript, Represents the decoding conversion function that transforms the predicted unit vectorized two-dimensional Gaussian distribution combined with the candidate anchor box into a two-dimensional Gaussian distribution in the actual scale space. Represents the predicted two-dimensional Gaussian distribution of the feature space unit, represents the candidate anchor box, Represents the abscissa component of the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the horizontal coordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unit, Indicates the actual scale width value of the candidate anchor box, Indicates the actual scale horizontal coordinate value of the candidate anchor box, Represents the abscissa component of the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the ordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unitization, Indicates the actual scale height value of the candidate anchor box, Indicates the actual scale ordinate value of the candidate anchor box, Represents the scaling mapping function The inverse function of Represents the output vector group of the prediction head.

[0059] Optionally, the input module is further configured to:

[0060] Convert the two-dimensional Gaussian distribution into a directed envelope as shown in the following equation (5):

[0061] (5)

[0062] Where, Represents the angle of the predicted directional envelope, Represents the predicted value, The first component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The second component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The third component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, Indicates the width corresponding to the equivalent OBB, , , Indicates the height corresponding to the equivalent OBB.

[0063] Optionally, the loss function of the oriented target detector is as follows (6):

[0064] (6)

[0065] Where, represents the loss function based on KL divergence, Represents the equivalent distance between the predicted Gaussian distribution based on KL divergence and the ground truth Gaussian distribution, represents the ground truth Gaussian distribution, represents the predicted Gaussian distribution, represents the loss function based on Bhattacharyya distance, Represents the equivalent distance between the predicted Gaussian distribution based on Bhattacharyya distance and the ground truth Gaussian distribution, represents the loss function based on Wasserstein distance, Represents the equivalent distance between the predicted Gaussian distribution and the ground truth Gaussian distribution based on the Wasserstein distance.

[0066] On the other hand, a target detection device is provided, which includes: a processor; a memory, wherein computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processor, any one of the target detection methods based on the boundary continuity-aware Gaussian encoder is implemented.

[0067] On the other hand, a computer-readable storage medium is provided, in which at least one instruction is stored. The at least one instruction is loaded and executed by a processor to implement any one of the above-mentioned target detection methods based on the boundary continuity-aware Gaussian encoder.

[0068] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0069] In an embodiment of the present invention, an encoder BCGE for boundary continuity is proposed, and the application of Gaussian distribution is extended to solve the boundary discontinuity problem. BCGE can be used as a plug-and-play module to generate candidate directed boxes that are not affected by angle values. In the absence of angle interference, BCGE can still output a high-precision transformation from the predicted Gaussian distribution to the actual two-dimensional rectangular area, providing a new paradigm for directed target detection. In order to utilize samples in the complex domain, the present invention optimizes the BCGE mechanism by extending the conversion theory from Gaussian distribution to actual two-dimensional directed boxes, and introduces corresponding loss function design strategies C2R and R2C. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0071] Figure 1 This is a flow chart of a target detection method based on a boundary continuity-aware Gaussian encoder provided by an embodiment of the present invention;

[0072] Figure 2 1 is a detection flow chart of a Gaussian encoder-decoder provided in an embodiment of the present invention;

[0073] Figure 3 is a diagram of a dynamic Gaussian loss function for a Gaussian encoder provided by an embodiment of the present invention;

[0074] Figure 4 This is a comparison diagram before and after the vector value provided by the embodiment of the present invention is processed by the Gaussian encoder;

[0075] Figure 5 Schematic diagram of the performance improvement brought about by the Gaussian encoder provided by an embodiment of the present invention;

[0076] Figure 6 Schematic diagram of the performance improvement brought about by the Gaussian encoder submodule provided in an embodiment of the present invention;

[0077] Figure 7 This is a block diagram of an object detection device based on a boundary continuity-aware Gaussian encoder provided by an embodiment of the present invention;

[0078] Figure 8 It is a structural diagram of a target detection device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0079] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0080] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0081] In the embodiments of the present invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same. The terms "of," "corresponding," and "corresponding" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same.

[0082] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0083] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0084] The embodiment of the present invention provides a target detection method based on a boundary continuity-aware Gaussian encoder. The method can be implemented by a target detection device, which can be a terminal or a server. Figure 1 The flow chart of the target detection method based on the boundary continuity perception Gaussian encoder is shown. The processing flow of the method may include the following steps:

[0085] S1. Acquire two-dimensional image data for directed target detection.

[0086] S2. Input the two-dimensional image data into the constructed oriented target detector.

[0087] Among them, the directed target detector can include a boundary continuity-aware Gaussian encoder module, a decoder, etc.

[0088] S3. According to the two-dimensional image data and the boundary continuity perception Gaussian encoder module, a periodically continuous two-dimensional Gaussian distribution is obtained, and a two-dimensional directional detection frame is obtained according to the two-dimensional Gaussian distribution.

[0089] Optionally, the construction process of the directional object detector in S2 may include the following steps S21-S25:

[0090] S21. Obtain two-dimensional image sample data.

[0091] S22. Obtain a candidate frame in the two-dimensional image sample data through a candidate frame generation network module.

[0092] S23. Input the candidate box and the two-dimensional image sample data into the boundary continuity-aware Gaussian encoder module to obtain a directed envelope box, and convert the directed envelope box into a two-dimensional Gaussian distribution.

[0093] In a feasible implementation mode, the present invention represents five parameters: Convert to Gaussian representation and design an encoder specifically adapted to Gaussian distribution representation, which includes value range conversion strategies (scaling and mapping) to improve network learning and prevent gradient explosion. Figure 2 shown.

[0094] S24. Input the two-dimensional Gaussian distribution into the decoder module to obtain the decoder module output.

[0095] In a feasible implementation manner, in the decoder part, the present invention converts the prediction vector output by the prediction head into a Gaussian distribution format.

[0096] S25. Design a loss function for the directed target detector, optimize the directed target detector according to the loss function, and obtain a constructed directed target detector.

[0097] Specifically, in the generation part, the present invention converts the Gaussian representation back into a directed box, breaking through the limitation that previous methods can only convert from a box to a Gaussian representation in one direction. In addition, since some results (prediction boxes) may fall into the complex domain, the present invention further extends the generation process to the complex domain. The present invention proposes the generation of the output envelope box from the complex domain to the real domain (C2R), so that the samples of the envelope box in the complex domain are effectively utilized. Finally, the present invention further proposes a strategy for converting prediction boxes in the complex domain to ensure that they are aligned in the same domain and more closely follow the Gaussian calculation principle, which is called the real domain to complex domain loss strategy (R2C). This method eliminates the case with negative values ​​in the loss function. It is worth noting that the method of the present invention is compatible with any Gaussian-based loss function.

[0098] In the forward process of the entire network, the present invention abandons the stereotyped use of angles, thereby completely reshaping the directed target detection problem into a vector-matrix regression problem, introducing a new paradigm for directed target detection.

[0099] The BCGE model can be used to learn rectangular regions as a differentiable two-dimensional Gaussian distribution matrix. By estimating the mean and variance of the target Gaussian distribution, the entire training process of each envelope box can be better supervised. The design strategy aims to minimize the error distance between the mean and variance matrices of the predicted box and the ground truth box to reconstruct the output two-dimensional oriented detection box. In previous research GWD, GWD used matrix transformation to construct the Gaussian distribution of the OBB, thus giving a new representation paradigm for describing oriented rectangles:

[0100] For any OBB , the corresponding Gaussian distribution The parameters can be calculated as follows:

[0101] (1)

[0102] Where, represents the variance of any two-dimensional Gaussian distribution, represents the mean of any two-dimensional Gaussian distribution, represents the first component of the variance of an arbitrary two-dimensional Gaussian distribution, represents the second component of the variance of an arbitrary two-dimensional Gaussian distribution, Represents the third component of the variance of an arbitrary two-dimensional Gaussian distribution.

[0103] Specifically, Figure 2 in It can be broken down into the following forms:

[0104] (2)

[0105] Where, represents the normalized two-dimensional Gaussian distribution of the feature space of the ground truth, represents the mean of the normalized two-dimensional Gaussian distribution of the feature space of the ground truth, The variance of the normalized two-dimensional Gaussian distribution representing the feature space of the ground truth, The set of vectors representing the ground truth values, represents the mean of the two-dimensional Gaussian distribution of the actual scale space of the ground truth, The variance of the two-dimensional Gaussian distribution representing the actual scale space of the ground truth, Represents the encoder encoding function that converts the two-dimensional Gaussian distribution of the actual scale space of the ground truth into the two-dimensional Gaussian distribution of the unitized feature space, represents the two-dimensional Gaussian distribution of the actual scale space, represents the candidate anchor box, represents the horizontal coordinate corresponding component in the ground truth vector group, Indicates the central horizontal coordinate of the directed envelope. represents the central horizontal coordinate of the ground truth directed envelope, represents the central horizontal coordinate of the candidate anchor box, The subscript anchor represents the anchor box. Indicates the width of the candidate anchor box, represents the ordinate corresponding component in the ground truth vector group, Indicates the central ordinate of the directed envelope. represents the central ordinate of the ground truth directed envelope, represents the central ordinate of the candidate anchor box, Indicates the height value of the candidate anchor box, represents the scale adjustment parameter, represents the scaling mapping function, Represents the three components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space, represents the scale-invariant adaptive factor component, Indicates the scale, represents the angle-dependent component, Represents the angle of the directed envelope, Indicates the first adjustable parameter, Represents the respective index numbers of the components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space.

[0106] In order to obtain the Gaussian distribution of the mean and variance of the predicted two-dimensional Gaussian matrix with respect to the actual scale of the prediction, Figure 2 The decoding process in The formula can be broken down into the following forms:

[0107] (3)

[0108] Where, represents the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the variance of the predicted two-dimensional Gaussian distribution of the actual scale space, Indicates the predicted subscript, Represents the decoding conversion function that transforms the predicted unit vectorized two-dimensional Gaussian distribution combined with the candidate anchor box into a two-dimensional Gaussian distribution in the actual scale space. Represents the predicted two-dimensional Gaussian distribution of the feature space unit, represents the candidate anchor box, Represents the abscissa component of the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the horizontal coordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unit, Indicates the actual scale width value of the candidate anchor box, Indicates the actual scale horizontal coordinate value of the candidate anchor box, Represents the abscissa component of the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the ordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unitization, Indicates the actual scale height value of the candidate anchor box, Indicates the actual scale ordinate value of the candidate anchor box, Represents the scaling mapping function The inverse function of Represents the output vector group of the prediction head.

[0109] Compared to traditional methods, the analysis method of the present invention directly estimates the exact state of the Gaussian distribution in a single analysis operation, bypassing any intermediate procedures. The method of the present invention does not rely on the singular value decomposition (SVD) algorithm to recover the matrix from the Gaussian distribution matrix to the actual two-dimensional detection box, as the SVD algorithm cannot effectively handle unreasonable prediction quantities that produce negative values. The present invention chooses to use the analytical solution of the reversible equation to perform the envelope generation process.

[0110] Based on the forward conversion relationship, the present invention considers the solution equation of the inverse conversion from the Gaussian distribution matrix to the OBB. First, based on the analysis of the coupling amount in the Gaussian distribution matrix, the formula of the extracted factor is as follows:

[0111] (4)

[0112] in and are the width and height of the equivalent OBB, which are migrated from the Gaussian distribution of the actual size predicted by the prediction head. is m, is n. The final rectangular frame generated is as follows:

[0113] (5)

[0114] Where, Represents the angle of the predicted directional envelope, Represents the predicted value, The first component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The second component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The third component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, , .

[0115] Because the 、 and are three independent outputs of the convolutional neural network, and their operational combination is not constrained by the network alone. Once the decision value of the prediction matrix becomes negative, the generated matrix is ​​no longer in the representation format of the Gaussian probability distribution matrix. In this case, and will be decoded as a negative value, so the parameter values ​​w and h obtained by solving the equation will fall into the complex domain. Therefore, the present invention extends the theory to the complex domain, making the detection box conversion generation process of the present invention effective in both the real and complex domains.

[0116] The present invention first assumes that for , for , where i is the imaginary unit. Under this condition, the conversion formula can be expanded as follows:

[0117] (6)

[0118] Among them, the clear Gaussian distribution in the real number domain can be mapped to an infinite number of corresponding complex number arrays through the complex probability distribution. This constitutes a one-way one-to-many process. The goal of this invention is to find any valid complex number solution from the infinite possibilities and ensure that it is uniquely mapped to the corresponding real number solution. In view of the previous proof of this invention, and Can only be decoded as pure positive or pure negative, which means that their imaginary part is zero. Under this premise, the present invention derives the following formula:

[0119] (7)

[0120] Considering that any complex number solution is a valid solution, the present invention selects a very special scenario, namely and , where k is an adjustable parameter (set to 0.5 here), and m and n are set the same as in the previous equation. In this scenario, the determinant of the complex Gaussian probability diffusion matrix derived from the original complex probabilities can be purely negative, without a complex part. This is because the corresponding prediction head predicts w2 and h2 as purely negative. The process involving the mixed domain can be further described as follows:

[0121] (8)

[0122] The present invention can thus convert it from the complex number domain to the real number domain and obtain and The process of mapping the predicted Gaussian distribution from the complex domain back to a box in the real domain is called C2R transformation.

[0123] Because the output is encoded into a vector space, the output of the proposed BCGE can be directly constrained by choosing a loss function, just like any previous encoder. The present invention first uses a smooth L1 loss to measure the distance between the ground-truth vector and the encoded vector of the predicted box to help narrow their differences in the vector space, as shown below:

[0124] (9)

[0125] Where, Represents the value of the total classification loss function in a batch, Indicates the total number of samples (how many candidate boxes), A function representing the chosen classification loss function, represents the i-th prediction box, represents the ground truth box corresponding to the i-th predicted box, Represents the value of the total box positioning loss function in a batch, A function representing the chosen classification loss function.

[0126] In addition to the L1 loss, considering that the output of the prediction head of the present invention is a Gaussian distribution, in order to reduce the error with the actual scale Gaussian distribution of the ground truth, the present invention still utilizes the geometric characteristics exhibited by the two-dimensional variance, takes the Gaussian distance metric as part of the loss function, and designs a dynamic adaptive balanced Gaussian loss function.

[0127] The present invention selects three loss functions based on Gaussian distance to measure the distance between the Gaussian distribution of the ground truth and the predicted Gaussian distribution by the difference in statistical probability.

[0128] Kullback-Leibler Divergence (KLD):

[0129] (10)

[0130] Where, represents the trace of the matrix.

[0131] Bhattacharyya distancex(BD):

[0132] (11)

[0133] Wasserstein Distance(WD):

[0134] (12)

[0135] (13)

[0136] (14)

[0137] (15)

[0138] Where, represents the trace of the matrix.

[0139] However, considering the conditions of the complex domain, in some cases, the traditional definition of matrices w and h is complex, making w2 and h2 positive numbers, which will be regarded as illegal calculations when calculating Gaussian-based loss values, thereby discarding a large number of valid learning samples.

[0140] like Figure 3 As shown on the right, since the predicted values ​​w and h and their corresponding imaginary and real parts are generated in four quadrants, the Gaussian encoder training process produces three pseudo-GTs, corresponding to four equivalent results. The training process adaptively selects the GT Gaussian distribution corresponding to the optimization objective in the same domain and calculates the minimum loss value, namely the real-to-complex Gaussian loss strategy (R2C).

[0141] Finally, the loss designed to constrain the shapes of the two probability distributions is defined in real space and vector space as follows:

[0142] (16)

[0143] Where, represents the loss function based on KL divergence, Represents the equivalent distance between the predicted Gaussian distribution based on KL divergence and the ground truth Gaussian distribution, represents the ground truth Gaussian distribution, represents the predicted Gaussian distribution, represents the loss function based on Bhattacharyya distance, Represents the equivalent distance between the predicted Gaussian distribution based on Bhattacharyya distance and the ground truth Gaussian distribution, represents the loss function based on Wasserstein distance, Represents the equivalent distance between the predicted Gaussian distribution and the ground truth Gaussian distribution based on the Wasserstein distance.

[0144] The present invention proposes a boundary continuity-aware Gaussian encoder (BCGE) for oriented object detectors in two-dimensional images to eliminate boundary discontinuity problems while accurately measuring distances for oriented object detection tasks.

[0145] Specifically, BCGE directly predicts the target 2D Gaussian distribution of the prediction box and learns the oriented bounding box as an integrated 2D matrix, effectively addressing boundary discontinuities. The present invention also proposes a conversion theory from the 2D Gaussian representation back to the box and extends this conversion theory to the complex domain to accommodate the learning characteristics of neural networks. Furthermore, BCGE is a versatile plug-and-play architecture encoder that can directly replace the standard encoding process in various oriented detectors and is adaptable. Experimental results on four popular datasets (DOTA, UCAS-AOD, HRSC2016, and HRSID) consistently demonstrate the effectiveness of the present invention.

[0146] like Figure 4-Figure 6 As shown in Figure 3, the proposed method is evaluated on several public large-scale oriented object detection datasets, including DOTA, HRSC2016, UCAS-AOD, and HRSID. The proposed BCGE achieves significant performance improvements over many methods.

[0147] In an embodiment of the present invention, an encoder BCGE for boundary continuity is proposed, and the application of Gaussian distribution is extended to solve the boundary discontinuity problem. BCGE can be used as a plug-and-play module to generate candidate directed boxes that are not affected by angle values. In the absence of angle interference, BCGE can still output a high-precision transformation from the predicted Gaussian distribution to the actual two-dimensional rectangular area, providing a new paradigm for directed target detection. In order to utilize samples in the complex domain, the present invention optimizes the BCGE mechanism by extending the conversion theory from Gaussian distribution to actual two-dimensional directed boxes, and introduces corresponding loss function design strategies C2R and R2C.

[0148] Figure 7 This is a block diagram of a target detection device based on a boundary continuity-aware Gaussian encoder according to an exemplary embodiment. The device is used in a target detection method based on a boundary continuity-aware Gaussian encoder. Figure 7 The device includes an acquisition module 310, an input module 320, and an output module 330.

[0149] The acquisition module 310 is used to acquire two-dimensional image data to be used for oriented target detection.

[0150] The input module 320 is used to input the two-dimensional image data into the constructed oriented object detector; wherein the oriented object detector includes a boundary continuity-aware Gaussian encoder module.

[0151] The output module 330 is used to obtain a periodically continuous two-dimensional Gaussian distribution based on the two-dimensional image data and the boundary continuity perception Gaussian encoder module, and obtain a two-dimensional directional detection frame based on the two-dimensional Gaussian distribution.

[0152] Optionally, the input module 320 is further configured to:

[0153] S21. Obtain two-dimensional image sample data.

[0154] S22. Obtain a candidate frame in the two-dimensional image sample data through a candidate frame generation network module.

[0155] S23. Input the candidate box and the two-dimensional image sample data into the boundary continuity-aware Gaussian encoder module to obtain a directed envelope box, and convert the directed envelope box into a two-dimensional Gaussian distribution.

[0156] S24. Input the two-dimensional Gaussian distribution into the decoder module to obtain the decoder module output.

[0157] S25. Design a loss function for the directed target detector, optimize the directed target detector according to the loss function, and obtain a constructed directed target detector.

[0158] Optionally, the encoding process of the boundary continuity-aware Gaussian encoder module is as shown in the following equation (1):

[0159] (1)

[0160] Subdividing formula (1), we get the following formula (2):

[0161] (2)

[0162] Where, represents the normalized two-dimensional Gaussian distribution of the feature space of the ground truth, represents the mean of the normalized two-dimensional Gaussian distribution of the feature space of the ground truth, The variance of the normalized two-dimensional Gaussian distribution representing the feature space of the ground truth, The set of vectors representing the ground truth values, represents the mean of the two-dimensional Gaussian distribution of the actual scale space of the ground truth, The variance of the two-dimensional Gaussian distribution representing the actual scale space of the ground truth, Represents the encoder encoding function that converts the two-dimensional Gaussian distribution of the actual scale space of the ground truth into the two-dimensional Gaussian distribution of the unitized feature space, represents the two-dimensional Gaussian distribution of the actual scale space, represents the candidate anchor box, represents the horizontal coordinate corresponding component in the ground truth vector group, Indicates the central horizontal coordinate of the directed envelope. represents the central horizontal coordinate of the ground truth directed envelope, represents the central horizontal coordinate of the candidate anchor box, The subscript anchor represents the anchor box. Indicates the width of the candidate anchor box, represents the ordinate corresponding component in the ground truth vector group, Indicates the central ordinate of the directed envelope. represents the central ordinate of the ground truth directed envelope, represents the central ordinate of the candidate anchor box, Indicates the height value of the candidate anchor box, represents the scale adjustment parameter, represents the scaling mapping function, Represents the three components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space, represents the scale-invariant adaptive factor component, Indicates the scale, represents the angle-dependent component, Represents the angle of the directed envelope, Indicates the first adjustable parameter, Represents the respective index numbers of the components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space.

[0163] Optionally, the decoding process of the decoder module is as shown in the following formula (3):

[0164] (3)

[0165] Subdividing formula (3), we get the following formula (4):

[0166] (4)

[0167] Where, represents the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the variance of the predicted two-dimensional Gaussian distribution of the actual scale space, Indicates the predicted subscript, Represents the decoding conversion function that transforms the predicted unit vectorized two-dimensional Gaussian distribution combined with the candidate anchor box into a two-dimensional Gaussian distribution in the actual scale space. Represents the predicted two-dimensional Gaussian distribution of the feature space unit, represents the candidate anchor box, Represents the abscissa component of the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the horizontal coordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unit, Indicates the actual scale width value of the candidate anchor box, Indicates the actual scale horizontal coordinate value of the candidate anchor box, Represents the abscissa component of the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, Represents the ordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unitization, Indicates the actual scale height value of the candidate anchor box, Indicates the actual scale ordinate value of the candidate anchor box, Represents the scaling mapping function The inverse function of Represents the output vector group of the prediction head.

[0168] Optionally, the input module 320 is further configured to:

[0169] Convert the two-dimensional Gaussian distribution into a directed envelope as shown in the following equation (5):

[0170] (5)

[0171] Where, Represents the angle of the predicted directional envelope, Represents the predicted value, The first component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The second component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The third component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, Indicates the width corresponding to the equivalent OBB, , , Indicates the height corresponding to the equivalent OBB.

[0172] Optionally, the loss function of the oriented target detector is as follows (6):

[0173] (6)

[0174] Where, represents the loss function based on KL divergence, Represents the equivalent distance between the predicted Gaussian distribution based on KL divergence and the ground truth Gaussian distribution, represents the ground truth Gaussian distribution, represents the predicted Gaussian distribution, represents the loss function based on Bhattacharyya distance, Represents the equivalent distance between the predicted Gaussian distribution based on Bhattacharyya distance and the ground truth Gaussian distribution, represents the loss function based on Wasserstein distance, Represents the equivalent distance between the predicted Gaussian distribution and the ground truth Gaussian distribution based on the Wasserstein distance.

[0175] In an embodiment of the present invention, an encoder BCGE for boundary continuity is proposed, and the application of Gaussian distribution is extended to solve the boundary discontinuity problem. BCGE can be used as a plug-and-play module to generate candidate directed boxes that are not affected by angle values. In the absence of angle interference, BCGE can still output a high-precision transformation from the predicted Gaussian distribution to the actual two-dimensional rectangular area, providing a new paradigm for directed target detection. In order to utilize samples in the complex domain, the present invention optimizes the BCGE mechanism by extending the conversion theory from Gaussian distribution to actual two-dimensional directed boxes, and introduces corresponding loss function design strategies C2R and R2C.

[0176] Figure 8 is a schematic structural diagram of a target detection device provided by an embodiment of the present invention, such as Figure 8 As shown, the target detection device may include the above Figure 7 Optionally, the object detection device 410 may include a first processor 2001 .

[0177] Optionally, the target detection device 410 may further include a memory 2002 and a transceiver 2003 .

[0178] The first processor 2001, the memory 2002 and the transceiver 2003 may be connected via a communication bus.

[0179] The following combination Figure 8 The components of the target detection device 410 are described in detail:

[0180] The first processor 2001 is the control center of the target detection device 410 and can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs) or one or more field programmable gate arrays (FPGAs).

[0181] Optionally, the first processor 2001 may execute various functions of the target detection device 410 by running or executing a software program stored in the memory 2002 and calling data stored in the memory 2002 .

[0182] In a specific implementation, as an embodiment, the first processor 2001 may include one or more CPUs, such as Figure 8 CPU0 and CPU1 are shown in FIG.

[0183] In a specific implementation, as an embodiment, the target detection device 410 may also include multiple processors, such as Figure 8 1 and 2. The first processor 2001 and the second processor 2004 are shown in FIG. Each of these processors can be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). A processor herein can refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0184] The memory 2002 is used to store the software program for executing the solution of the present invention, and is controlled by the first processor 2001 for execution. The specific implementation method can refer to the above method embodiment and will not be repeated here.

[0185] Alternatively, the memory 2002 may be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 may be integrated with the first processor 2001 or exist independently and accessed through the interface circuit ( Figure 8 (not shown) is coupled to the first processor 2001, which is not specifically limited in this embodiment of the present invention.

[0186] The transceiver 2003 is used to communicate with a network device or a terminal device.

[0187] Optionally, the transceiver 2003 may include a receiver and a transmitter ( Figure 8 The receiver is used to implement a receiving function, and the transmitter is used to implement a sending function.

[0188] Optionally, the transceiver 2003 may be integrated with the first processor 2001, or may exist independently and communicate with the target detection device 410 through an interface circuit ( Figure 8 (not shown) is coupled to the first processor 2001, which is not specifically limited in this embodiment of the present invention.

[0189] It should be noted that Figure 8 The structure of the target detection device 410 shown in the figure does not constitute a limitation on the router. The actual knowledge structure recognition device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0190] In addition, the technical effects of the target detection device 410 can refer to the technical effects of the target detection method based on the boundary continuity-aware Gaussian encoder described in the above method embodiment, and will not be repeated here.

[0191] It should be understood that the first processor 2001 in the embodiment of the present invention may be a central processing unit (CPU), or may be 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. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, etc.

[0192] It should also be understood that the memory in the embodiments of the present invention may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0193] The above embodiments can be implemented in whole or in part via software, hardware (e.g., circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in accordance with the embodiments of the present invention are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0194] It should be understood that the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. A and B can be singular or plural. Furthermore, the character " / " as used herein generally indicates an "or" relationship between the associated objects, but it may also indicate an "and / or" relationship. For specific understanding, please refer to the context.

[0195] In this disclosure, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, "at least one of a, b, or c" can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or plural.

[0196] It should be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes 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 the present invention.

[0197] 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. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0198] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0199] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interface, indirect coupling or communication connection of the device or unit, which can be electrical, mechanical or other forms.

[0200] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0201] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0202] If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical disks.

[0203] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A target detection method based on boundary continuity-aware Gaussian encoder, characterized in that: The method comprises: S1. Acquire two-dimensional image data for directed target detection; S2. Inputting the two-dimensional image data into a constructed oriented object detector; wherein the oriented object detector includes a boundary continuity-aware Gaussian encoder module; S3. Obtaining a periodically continuous two-dimensional Gaussian distribution according to the two-dimensional image data and the boundary continuity-aware Gaussian encoder module, and obtaining a two-dimensional directional detection frame according to the two-dimensional Gaussian distribution; The construction process of the directed target detector in S2 includes: S21, obtaining two-dimensional image sample data; S22, obtaining a candidate frame in the two-dimensional image sample data through a candidate frame generation network module; S23, inputting the candidate box and the two-dimensional image sample data into a boundary continuity-aware Gaussian encoder module to obtain a directed envelope box, and converting the directed envelope box into a two-dimensional Gaussian distribution; S24, inputting the two-dimensional Gaussian distribution into a decoder module to obtain a decoder module output; S25. Design a loss function for the directed target detector, and optimize the directed target detector according to the loss function to obtain a constructed directed target detector.

2. The target detection method based on boundary continuity-aware Gaussian encoder according to claim 1, characterized in that: The encoding process of the boundary continuity-aware Gaussian encoder module in S23 is shown in the following formula (1): Subdividing formula (1) into the following formula (2): Where, Represents the two-dimensional Gaussian distribution of the ground truth feature space unit, μ t ′ represents the mean of the two-dimensional Gaussian distribution of the ground truth feature space, Σ t ′ represents the variance of the two-dimensional Gaussian distribution of the ground truth feature space, t represents the vector group of the ground truth, μ t The mean of the two-dimensional Gaussian distribution of the actual scale space representing the ground truth, Σ t The variance of the two-dimensional Gaussian distribution representing the actual scale space of the ground truth, Represents the encoder encoding function that converts the two-dimensional Gaussian distribution of the actual scale space of the ground truth into the two-dimensional Gaussian distribution of the unitized feature space, Represents the two-dimensional Gaussian distribution of the actual scale space, Anchor represents the candidate anchor box, t x Represents the horizontal coordinate corresponding component in the ground truth vector group, x represents the central horizontal coordinate of the directed envelope, x t Represents the central horizontal coordinate of the ground truth directed envelope, x a Represents the central horizontal coordinate of the candidate anchor box, a represents the subscript anchor of the anchor box, w a Indicates the width of the candidate anchor box, t y Represents the ordinate corresponding component in the ground truth vector group, y represents the central ordinate of the directed envelope, y t Represents the central ordinate of the ground truth directed envelope, y a Indicates the central ordinate of the candidate anchor box, h a represents the height value of the candidate anchor box, α represents the scale scaling adjustable parameter, represents the scaling mapping function, The three components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space, Ψ s Represents the scale-invariant adaptive factor component, s represents the scale, Φ θ represents the angle-related component, θ represents the angle of the directional envelope, ε i Represents the first adjustable parameter, and i represents the respective index numbers of the components of the variance matrix of the two-dimensional Gaussian distribution in the actual scale space.

3. The target detection method based on boundary continuity-aware Gaussian encoder according to claim 2, characterized in that: The decoding process of the decoder module in S24 is shown in the following formula (3): Subdividing formula (3) into the following formula (4): Where, Represents the predicted two-dimensional Gaussian distribution of the actual scale space, μ p Represents the mean of the predicted two-dimensional Gaussian distribution of the actual scale space, ∑ p Represents the variance of the predicted two-dimensional Gaussian distribution of the actual scale space, and p represents the subscript of the prediction. Represents the decoding conversion function that transforms the predicted unit vectorized two-dimensional Gaussian distribution combined with the candidate anchor box into a two-dimensional Gaussian distribution in the actual scale space. Represents the predicted two-dimensional Gaussian distribution of the feature space unit, Anchor represents the candidate anchor box, x p Represents the abscissa component of the mean of the two-dimensional Gaussian distribution of the predicted actual scale space, δ x Represents the horizontal coordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unit, w a Indicates the actual scale width value of the candidate anchor box, x a Indicates the actual scale horizontal coordinate value of the candidate anchor box, y p Represents the abscissa component of the mean of the two-dimensional Gaussian distribution of the predicted actual scale space, δ y Represents the ordinate component of the mean of the two-dimensional Gaussian distribution of the predicted feature space unit, h a Indicates the actual scale height value of the candidate anchor box, y a Indicates the actual scale ordinate value of the candidate anchor box, represents the inverse function of the scaling mapping function S, δ i Represents the output vector group of the prediction head.

4. The object detection method based on boundary continuity-aware Gaussian encoder according to claim 1, characterized in that The construction process of the directed target detector further includes: Convert the two-dimensional Gaussian distribution into a directed envelope as shown in the following formula (5): Where θ p Represents the angle of the predicted directional envelope, p represents the predicted value, The first component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The second component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, The third component of the variance of a two-dimensional Gaussian distribution representing the actual scale of the predicted quantity, w p Indicates the width corresponding to the equivalent OBB, h p Indicates the height corresponding to the equivalent OBB.

5. The object detection method based on boundary continuity-aware Gaussian encoder according to claim 1, characterized in that The loss function of the directional target detector in S25 is shown in the following formula (6): Where, Denotes the loss function based on KL divergence, D K Represents the equivalent distance between the predicted Gaussian distribution based on KL divergence and the ground truth Gaussian distribution, represents the ground truth Gaussian distribution, represents the predicted Gaussian distribution, Denotes the loss function based on Bhattacharyya distance, D B Represents the equivalent distance between the predicted Gaussian distribution based on Bhattacharyya distance and the ground truth Gaussian distribution, Denotes the loss function based on Wasserstein distance, D W Represents the equivalent distance between the predicted Gaussian distribution and the ground truth Gaussian distribution based on the Wasserstein distance.

6. A target detection device based on a boundary continuity-aware Gaussian encoder, wherein the target detection device based on a boundary continuity-aware Gaussian encoder is used to implement the target detection method based on a boundary continuity-aware Gaussian encoder according to any one of claims 1 to 5, characterized in that: The device comprises: An acquisition module, used to acquire two-dimensional image data to be used for oriented target detection; An input module, configured to input the two-dimensional image data into a constructed oriented object detector; wherein the oriented object detector includes a boundary continuity-aware Gaussian encoder module; An output module, configured to obtain a periodically continuous two-dimensional Gaussian distribution based on the two-dimensional image data and the boundary continuity perception Gaussian encoder module, and obtain a two-dimensional directional detection frame based on the two-dimensional Gaussian distribution; The construction process of the directed target detector includes: S21, obtaining two-dimensional image sample data; S22, obtaining a candidate frame in the two-dimensional image sample data through a candidate frame generation network module; S23, inputting the candidate box and the two-dimensional image sample data into a boundary continuity-aware Gaussian encoder module to obtain a directed envelope box, and converting the directed envelope box into a two-dimensional Gaussian distribution; S24, inputting the two-dimensional Gaussian distribution into a decoder module to obtain a decoder module output; S25. Design a loss function for the directed target detector, and optimize the directed target detector according to the loss function to obtain a constructed directed target detector.

7. A target detection device, characterized in that: The target detection device comprises: processor; A memory having computer-readable instructions stored thereon, wherein when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program code, which can be called by a processor to execute the method according to any one of claims 1 to 5.

Images