An ellipse distortion perception-based target detection model and method

Abstract

The present application relates to the field of vision-based target image detection, in particular to an ellipse distortion perception-based target detection model and method. The target detection model is a model structure formed based on improvement of a YOLOv11 model, and the improvement at least includes a distortion perception detection head, a circular convolution dynamic ring-shaped sampling structure, a circular parameter modeling method, and a circular improved distance intersection over union loss function with fusion of physical consistency constraints. The target detection method uses the target detection model to detect a target placement image, outputs target categories, structure parameters and confidence information, and judges the target placement state accordingly.

Description

Technical Field

[0001] This invention relates to the field of vision-based target image detection, and more particularly to a target detection model and method based on elliptic distortion perception. Background Technology

[0002] With the rapid development of artificial intelligence technology, machine vision inspection technology is increasingly widely used in industrial production. YOLO series object detection algorithms based on deep learning can efficiently handle object recognition tasks in industrial scenarios. However, existing general-purpose inspection technologies still suffer from problems in aluminum alloy die-casting production lines, such as insufficient modeling of filter screen elliptical distortion features, lack of physical constraints leading to positioning errors, and weak anti-interference capabilities in complex industrial environments. These issues result in inspection accuracy and robustness failing to meet production requirements.

[0003] Therefore, there is an urgent need for a method that combines elliptic distortion perception with physical consistency constraints to achieve high-precision detection and stable judgment of the filter screen placement status in aluminum alloy die-casting production lines. Summary of the Invention

[0004] One of the objectives of this invention is to provide a target detection model that improves detection accuracy and stability by using the detection model.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a target detection model, based on an improvement to the YOLOv11 model, comprising: The distortion-aware detection head constructs a dynamic ring sampling structure based on circular convolution. It generates a ring sampling point set at the center of each pixel in the feature map and uses offset prediction convolution to learn the two-dimensional position offset of the sampling points. This enables the sampling structure to adapt to the target shape and thus enhances the model's response to elliptical distortion boundary structures.

[0006] Among them, circular convolution refers to a convolutional structure that uses the ideal circular outline of the target as a geometric prior, adjusts the regular sampling method of the traditional rectangular convolution kernel to a ring-shaped reference sampling method, and uses a learnable two-dimensional offset to make the ring-shaped sampling points adaptively deform along the target boundary. For near-circular targets such as aluminum alloy die-cast filter screens, the target is usually projected from an ideal circle to an ellipse when the target is placed skewed or the viewing angle changes; therefore, circular convolution first covers the key boundary region near the circular outline, and then learns the boundary changes in the major and minor axes of the ellipse through offset learning, thus being able to extract circular or elliptical boundary features more concentratedly than ordinary rectangular convolution kernels.

[0007] Furthermore, the dynamic ring sampling structure includes: constructing a set of sampling points that satisfy the ring constraint conditions within a preset radius; constraining the predicted offset by batch normalization and nonlinear activation functions; obtaining the feature response of the deformed sampling position based on bilinear interpolation; and completing the feature enhancement expression by aggregating the ring sampling features.

[0008] Furthermore, the target detection model uses a circular parameter modeling method to geometrically represent the target, representing the target as the coordinates of the center of a circle and the radius parameter, and then restores it to the original image space through scale mapping, thus achieving a geometric modeling method consistent with the ring perception structure.

[0009] Furthermore, during the training phase, the target detection model employs a circular-like improved distance intersection-union ratio (DIU) loss function that incorporates physical consistency constraints to optimize the prediction results. Based on the circular DIU calculation, the loss function introduces a center distance penalty term and a radius difference penalty term to construct geometric consistency constraints, and combines the elliptical axis ratio physical parameter to construct a physical consistency constraint term, so that the prediction results meet the physical structural rationality requirements while achieving geometric matching.

[0010] Furthermore, the improved circular distance intersection-union ratio loss function includes an internal region intersection-union ratio constraint mechanism and an interval modulation mechanism, which are used to enhance the optimization capability of the matching degree of the target center region and to perform differentiated optimization for regions with different degrees of overlap, so as to improve training stability.

[0011] Specifically, the internal region intersection-union ratio (IUGR) constraint mechanism involves introducing a preset scaling factor based on the original radii of the predicted and actual circles to obtain the corresponding inner circular regions, and calculating the IUGR between these inner circular regions. This inner IUGR, together with the overall circular IUGR, constrains the model, ensuring that the model considers not only whether the outer contours overlap but also whether the target center regions are aligned. The interval modulation mechanism involves dividing the samples into low-overlap, medium-overlap, and high-overlap intervals according to the IUGR, and applying nonlinear compression or weight adjustment to the loss contribution of different intervals to avoid extreme samples dominating the training process.

[0012] Furthermore, the bounding box regression branch of the original YOLOv11 detector head is replaced with a circle-like module, which pre-generates a ring-shaped reference sampling coordinate with a radius of 2 to simulate the ideal circular contour of the filter.

[0013] Another objective of this invention is to provide a target detection method based on elliptic distortion perception, which can improve detection accuracy and stability.

[0014] To achieve the above objectives, the present invention provides the following technical solution: a target detection method based on elliptic distortion perception, comprising acquiring an image of the placement of a filter screen on an aluminum alloy component; The target image is input into the target detection model in the above scheme, and the target category, structural parameters and confidence information are output. Furthermore, this also includes extracting elliptical distortion physical parameters; where the target segmentation mask refers to a binary mask obtained by manual annotation or segmentation algorithms to represent the pixel region where the target is located. By statistically analyzing the spatial distribution of target pixels in this mask, the circumscribed contour, center position, principal axis direction, secondary axis direction, principal axis length, and secondary axis length of the target region are obtained, and the ratio of the principal axis to the secondary axis is calculated as a parameter characterizing the degree of elliptical distortion, used to reflect the degree of structural stretching or compression of the target during the imaging process. Furthermore, by introducing the inner circle IoU, the perception of distortion in the central region of the ellipse is enhanced; Calculate the geometric penalty terms, including the center distance penalty ((cxgt-cxpr)2+(cygt-cypr)2) and the radius error penalty ((rgt-rpr)2), and normalize them to the image size range.

[0015] Furthermore, bilinear interpolation sampling is performed based on the deformed coordinates to obtain key feature points on the elliptical contour. After feature aggregation through grouped convolution, the basic elliptical distortion features are output after GroupNorm normalization and SiLU activation.

[0016] in, For deformation sampling coordinates The four nearest integer coordinates, This is the decimal offset of the coordinates. These are the characteristic values ​​of the corresponding coordinates.

[0017] In summary, the present invention has the following advantages compared with the prior art: 1. This invention introduces a dynamic ring sampling structure based on circular convolution, which can adaptively match the distorted filter contour in actual production, solves the problem of insufficient modeling of elliptical distortion features by conventional detection models, and improves the model's feature extraction capability for target distortion boundaries. 2. A geometric modeling approach combining circular parameters with elliptical physical parameters is adopted, along with an improved loss function that incorporates physical consistency constraints. This introduces explicit physical structural constraints into the model training, effectively reducing positioning bias and improving the rationality of the prediction results. The overall improvement is based on the YOLOv11 model structure. Only the detection head and loss function are modified, without the need for major adjustments to the overall network architecture. The inference speed meets the real-time detection requirements of industrial production lines. For the detection task of aluminum alloy die-casting filter screen in complex industrial environments, it has higher detection accuracy and stronger robustness, and can meet the detection needs of actual production. Attached image description: Figure 1 Here is a flowchart of the method in this embodiment. Figure 2 Here is a schematic diagram of a circular convolution-like dynamic ring sampling structure. Figure 3 Here is a schematic diagram illustrating the principle of the elliptic distortion-perceived consistency loss function. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in 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 may be executed in a different order than that shown here.

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] like Figures 1-3 As shown in Example 1, an object detection method based on elliptic distortion perception includes an object detection model construction process and an object detection detection process using the object detection model. The object detection model construction process is a structural improvement on the YOLOv11 model, mainly including a distortion-aware detection head, a circular convolutional dynamic ring sampling structure, a circular parameter modeling method, and an elliptic distortion-aware consistency loss function. The object detection detection process includes image acquisition, image input, model inference, result output, and placement state determination.

[0022] The target detection method in this embodiment takes an aluminum alloy die-cast filter screen as the target to be tested as an example, and specifically includes: Acquire images of the target object using an aluminum alloy die-cast filter screen; The image is input into the detection model based on YOLOv11 (the object detection model mentioned above), and multi-scale features are extracted through the backbone network; In the distortion-aware detection head, dynamic ring sampling and adaptive feature extraction are performed through a circular convolution module, so that the sampling points are adaptively shifted from near the ideal circular contour to the actual elliptical distortion boundary, thereby sensing the elliptical distortion features of the filter. Based on the Elliptic Distortion Aware Consistency Loss (EDC Loss), geometric constraints and elliptic axis ratio physical consistency constraints are integrated to enhance the perception and correction of elliptic distortion at the loss level. Output the test results and determine whether the filter is qualified (correctly placed) or unqualified (incorrectly placed).

[0023] Furthermore, the image acquired by the aluminum alloy die-cast filter is a high-resolution image acquired by industrial vision inspection equipment. Due to factors such as manual installation tilt, the ideal circular outline of the filter appears as an elliptical distortion in the image. Specifically, the acquired images are scaled to 640×640 resolution and normalized to eliminate the interference of illumination differences on the perception of elliptical distortion, thus preparing for subsequent model input.

[0024] Furthermore, the image is input into a detection model based on an improved YOLOv11, and multi-scale features are extracted through the backbone network, including: The YOLOv11 CSPDarknet backbone network is adopted, and the multi-scale features are upsampled and concatenated through the feature fusion module to enhance the expressive power of target features at different scales. Furthermore, such as Figure 3 As shown, the circular convolutional module in the distortion-aware detection head performs dynamic circular sampling and adaptive feature extraction, which is a structural improvement to the YOLOv11 model detection head, specifically including: The bounding box regression branch of the original YOLOv11 detector head is replaced with a circle-like module. This module pre-generates annular reference sampling coordinates with a radius of 2, which are used to simulate the ideal circular contour of the filter screen. ; The input features are processed by offset prediction convolution to adaptively learn the positional offset of each sampling point within the annular sampling region. This offset is constrained to the range [-1, 1] by the tanh function and multiplied by an expansion coefficient to control the offset range. It is then superimposed with the annular reference sampling coordinates to obtain the deformed sampling region. The deformed sampling coordinates are calculated as shown in the formula below, enabling the sampling points to adaptively fit the elliptical contour in the image.

[0025] in, The circular reference sampling coordinates, Here, α represents the original offset predicted by the model, and α is the offset expansion coefficient. To adapt the deformation sampling coordinates of the elliptical contour.

[0026] Bilinear interpolation sampling is performed based on deformed coordinates to obtain key feature points on the elliptical contour. After feature aggregation through grouped convolution, the basic elliptical distortion features are output after GroupNorm normalization and SiLU activation.

[0027] in, For deformation sampling coordinates The four nearest integer coordinates, This is the decimal offset of the coordinates. These are the characteristic values ​​of the corresponding coordinates.

[0028] Specifically, the circular adaptive convolution module first constructs a basic coordinate system for annular sampling using a baseline coordinate generation logic, selecting only sampling points in the annular region with a preset distance from the center pixel. Then, using a deformable coordinate generation logic, it fuses the basic coordinates with the offset learned by the model to obtain deformable sampling coordinates adapted to the target shape. Finally, relying on bilinear interpolation sampling logic, it extracts feature values ​​at corresponding positions in the feature map based on the deformed sampling coordinates, ultimately outputting elliptical distortion features with a fixed number of channels and spatial resolution consistent with the input feature map. This module dynamically adjusts the spatial distribution of the sampling kernel, enabling the original standard circular sampling region to adaptively fit the elliptical distortion contour of the filter in the image. This overcomes the limitations of traditional rectangular convolution kernels in capturing circular / elliptical structural features, achieving precise front-end perception of the elliptical distortion features of the filter.

[0029] like Figure 2 As shown in the figure, the core principle of the ellipse distortion perception consistency loss function of the present invention is: taking the inherent red circle on the die casting as the physical benchmark, the green ellipse detected by the model as the actual filter shape, and realizing the perception of distortions such as filter misalignment and tilting by comparing the differences between the two in multiple dimensions.

[0030] The red circle represents the ideal shape of the filter when it is undistorted, with its center and radius serving as fixed evaluation criteria. The green ellipse, on the other hand, represents the distorted shape of the filter when it is misaligned, described by its center, major axis, and minor axis. By comparing the deviations of the ellipse and the circle in three dimensions—center position, size, and major / minor axis differences—a consistency loss function is constructed to guide the model towards optimizing towards the "distortion-free circle," thereby improving the detection accuracy of filter posture defects.

[0031] Furthermore, based on the elliptic distortion-aware consistency loss function, and by integrating geometric constraints and physical consistency constraints on elliptic axis ratio, the perception and correction of elliptic distortion are enhanced at the loss level, including: The predicted circle parameters (cx, cy, r) output by the detection head are decoded into circle coordinates in actual space through a function and matched with the real circle parameters extracted from the label mask. The ellipse distortion-perceived consistency loss function is calculated as a basic metric for geometric consistency. Optionally, the perception of distortion in the central region of the ellipse can be enhanced by introducing the inner circle IoU (default scaling factor 0.7). Compute the geometric penalty term, including the center-distance penalty. Radius error penalty and normalized to the image size range; By extracting the ellipse axis ratio (b / a) from the labeled mask as a physical prior, and calculating the difference between the predicted axis ratio and the true axis ratio as a physical consistency penalty term, the perception of ellipse shape distortion is enhanced; among them, the principal axis length a and secondary axis length b of the target pixel region in the labeled mask can be obtained by contour fitting, minimum bounding ellipse or principal component direction statistics. The geometric penalty term and the physical consistency penalty term of the loss function are weighted and fused together. The elliptic distortion-aware consistency loss function is calculated by function calculation, and the final output loss value (1 − EDC Loss) is used for backpropagation to update the network. Specifically, the EDC Loss calculation process is as follows: 1. Spatial decoding is performed on the predicted contour parameters and the true contour parameters to obtain the center coordinates (cx, cy) and radius r in the actual scene, and then mapped to the true scale of the image.

[0032] 2. Calculate the Circle-IoU ratio based on the relative positional relationship between the two circular outlines:

[0033] When there is no overlap When the area is completely contained, take the smaller circle area; when there is partial overlap, calculate the exact solution by subtracting the triangle area from the sector area.

[0034] 3. Optionally, an Inner-Circle-IoU ratio can be introduced. By using a radius scaling factor k=0.7 to weight the inner and outer regions, the model's ability to perceive distortion in the central region of the ellipse can be enhanced.

[0035] 4. Construct a normalized geometric penalty term, including center distance penalty and radius deviation penalty:

[0036] It is then normalized to the image scale range [0,1] to ensure stable training at different resolutions.

[0037] Introducing the elliptic axis ratio e=b / a as a physical consistency constraint, the axis ratio difference penalty term is calculated:

[0038] Incorporate into the total penalty with a weighting coefficient λ=0.5:

[0039] 6. The final ellipse distortion perception consistency loss EDC Loss is defined as:

[0040] The value range is constrained between [0, 1], which is used for network backpropagation and parameter iterative updating.

[0041] Further, outputting the detection result to determine whether the filter screen is qualified or unqualified includes: Perform Non-Maximum Suppression (NMS) on the ellipse parameters (center, radius, axis ratio) and class probabilities output by the detection head to filter redundant detection frames; Determine whether the placement posture of the detected filter screen meets the process requirements according to the detected center offset, radius deviation and ellipse axis ratio of the filter screen; If both the offset and the deviation are within the threshold range and the ellipse axis ratio is close to 1, it is determined as qualified (correct placement); otherwise, it is determined as unqualified (incorrect and skewed placement); Specifically, in the inference phase, the model outputs the (cx, cy, r, score, class) information of each detected target. After filtering duplicate detections through NMS (IoU threshold 0.5), the center coordinates and radius of the filter screen are extracted. If the offset of the center from the image center is less than 5 pixels, the radius deviation is less than 3% of the nominal value, and the ellipse axis ratio is greater than 0.95, the filter screen is determined to be qualified; otherwise, it is determined to be unqualified, and the skew angle and offset are output to guide manual adjustment.

[0042] A specific application example of the present invention is as follows: On the aluminum alloy die-casting filter screen production line, an industrial camera collects the front image after the filter screen is installed, and the image resolution is 640×640. Due to the inaccuracy of manually placing the filter screen, the outline of the filter screen presents ellipse distortion (non-standard circle) in the image.

[0043] 1. After the image is preprocessed by scaling, normalization and other operations, it is input into the improved YOLOv11 detection model. After the backbone network completes multi-scale feature extraction and feature fusion, the result is sent to the distortion perception detection head; wherein the quasi-circular convolution module learns the dynamic offset of sampling points by generating annular reference sampling coordinates, so that the sampling kernel adaptively deforms from a standard circle to an ellipse shape, accurately extracts the ellipse distortion features of the filter screen and outputs ellipse parameters; 2. Parse the true contour parameters from the labeled mask, including the center coordinates, radius, and ellipse ratio; the ellipse ratio is obtained by statistically analyzing the principal axis length and secondary axis length of the pixels within the target mask area, and serves as the true supervision information for model training; 3. In the loss calculation stage, Elliptical Distortion Perceived Consistency Loss (EDC Loss) is adopted to comprehensively consider the intersection-union ratio of the predicted contour and the true contour, the deviation of the center position, the radius deviation, and the difference of the ellipse axis ratio to construct a geometric penalty term and a physical consistency penalty term. The final loss value is obtained by weighted fusion. 4. Backpropagation is performed using the final loss to continuously update network parameters and enhance the model's ability to perceive and locate elliptical distortion. During the inference phase, if the filter center is detected to be 8 pixels off-center and the elliptical axis ratio is 0.88, it is judged as unqualified, and "Incorrect placement" is output with a skew angle prompt. If the center is only 2 pixels off-center and the elliptical axis ratio reaches 0.97, it is judged as qualified, and "Correct placement" is output.

[0044] This invention proposes a method for detecting aluminum alloy die-cast filter screens based on elliptic distortion perception. It achieves precise front-end feature-level perception of elliptic distortion through a circular convolution module, and constructs geometric and physical consistency constraints at the loss level using Elliptic Distortion Perception Consistency Loss (EDC Loss), forming a full-link elliptic distortion perception mechanism from feature extraction to loss supervision. This enables high-precision detection of filter screen installation posture and contour distortion. This method effectively improves the robustness, accuracy, and efficiency of filter screen defect detection in industrial scenarios, and is suitable for fields with high installation accuracy requirements, such as aluminum alloy filter press equipment production, precision component assembly, and automated visual inspection.

[0045] The proposed method for detecting aluminum alloy die-cast filter screens based on elliptic distortion perception has significant technical advantages and broad application prospects. This method overcomes the limitation of traditional rectangular convolution kernels in accurately perceiving circular / elliptical structures. By using a circular convolution module to achieve dynamic adaptive deformation of the sampling kernel, it can efficiently capture the elliptic distortion features of the filter screen caused by installation deviations and viewing angle changes. Simultaneously, the Elliptic Distortion Perception Consistency Loss (EDC Loss) introduces the elliptic axis ratio as a physical prior, enabling the model to deeply perceive distortion from both geometric matching and physical morphology perspectives, effectively overcoming the shortcomings of traditional loss functions that only focus on overlapping regions. Furthermore, this invention seamlessly embeds the improved structure into the YOLOv11 framework, maintaining high inference speed while ensuring detection accuracy, meeting the real-time detection needs of industrial production lines. Compared with traditional manual detection and template matching methods, this method has stronger generalization ability and is more robust to changes in lighting and pose, significantly reducing labor costs and false detection rates, and improving the level of automated detection.

[0046] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0047] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications and improvements made by those skilled in the art without departing from the technical concept of the present invention should fall within the scope of protection defined by the claims of the present invention.

[0048] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A target detection model, characterized in that, Based on improvements to the YOLOv11 model, the invention includes: a distortion-aware detection head, which constructs a dynamic ring-shaped sampling structure based on circular convolution; the circular convolution refers to a convolution structure that uses the ideal circular contour of the target as a priori, generates ring-shaped reference sampling points around the center of the feature map pixels, and learns the two-dimensional offset of each sampling point through offset prediction convolution; the distortion-aware detection head utilizes this circular convolution to enable the sampling structure to adaptively deform according to the target shape, thereby enhancing the model's responsiveness to elliptical distortion boundary structures.

2. The target detection model according to claim 1, characterized in that, The dynamic ring sampling structure includes: constructing a set of sampling points that satisfy the ring constraint conditions within a preset radius; constraining the predicted offset by batch normalization and nonlinear activation functions; obtaining the feature response of the deformed sampling position based on bilinear interpolation; and completing the feature enhancement expression by aggregating the ring sampling features.

3. The target detection model according to claim 1, characterized in that, The target detection model uses a circular parameter modeling method to geometrically represent the target, representing the target as the coordinates of the center of a circle and the radius parameter, and then restores it to the original image space through scale mapping, thus achieving a geometric modeling method consistent with the ring sensing structure.

4. The target detection model according to claim 1, characterized in that, During the training phase, the target detection model uses a circular-like improved distance intersection-union ratio (DIU) loss function that incorporates physical consistency constraints to optimize the prediction results. Based on the circular DIU calculation, the loss function introduces a center distance penalty term and a radius difference penalty term to construct geometric consistency constraints, and combines the elliptical axis ratio physical parameter to construct a physical consistency constraint term, so that the prediction results meet the physical structural rationality requirements while achieving geometric matching.

5. The target detection model according to claim 4, characterized in that, The improved circular crossover ratio (CUP) loss function includes an internal region CUP constraint mechanism and an interval modulation mechanism. The internal region CUP constraint mechanism calculates the inner circle CUP by scaling the radii of the predicted circle and the real circle according to a preset scaling factor, and uses it together with the outer circle CUP as an overlap constraint. The interval modulation mechanism performs nonlinear compression or weighting on the loss value according to the interval in which the CUP is located, so that low overlap, medium overlap and high overlap regions have different optimization intensities to improve training stability.

6. The target detection model according to claim 1, characterized in that, The distortion-aware detection head replaces the bounding box regression branch of the original YOLOv11 detection head with a circular module. This circular module pre-generates annular reference sampling coordinates with a radius of 2 to simulate the ideal circular contour of the filter, and adjusts the annular reference sampling coordinates to deformable sampling coordinates that adapt to the elliptical distortion contour through a learnable offset.

7. A target detection method based on elliptic distortion sensing, characterized in that, include: Acquire the target placement image; The target placement image is input into the target detection model according to any one of claims 1-6, and the target detection model outputs the target category, structural parameters and confidence information; the target placement state is determined according to the structural parameters.

8. The target detection method based on elliptic distortion perception according to claim 7, characterized in that, It also includes extracting elliptical distortion physical parameters; wherein, the target segmentation mask is a binary image of the target region obtained by annotation or segmentation. By statistically analyzing the coordinates of the pixels belonging to the target region in the target segmentation mask, the principal axis direction, secondary axis direction, principal axis length and secondary axis length of the target region are determined, and the ratio of the principal axis to the secondary axis is calculated as a parameter representing the degree of elliptical distortion, which is used to reflect the degree of structural stretching or compression of the target during the imaging process.

9. The target detection method based on elliptic distortion perception according to claim 7, characterized in that, By introducing the inner circle IoU, the perception of distortion in the central region of the ellipse is enhanced; Compute the geometric penalty term, including the center-distance penalty. Radius error penalty And normalize to the image size range.

10. The target detection method based on elliptic distortion perception according to claim 7, characterized in that, Bilinear interpolation sampling is performed based on deformed coordinates to obtain key feature points on the elliptical contour. After feature aggregation through grouped convolution, GroupNorm normalization and SiLU activation are applied to output the basic features of elliptical distortion. in, For deformation sampling coordinates The four nearest integer coordinates, This is the decimal offset of the coordinates. These are the characteristic values ​​of the corresponding coordinates.

Images