Illumination adaptive underwater checkerboard ordered sub-pixel angular point detection method

Abstract

The invention discloses an illumination adaptive underwater checkerboard ordered sub-pixel angular point detection method, and relates to the technical field of computer vision and underwater robots. The method comprises the following steps: firstly, carrying out feature enhancement on an input image through an illumination adaptive front end based on a Retinex theory and gradient fusion, and extracting a global illumination embedding vector; and dynamically modulating the characteristics of the backbone network according to an illumination embedding vector by using a condition batch normalization mechanism so as to adapt to different underwater illumination conditions. Secondly, outputting an angular point confidence map, a sub-pixel offset field, a row sequence classification thermodynamic diagram and a column sequence classification thermodynamic diagram in parallel through a multi-head prediction decoder, and introducing angular point auxiliary classification loss in training to strengthen row and column semantic separability; and finally, constructing a global cost matrix fusing angular point confidence and row and column response, solving optimal topological mapping by using a bipartite graph matching algorithm, and directly outputting ordered sub-pixel angular points conforming to a physical topological structure. According to the method, the detection robustness and the positioning precision in an underwater complex scene are remarkably improved, and the problem of disordered angular point sorting is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of computer vision and camera calibration technology, specifically to an illumination-adaptive ordered checkerboard corner detection method for visual navigation and precision measurement of underwater robots. Background Technology

[0002] With the rapid development of applications such as machine vision, robot localization and navigation, industrial dimensional measurement, and augmented reality, camera calibration, as a crucial foundational technology for recovering realistic 3D spatial geometry from image space, has been widely researched and applied. Among existing camera calibration methods, using a planar checkerboard calibration plate, which detects corner points within the checkerboard and performs geometric optimization, has become a mainstream solution that is convenient, low-cost, and highly accurate. In underwater environments, due to the absorption and scattering of light by water, images often exhibit blue-green color casts, low contrast, and fogging. Simultaneously, underwater cameras are typically equipped with waterproof covers (planar or hemispherical), and the refraction of light at the water-air interface can lead to severe pincushion distortion or nonlinear distortion. Existing general corner detection methods rely on clear gradient features, which are prone to failure at underwater blurry and strongly distorted edges, and it is difficult to guarantee sub-pixel accuracy. Therefore, how to stably and accurately detect checkerboard corner points under complex underwater imaging conditions, especially obtaining sub-pixel level precise positions, is a key aspect of improving calibration accuracy.

[0003] In existing technologies, checkerboard corner detection typically employs a two-stage process: first, candidate corners are found at the integer pixel level using traditional feature operators such as Harris corner, SUSAN, and FAST, or rules based on gradient extrema; then, sub-pixel interpolation and refinement are performed near the candidate locations using methods such as quadratic surface fitting, image moments, or least squares fitting. These methods are simple to implement, computationally inexpensive, and can obtain relatively reliable corner coordinates when imaging conditions are good, noise is low, and the checkerboard is fully visible. However, when the image has uneven illumination, blurriness, noise interference, occlusion, or reflected highlights, traditional integer corner detection is prone to missed detections and false detections. Furthermore, because it does not fully utilize the overall grid topology of the checkerboard, isolated local false detections are difficult to effectively eliminate, thus affecting the accuracy and robustness of subsequent calibration.

[0004] In recent years, with the widespread application of deep learning in object detection and keypoint localization, some methods have introduced convolutional neural networks to detect checkerboard corners, improving detection robustness by regressing corner heatmaps or directly regressing corner coordinates. These methods typically output corner response maps across the entire image, then combine them with non-maximum suppression to obtain corner positions. Although deep learning methods have certain advantages over traditional methods in complex backgrounds and noisy conditions, existing deep learning methods, while improving feature extraction capabilities, often weaken explicit constraints on the checkerboard topology, easily leading to corner order disorder and thus calibration failure. Furthermore, existing methods struggle to simultaneously achieve cross-domain robustness, sub-pixel accuracy, and topology recovery within a unified framework. Summary of the Invention

[0005] This invention aims to solve the following problems in existing underwater camera calibration techniques: uneven illumination in the underwater environment, large-angle checkerboard images, and dynamic blur lead to low corner detection recall; existing heatmap-based methods are limited by grid quantization errors, resulting in insufficient sub-pixel positioning accuracy, and deep learning methods lack explicit constraints on the global topology of the checkerboard, leading to disordered corner sorting under occlusion or noise interference, causing calibration failure.

[0006] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0007] An illumination-adaptive underwater checkerboard ordered subpixel corner detection method includes the following steps:

[0008] 1) Acquire checkerboard images of the underwater environment and enhance the images using a lighting adaptive module based on Retinex theory;

[0009] 2) Construct a conditional modulation backbone network, extract features from the enhanced image using this backbone network, and use the conditional modulation mechanism to normalize the illumination at the feature level;

[0010] 3) Construct a multi-head predictive decoder to decode the image features output by the conditional modulation backbone network to obtain a single-channel heatmap. Subpixel offset field , row sequence heatmap and sequence heatmap ;

[0011] 4) Construct a training set containing truth labels;

[0012] 5) Train the detection model consisting of an illumination adaptation module, a conditional modulation backbone network, and a multi-head predictive decoder using the training set data;

[0013] 6) Use the trained detection model to perform ordered corner inference based on global bipartite graph matching, and output ordered chessboard corners through a global optimization strategy.

[0014] Step 1) includes the following steps:

[0015] 1.1) The acquired underwater RGB image Convert to grayscale and perform multi-scale average pooling to estimate illumination components. ;

[0016] 1.2) Reflection component of the image in the logarithmic domain :

[0017] ;

[0018] in, and These are the mean and standard deviation of the reflection component, respectively;

[0019] 1.3) Extracting the reflection component using the Sobel operator gradient magnitude And the gradient magnitude is converted through a convolutional layer. With reflection component Residual fusion is performed, and the features fused by the residual module are superimposed back onto the original reflection components to obtain the edge-compensated image. :

[0020] ;

[0021] in, This indicates a convolution operation.

[0022] Step 2) includes the following steps:

[0023] 2.1) Downsample the image using a convolutional layer with a stride of 2;

[0024] 2.2) Replace the standard batch normalization layer in the backbone network with a conditional batch normalization layer, and pass through the backbone network in four stages to allow feature branches of different resolutions to interact in parallel.

[0025] The conditional batch normalization layer receives the illumination embedding vector. Affine transformation parameters are dynamically generated through fully connected layers. and And calculate the actual scaling factor. And then through Adaptive modulation is applied to the feature map, where,

[0026] ;

[0027] in, Represents the input feature map, This represents the output feature map after conditional modulation. This is the arithmetic mean of the pixel values ​​of all images in the current batch. It is the standard deviation of all pixel values ​​of all images in the current batch;

[0028] The illumination embedding vector By combining grayscale images and illumination components The mean and standard deviation are concatenated and then mapped using a multilayer perceptron.

[0029] Step 3) includes the following steps:

[0030] 3.1) The four feature maps output by the backbone network at different scales are upsampled to a uniform resolution and then fused.

[0031] 3.2) Utilize a shared backbone decoder to extract high-dimensional features from the fused features through convolutional layers and adjust the number of channels;

[0032] 3.3) The processed features are predicted using four parallel prediction heads, and single-channel heatmaps representing the confidence of corner points are output respectively. Characterizes the offset of the actual sub-pixel corner. The pixel represents the first pixel. Row order heatmap of row probability and the characterization of the pixel belonging to the first Heatmap of column order of column probabilities .

[0033] Step 4) includes the following steps:

[0034] 4.1) Based on the true corner coordinates of the original image, generate Gaussian-distributed corner heatmap labels using inverse distance transformation. :

[0035] ;

[0036] in, To control the hyperparameters of the Gaussian distribution width, The pixel coordinates in the image space. These are the coordinates of the actual corner points;

[0037] 4.2) Generate subpixel offset field labels with the same scale and resolution as the original image. ;

[0038] 4.3) Generate row category labels Category tags The number of channels for the row category labels corresponds to the number of rows in the chessboard, and the number of channels for the column category labels corresponds to the number of columns in the chessboard. For those belonging to the first Line 1 The corner of the column, only in the first row label. The first channel and column label The spatial location corresponding to the channel generates a Gaussian response.

[0039] Step 4.2) specifically refers to:

[0040] Initialize a zero vector field with the same size as the original image;

[0041] Iterate through the coordinates of each real corner point in the image. Determined by rounding down. Integer pixel anchor point ; Calculate the sub-pixel residual vector between the true coordinates and the integer anchor point. ;

[0042] Assign the residual vector to the offset field. Pixel value at;

[0043] In the offset field, except for the integer anchor points corresponding to the corner points, the values ​​of all other background positions are kept to zero.

[0044] In step 5), during training, the training images are input into the model, and the multi-task joint loss between the predicted output and the supervision labels generated in step S2 is calculated. The model is then trained using this loss function. The multi-task joint loss... for:

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] in, For corner regression loss, For mask offset loss, For sequential regression loss, To aid in classification loss, , , , These are the weights corresponding to different losses. This represents the total number of pixels in the feature map. For the foreground corner region, To prevent extremely small constants with a denominator of zero, Foreground weighted mask, For the heatmap of the sequence predicted by the model, This is a real sequence tag heatmap. This indicates an indicator function that takes the value 1 when the condition within the parentheses is met, and 0 otherwise. Representing corner point The corresponding actual row and column index labels, The total number of categories represents the total number of rows or columns of the chessboard. For category indexes, it indicates the c-th row or c-th column currently being traversed.

[0051] Step 6) includes the following steps:

[0052] 6.1) Perform non-maximum suppression on the output of the corner prediction head and extract local peak values ​​as candidate points. By calculating sub-pixel precise coordinates The sub-pixel coordinates are obtained by combining the offset field output, and all sub-pixel coordinates are used to form a candidate corner point set. ,in:

[0053] ;

[0054] in, Candidate points The corresponding prediction vector in the offset field;

[0055] 6.2) Construct the global cost matrix and define the logical topological node set of the chessboard grid as follows: With the Matching score of each candidate point :

[0056] ;

[0057] in, and These represent the responses of the row and column heatmaps in their respective channels. These are weighting coefficients, which control the confidence levels of the corner points. Row classification probability Weights of column classification probabilities ;

[0058] 6.3) Solving the objective function using the Hungarian algorithm :

[0059] ;

[0060] in, For the constructed logical grid node set, For the logical grid node set to candidate point set The mapping relationship;

[0061] 6.4) Based on the solution results, obtain the global matching result that maximizes the total compatibility score, and then directly output the ordered sub-pixel corner coordinates that conform to the grid topology.

[0062] The present invention has the following beneficial effects and advantages:

[0063] 1. By introducing illumination estimation and conditional modulation mechanisms, the model's cross-domain generalization ability under underwater strong illumination changes, shadows, and blurred scenes is significantly improved.

[0064] 2. By using decoupled subpixel offset prediction, the quantization error caused by feature map discretization is eliminated, achieving extremely low reprojection error in real underwater datasets, which is superior to existing mainstream methods.

[0065] 3. By using row / column factorized heatmaps and a global matching strategy, ordered corner points that conform to the physical topology can be directly output, avoiding calibration failures caused by disordered corner point order, and significantly improving effective viewing rate and end-to-end correctness. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method of the present invention.

[0067] Figure 2 This is a schematic diagram of the illumination-adaptive ordered sub-pixel corner detection network structure in this invention;

[0068] Figure 3 This is a schematic diagram of the adaptive lighting module in this invention;

[0069] Figure 4 This is a schematic diagram of the conditional batch normalization layer in this invention;

[0070] Figure 5 These are the model training images, corner heatmaps, and row and column sequence heatmaps used in this invention.

[0071] Figure 6 This is a schematic diagram of the corner detection results of an underwater checkerboard image in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0073] like Figure 1As shown, an illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method is described, which includes:

[0074] Step S1: Image Acquisition and Illumination Adaptive Preprocessing

[0075] The process involves acquiring checkerboard images captured in an underwater environment as input, and then constructing an illumination-adaptive module based on Retinex theory to enhance the images. Specifically, this step first uses multi-scale filtering to estimate the illumination components of the input image. And through the formula Restore reflection component Subsequently, the reflection component was extracted using the Sobel operator. gradient magnitude Furthermore, gradient features and reflection features are residually fused through convolutional layers to obtain edge-enhanced image features. This compensates for the loss of high-frequency edge information caused by scattering from the water medium during the front-end preprocessing stage, providing a clear structural input for subsequent feature extraction.

[0076] Step S2: Illumination-sensing feature extraction and conditional modulation

[0077] The preprocessed features from step S1 are input into the backbone of a deep convolutional neural network for multi-scale feature extraction, and illumination normalization at the feature level is achieved using a conditional modulation mechanism. This step utilizes an illumination estimator to extract illumination embedding vectors from global image statistics. Furthermore, a Conditional Batch Normalization (CBS) mechanism is introduced, based on the extracted illumination embedding vector. Dynamically generate affine parameters and These parameters are used to perform graph-based adaptive modulation of the feature layers of the backbone network, and the calculation formula is as follows: This allows the geometric features extracted by the network to dynamically adapt to changes in strong light, shadow interference, and exposure shift in the underwater environment, thereby improving cross-domain robustness.

[0078] Step S3: Structured Multi-Head Predictive Decoding

[0079] The multi-scale features output from the backbone network are input into a parallel multi-head decoder to simultaneously predict the geometric location and topological attribute information of corner points. This step outputs a corner existence probability heatmap through the corner prediction head. To indicate the approximate location of corner points; to output a dense vector field aligned with the image space by offsetting the prediction head output. This vector field encodes the residual displacement from the grid center to the true sub-pixel corner, which is used to subsequently eliminate quantization errors caused by feature map discretization; simultaneously, the row prediction head and column prediction head output the number of channels as the number of rows, respectively. heatmap The number of channels is equal to the number of columns. Heatmap The complex topological order of the chessboard is decomposed into independent row and column channel responses, thus achieving a factorized representation of the chessboard structure.

[0080] Step S4: Calculation of sub-pixel coordinates combining coarse and fine coordinates

[0081] Combining the corner heatmap and offset field output from step 3, high-precision sub-pixel coordinates are calculated using a "coarse detection + fine correction" approach. This step first uses the corner heatmap... Extract local maxima points and use them as coarse positioning anchor points on an integer scale. Then, extract the prediction vector corresponding to the anchor point position in the offset field. Using the formula The coarse coordinates are added to the offset to obtain the final sub-pixel accurate coordinates after eliminating quantization errors. .

[0082] Step S5: Ordered topology restoration based on the global cost matrix

[0083] Using the candidate corner points and row / column heatmaps obtained in step 3, an ordered set of corner points conforming to the physical structure is recovered through a global graph matching algorithm. This step first defines the logical grid node set. and the set of detected candidate corner points And construct a bipartite graph compatibility cost matrix, where logical nodes are connected. and candidate points Score Fusion Point It is the confidence level of the corner point, point In the The response value of each channel and the point In the The algorithm takes the response values ​​of each column channel as its three parts; finally, it uses the Hungarian algorithm to maximize the overall compatibility score and solve for the optimal matching mapping, thereby obtaining the correct row and column indexes in one go. An ordered set of corner points effectively eliminates outliers and resolves local sorting conflicts.

[0084] Example

[0085] This invention provides an illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method, which mainly includes the following steps:

[0086] Step S1: Construct a lighting-adaptive deep neural network model

[0087] The model constructed in this embodiment includes an illumination adaptation module, a conditional modulation backbone network, and a multi-head prediction decoder. The specific process of model construction is as follows:

[0088] First, an illumination-adaptive module is designed for feature enhancement, specifically as follows: Figure 3 As shown, the input RGB image is first converted to a grayscale image. Based on Retinex theory, this module performs multi-scale average pooling on the grayscale image to estimate the illumination components. The reflection component is recovered in the logarithmic domain using the following formula.

[0089]

[0090] in and These represent the mean and standard deviation of the reflection component, respectively; simultaneously, the gradient magnitude of the reflection component is calculated using the Sobel operator. ,like Figure 3 As shown, the reflection component With gradient magnitude After concatenation, the concatenated components are superimposed back onto the original reflection components using the residual module shown in the diagram to compensate for edge information. The following formula is used to superimpose the reflection components back onto the convolutional layer as residuals to compensate for edge information:

[0091]

[0092] in, This indicates a convolution operation used to map concatenated features back to image space or to perform feature fusion. This module also includes an illumination estimator that takes an input grayscale image and illumination components. The mean and standard deviation are concatenated (Concat) and then input into a multilayer perceptron (MLP) to be mapped into an illumination embedding vector. .

[0093] Secondly, a conditional modulation backbone network is constructed, which adopts a high-resolution network architecture (such as HRNet-W32). Specifically... Figure 2 As shown, the input features first pass through a pre-sampling (Stem) module, which contains convolutional layers with a stride of 2 for initial dimensionality reduction. They then enter the backbone network, which comprises four stages (Stage 1 to Stage 4), where feature branches of different resolutions interact in parallel at each stage. The standard batch normalization layer in this network is replaced with a conditional batch normalization layer, which receives the illumination embedding vector. Through a fully connected layer, Figure 4 Linear dynamically generated affine transformation parameters (marked in the middle) and .like Figure 4 As shown, in order to maintain the identity mapping property of the features, the actual scaling factor is calculated. Using formulas Adaptive modulation of the feature map, specifically as follows: Figure 4 As shown.

[0094] Finally, a multi-head prediction decoder is constructed. For example... Figure 2 As shown, firstly, the feature maps at four different scales output by the backbone network are upsampled to a uniform resolution and fused through a multi-scale fusion and upsampling module. The fused features then enter the shared trunk decoder, where high-dimensional features are further extracted and the number of channels is adjusted through convolutional layers. Finally, the processed features are split and input to four parallel prediction heads, which output single-channel heatmaps representing the confidence of corner points. Characterizes the offset of the actual sub-pixel corner. The pixel represents the first pixel. Row order heatmap of row probability and the characterization of the pixel belonging to the first Heatmap of column order of column probabilities The overall network structure is as follows Figure 2 As shown, the training image, corner heatmap, and row / column sequence heatmap are as follows: Figure 5 As shown.

[0095] Step S2: Generate physically consistent supervisory data

[0096] To train the above model, a training set containing ground truth labels needs to be constructed. For each image with a resolution of... From the training images, the following supervision information is generated: First, based on the real corner coordinates, a Gaussian-distributed corner heatmap label is generated using inverse distance transformation. , specifically Figure 4 As shown:

[0097]

[0098] in, It is a hyperparameter that controls the width of the Gaussian distribution. In the example, Secondly, the generated size is Subpixel offset field tags The specific generation process is as follows: First, initialize a zero vector field with the same size as the image; then, iterate through the coordinates of each real corner point in the image. The integer pixel anchor point is determined by rounding down. ; Calculate the sub-pixel residual vector between the true coordinates and the integer anchor point. Assign the residual vector to the migration field. The pixel values ​​at each location are set to zero. In the offset field, except for the integer anchor points corresponding to the corners, the values ​​at all other background locations remain zero. This sparse label generation method ensures that the model learns sub-pixel corrections only in the local regions where the corners exist, avoiding interference from background noise. Next, row classification labels are generated. Category tags The number of channels for the row category labels corresponds to the number of rows in the chessboard, and the number of channels for the column category labels corresponds to the number of columns in the chessboard. For those belonging to the first Line 1 The corner of the column, only in the first row label. The first channel and column label The spatial location corresponding to the channel generates a Gaussian response, specifically as follows: Figure 4 As shown.

[0099] Step S3: Model training based on multi-task joint loss

[0100] Input the training images into the model, and calculate the multi-task joint loss between the predicted output and the supervision labels generated in step S2. Defined as:

[0101]

[0102] Among them, corner regression loss Mean square error , Represents the total number of pixels in the feature map; mask offset loss The L2 error is calculated only in the corner region, i.e. Sequence regression loss The weighted mean square error is applied to the row and column prediction heatmaps, and the formula is as follows: ,in Foreground weighted mask; auxiliary classification loss function Calculate the cross-entropy at the corner points using the following formula: ,in This indicates an indicator function that takes the value 1 when the condition within the parentheses is met, and 0 otherwise. Representing corner point The corresponding real row and column index labels. The auxiliary classification loss function forces the network to... At the physical corner point, it must be in the first position of the row prediction header. The first channel and column prediction header The channel outputs the maximum response, thereby explicitly enhancing the separability of row and column semantics, as shown in this embodiment. All are 1.

[0103] Step S4: Ordered Corner Reasoning Based on Global Bipartite Graph Matching

[0104] During the model inference phase, an ordered chessboard corner point is output through a global optimization strategy. Specifically, the process involves first performing non-maximum suppression on the corner point prediction head output, and then extracting local peaks as candidate points. And through the formula Sub-pixel coordinates are obtained by combining the offset field output; then, a global cost matrix is ​​constructed, and the logical topological node set of the chessboard is defined as follows. With the Matching score of each candidate point The calculation formula is:

[0105] ;

[0106] in and These represent the responses of the row and column heatmaps in the corresponding channels, respectively, in this embodiment. Finally, the objective function is solved using the Hungarian algorithm. To obtain the global matching result that maximizes the overall compatibility score, the ordered sub-pixel corner coordinates that conform to the grid topology are directly output. The corner detection results of the underwater checkerboard image are as follows: Figure 6 As shown.

Claims

1. A method for detecting ordered sub-pixel corner points in an underwater checkerboard pattern with adaptive illumination, characterized in that, Includes the following steps: 1) Acquire checkerboard images of the underwater environment and enhance the images using a lighting adaptive module based on Retinex theory; 2) Construct a conditional modulation backbone network, extract features from the enhanced image using this backbone network, and use the conditional modulation mechanism to normalize the illumination at the feature level; 3) Construct a multi-head predictive decoder to decode the image features output by the conditional modulation backbone network to obtain a single-channel heatmap. Subpixel offset field , row sequence heatmap and sequence heatmap ; 4) Construct a training set containing truth labels; 5) Train the detection model consisting of an illumination adaptation module, a conditional modulation backbone network, and a multi-head predictive decoder using the training set data; 6) Use the trained detection model to perform ordered corner inference based on global bipartite graph matching, and output ordered chessboard corners through a global optimization strategy.

2. The illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method according to claim 1, characterized in that, Step 1) includes the following steps: 1.1) The acquired underwater RGB image Convert to grayscale and perform multi-scale average pooling to estimate illumination components. ; 1.2) Reflection component of the image in the logarithmic domain : ； in, and These are the mean and standard deviation of the reflection component, respectively; 1.3) Extracting the reflection component using the Sobel operator gradient magnitude And the gradient magnitude is converted through a convolutional layer. With reflection component Residual fusion is performed, and the features fused by the residual module are superimposed back onto the original reflection components to obtain the edge-compensated image. : ； in, This indicates a convolution operation.

3. The illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method according to claim 1, characterized in that, Step 2) includes the following steps: 2.1) Downsample the image using a convolutional layer with a stride of 2; 2.2) Replace the standard batch normalization layer in the backbone network with a conditional batch normalization layer, and pass through the backbone network in four stages to allow feature branches of different resolutions to interact in parallel.

4. The illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method according to claim 3, characterized in that, The conditional batch normalization layer receives the illumination embedding vector. Affine transformation parameters are dynamically generated through fully connected layers. and And calculate the actual scaling factor. And then through Adaptive modulation is applied to the feature map, where, ； in, Represents the input feature map, This represents the output feature map after conditional modulation. This is the arithmetic mean of the pixel values ​​of all images in the current batch. It is the standard deviation of all pixel values ​​of all images in the current batch; The illumination embedding vector By combining grayscale images and illumination components The mean and standard deviation are concatenated and then mapped using a multilayer perceptron.

5. The illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method according to claim 1, characterized in that, Step 3) includes the following steps: 3.1) The four feature maps output by the backbone network at different scales are upsampled to a uniform resolution and then fused. 3.2) Utilize a shared backbone decoder to extract high-dimensional features from the fused features through convolutional layers and adjust the number of channels; 3.3) The processed features are predicted using four parallel prediction heads, and single-channel heatmaps representing the confidence of corner points are output respectively. Characterizes the offset of the actual sub-pixel corner. The pixel represents the first pixel. Row order heatmap of row probability and the characterization pixel belongs to the first Heatmap of column order of column probabilities .

6. The illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method according to claim 1, characterized in that, Step 4) includes the following steps: 4.1) Based on the true corner coordinates of the original image, generate Gaussian-distributed corner heatmap labels using inverse distance transformation. : ； in, To control the hyperparameters of the Gaussian distribution width, The pixel coordinates in the image space. These are the coordinates of the actual corner points; 4.2) Generate subpixel offset field labels with the same scale and resolution as the original image. ; 4.3) Generate row category labels Category tags The number of channels for the row category labels corresponds to the number of rows in the chessboard, and the number of channels for the column category labels corresponds to the number of columns in the chessboard. For those belonging to the first Line number The corner of the column, only in the first row label. The first channel and column label The spatial location corresponding to the channel generates a Gaussian response.

7. The illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method according to claim 1, characterized in that, Step 4.2) specifically refers to: Initialize a zero vector field with the same size as the original image; Iterate through the coordinates of each real corner point in the image. Determined by rounding down. Integer pixel anchor point ; Calculate the sub-pixel residual vector between the true coordinates and the integer anchor point. ; Assign the residual vector to the offset field. Pixel value at; In the offset field, except for the integer anchor points corresponding to the corner points, the values ​​of all other background positions are kept to zero.

8. The illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method according to claim 1, characterized in that, In step 5), during training, the training images are input into the model, and the multi-task joint loss between the predicted output and the supervision labels generated in step S2 is calculated. The model is then trained using this loss function. The multi-task joint loss... for: ； ； ； ； ； in, For corner regression loss, For mask offset loss, For sequential regression loss, To aid in classification loss, , , , These are the weights corresponding to different losses. This represents the total number of pixels in the feature map. For the foreground corner region, To prevent extremely small constants with a denominator of zero, Foreground weighted mask, For the heatmap of the sequence predicted by the model, This is a real sequence tag heatmap. This indicates an indicator function that takes the value 1 when the condition within the parentheses is met, and 0 otherwise. Representing corner points The corresponding actual row and column index labels, The total number of categories represents the total number of rows or columns of the chessboard. For category indexes, it indicates the c-th row or c-th column currently being traversed.

9. The illumination-adaptive underwater checkerboard ordered sub-pixel corner detection method according to claim 1, characterized in that, Step 6) includes the following steps: 6.1) Perform non-maximum suppression on the output of the corner prediction head and extract local peak values ​​as candidate points. By calculating sub-pixel precise coordinates The sub-pixel coordinates are obtained by combining the offset field output, and all sub-pixel coordinates are used to form a candidate corner point set. ,in: ； in, Candidate points The corresponding prediction vector in the offset field; 6.2) Construct the global cost matrix and define the logical topological node set of the chessboard grid as follows: With the Matching score of each candidate point : ； in, and These represent the responses of the row and column heatmaps in their respective channels. These are weighting coefficients, which control the confidence levels of the corner points respectively. Row classification probability Weights of column classification probabilities ; 6.3) Solving the objective function using the Hungarian algorithm : ； in, For the constructed logical grid node set, For the logical grid node set to candidate point set The mapping relationship; 6.4) Based on the solution results, obtain the global matching result that maximizes the total compatibility score, and then directly output the ordered sub-pixel corner coordinates that conform to the grid topology.

Images