Search

CN-122023518-A - Illumination self-adaptive underwater checkerboard ordered sub-pixel corner detection method

CN122023518ACN 122023518 ACN122023518 ACN 122023518ACN-122023518-A

Abstract

The invention discloses a light self-adaptive underwater checkerboard ordered sub-pixel corner detection method, and relates to the technical field of computer vision and underwater robots. The method comprises the steps of carrying out feature enhancement on an input image through an illumination self-adaptive front end based on Retinex theory and gradient fusion, extracting a global illumination embedded vector, and dynamically modulating backbone network features according to the illumination embedded vector by utilizing a condition batch normalization mechanism so as to adapt to different underwater illumination conditions. And secondly, outputting a corner 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 predictive decoder, and introducing corner auxiliary classification loss in training to strengthen row and column semantic separability. And finally, constructing a global cost matrix integrating the angular point confidence and the row-column response, solving the optimal topology map by utilizing a bipartite graph matching algorithm, and directly outputting ordered sub-pixel angular points conforming to a physical topology structure. The invention obviously improves the detection robustness and the positioning accuracy under the underwater complex scene, and effectively solves the problem of disorder of angular point ordering.

Inventors

  • ZHANG QIFENG
  • WANG GUODONG
  • ZHANG YUNXIU

Assignees

  • 中国科学院沈阳自动化研究所

Dates

Publication Date
20260512
Application Date
20260112

Claims (9)

  1. 1. The illumination self-adaptive underwater checkerboard ordered sub-pixel corner detection method is characterized by comprising the following steps of: 1) Acquiring a checkerboard image shot in an underwater environment, and enhancing the image through an illumination self-adaptive module based on a Retinex theory; 2) Constructing a conditional modulation backbone network, extracting features of the enhanced image through the backbone network, and carrying out illumination normalization on a feature level by utilizing a conditional modulation mechanism; 3) Constructing a multi-head predictive decoder, decoding image characteristics output by a conditional modulation backbone network to obtain a single-channel thermodynamic diagram Sub-pixel offset field Thermodynamic diagram of line sequence Thermodynamic diagram of sequence ; 4) Constructing a training set containing true value labels; 5) Training a detection model consisting of an illumination self-adaptation module, a conditional modulation backbone network and a multi-head predictive decoder by using training set data; 6) And (3) using a trained detection model to perform ordered corner reasoning based on global bipartite graph matching, and outputting ordered checkerboard corner through a global optimization strategy.
  2. 2. The illumination-adaptive underwater checkerboard ordered subpixel corner detection method according to claim 1, wherein the step 1) comprises the following steps: 1.1 Underwater RGB image to be acquired) Converting into gray level diagram, and performing multi-scale average pooling to estimate illumination component ; 1.2 Restoring the reflected component of an image in the log domain : ; Wherein, the And The mean value and standard deviation of the reflected components are respectively; 1.3 Extracting reflected components using Sobel operator Gradient amplitude of (2) And the gradient amplitude is transmitted through the convolution layer And the reflected component Residual fusion is carried out, and the original reflection component is overlapped through the characteristics fused by the residual module to obtain an image after edge compensation : ; Wherein, the Representing a convolution operation.
  3. 3. The illumination-adaptive underwater checkerboard ordered subpixel corner detection method according to claim 1, wherein the step 2) comprises the following steps: 2.1 Downsampling the image by a convolution layer with a step size of 2; 2.2 The standard batch normalization layer in the backbone network is replaced by the conditional batch normalization layer, and feature branches with different resolutions are interacted in parallel through the backbone network of four stages in sequence.
  4. 4. A method for detecting corner points of ordered sub-pixels of an illumination-adaptive underwater checkerboard according to claim 3, wherein said conditional normalization layer receives illumination embedded vectors Dynamic generation of affine transformation parameters through fully connected layers And And calculates the actual scaling factor Further by The signature is adaptively modulated, wherein, ; Wherein, the A feature map representing the input is presented, The output characteristic diagram after the conditional modulation is shown, Being the arithmetic average of all image pixel values in the current batch, Is the standard deviation of all pixel values of all images in the current batch; The illumination embedding vector By combining gray-scale images with illumination components And (3) splicing the mean value and the standard deviation, and mapping through a multi-layer perceptron to obtain the target product.
  5. 5. The illumination-adaptive underwater checkerboard ordered subpixel corner detection method according to claim 1, wherein the step 3) comprises the following steps: 3.1 The four feature images with different scales output by the backbone network are adjusted to the uniform resolution through up-sampling and are fused; 3.2 Extracting high-dimensional features in the fusion features through a convolution layer and adjusting the number of channels by using a shared trunk decoder; 3.3 Predicting the processed characteristics through four parallel prediction heads, and respectively outputting single-channel thermodynamic diagrams representing the confidence coefficient of the angular points Characterizing the angular point offset to the real subpixel Characterization pixel belongs to Line sequential thermodynamic diagram of line probability Characterization pixel belongs to Column sequential thermodynamic diagram of column probabilities 。
  6. 6. The illumination-adaptive underwater checkerboard ordered subpixel corner detection method according to claim 1, wherein the step 4) comprises the following steps: 4.1 Based on the real angular point coordinates of the original image, generating angular point thermodynamic diagram labels in Gaussian distribution by using inverse distance transformation : ; Wherein, the In order to control the hyper-parameters of the gaussian distribution width, For pixel coordinate locations in image space, Coordinates of the real corner points; 4.2 Generating a subpixel offset field label having the same scale as the original image resolution ; 4.3 Generating row classification labels Sum column class labels Wherein the number of channels of the row classification labels corresponds to the number of checkerboard lines, and the number of channels of the column classification labels corresponds to the number of checkerboard lines For belonging to the first Line 1 Corner points of columns, only at row labels Channel and column labels The spatial locations corresponding to the channels generate gaussian responses.
  7. 7. The illumination-adaptive underwater checkerboard ordered subpixel corner detection method according to claim 1, wherein the step 4.2) is specifically: Initializing an all-zero vector field with the same size as the original image; Traversing each real corner coordinate in the image Determination by a rounding-down operation Integer pixel anchor point where Calculating a sub-pixel residual vector between the real coordinates and the integer anchor point ; Assigning residual vectors to offset fields Pixel values at; and keeping the values of all the rest background positions except the integer anchor point positions corresponding to the corner points in the offset field to be zero.
  8. 8. The illumination-adaptive underwater checkerboard ordered subpixel corner detection method according to claim 1, wherein in the training process of step 5), a training image is input into a model, a multi-task joint loss between a prediction output and the supervision label generated in step S2 is calculated, the model is trained through the loss function, and the multi-task joint loss The method comprises the following steps: ; ; ; ; ; Wherein, the For the corner point regression loss, In order for the mask to be offset lost, For the loss of the sequence regression, In order to assist in the classification loss, 、 、 、 The weights corresponding to the different losses are respectively given, As the total number of pixels in the feature map, A set of masks for the foreground corner regions, To prevent the denominator from being a very small constant of zero, The mask is weighted for the foreground and, For the sequence thermodynamic diagram predicted by the model, For a true sequence tag thermodynamic diagram, Indicating a function, which takes a value of 1 when the condition in brackets is satisfied, otherwise 0, Representing corner points The corresponding true rank index tag is used to determine, For the total number of categories, the total number of rows or columns of the checkerboard is represented, For category index, the c-th row or c-th column currently traversed is represented.
  9. 9. The illumination-adaptive underwater checkerboard ordered subpixel corner detection method according to claim 1, wherein the step 6) comprises the following steps: 6.1 Non-maximum suppression of the output of the angle point pre-measurement head, and extracting local peak values as candidate points By calculating the precise coordinates of the sub-pixels Combining the offset field output to obtain sub-pixel coordinates, and forming a candidate corner set from all the sub-pixel coordinates Wherein: ; Wherein, the As candidate points A corresponding prediction vector in the offset field; 6.2 Constructing a global cost matrix, and defining a logic topological node set of a checkerboard as And the first Matching score for individual candidate points : ; Wherein, the And Representing the response of the row and column thermodynamic diagrams in the corresponding channels, Is a weighting coefficient for controlling the confidence coefficient of the angular point respectively Line classification probability Weighting of column classification probabilities ; 6.3 Using hungarian algorithm to solve the objective function : ; Wherein, the For the set of logical grid nodes to be constructed, To be a slave logic grid node set To a candidate point set Mapping relation of (3); 6.4 Obtaining a global matching result of the maximum total compatibility score according to the solving result, thereby directly outputting the ordered sub-pixel corner coordinates conforming to the grid topological structure.

Description

Illumination self-adaptive underwater checkerboard ordered sub-pixel corner detection method Technical Field The invention relates to the technical field of computer vision and camera calibration, in particular to an illumination self-adaptive ordered checkerboard corner detection method for underwater robot vision navigation and precise measurement. Background With the rapid development of applications such as machine vision, robot positioning and navigation, industrial size measurement, and augmented reality, camera calibration is widely studied and applied as an important basic technology for recovering a real three-dimensional space geometrical relationship from an image space. In the existing camera calibration method, a plane checkerboard calibration plate is adopted, and the internal corners of the checkerboard are detected and geometrically optimized, so that the method has become a mainstream scheme with convenient use, low cost and high precision. In an underwater environment, images tend to appear bluish green, low in contrast and hazy due to absorption and scattering of light by a water body, and meanwhile, an underwater camera is usually provided with a waterproof cover (plane or hemispherical), so that serious pincushion distortion or nonlinear distortion can be caused by refraction of light at a water-gas interface. The existing universal corner detection method depends on clear gradient characteristics, is easy to fail at the underwater blurred and strong distorted edges, and is difficult to ensure the sub-pixel precision. Therefore, how to stably and accurately detect the angular points of the checkerboard under the underwater complex imaging condition, in particular to obtain the accurate position of the sub-pixel level, is a key link for improving the calibration precision. In the prior art, checkerboard corner detection generally adopts a two-stage process, namely, firstly, candidate corner points are found at integer pixel level by utilizing traditional feature operators such as Harris corner points, SUSAN, FAST and the like or rules based on gradient extremum, and then sub-pixel interpolation and refinement are carried out nearby candidate positions by using methods such as quadric surface fitting, image moment or least square fitting and the like. The method is simple to implement, has small calculated amount, and can obtain more reliable angular point coordinates when the imaging condition is good, the noise is low and the checkerboard is completely visible. However, when the image has the conditions of uneven illumination, blurring, noise interference, shielding or reflecting high light and the like, the conventional integer angular point detection is easy to generate missed detection and false detection, and the isolated local false detection is difficult to be effectively removed due to the fact that the grid topological structure of the whole checkerboard is not fully utilized, so that the subsequent calibration precision and robustness are affected. In recent years, with the wide application of deep learning in target detection and key point positioning, some methods introduce convolutional neural networks to detect checkerboard corner points, and the robustness of detection is improved by regression corner point thermodynamic diagrams or direct regression of corner point coordinates. The method generally outputs a corner response diagram on the whole image, and combines non-maximum suppression to obtain the corner position. Although the deep learning method has certain advantages over the traditional method under the complex background and noise conditions, the existing deep learning method improves the feature extraction capability, but weakens the explicit constraint on the checkerboard topological structure, and is easy to cause the angular point sequence disorder, thereby causing calibration failure. In addition, the existing method is difficult to achieve cross-domain robustness, sub-pixel precision and topology recovery in a unified frame. Disclosure of Invention The invention aims to solve the problems that the recall rate of corner detection is low due to uneven illumination, large-inclination checkerboard images and dynamic blurring of an underwater environment in the existing underwater camera calibration technology, the existing thermodynamic diagram-based method is limited by grid quantization errors, the sub-pixel positioning precision is insufficient, and the depth learning method lacks explicit constraint on the global topology of the checkerboard, so that the corner ordering is disordered under shielding or noise interference, and calibration failure is caused. The technical scheme adopted by the invention for achieving the purpose is as follows: an illumination self-adaptive underwater checkerboard ordered sub-pixel corner detection method comprises the following steps: 1) Acquiring a checkerboard image shot in an underwater environment, and enhancing the image