A Calibration-Free Zoom Camera Self-Calibration Method and System Based on Deep Learning

Abstract

The present invention discloses a calibration-free zoom camera self-calibration method and system based on deep learning, belonging to the technical field of camera calibration, including: obtaining an image group with the same viewing angle and different focal lengths and an image group with the same focal length and different viewing angles; using an improved lightweight ASTR algorithm to perform feature point matching on the two image groups to obtain a first group of matching point pairs and a second group of matching point pairs; based on the first group of matching point pairs, using the least squares method to calculate the scaling center of the image group with the same viewing angle and different focal lengths; based on the second group of matching point pairs, calculating the camera internal parameter matrix of the image group with the same focal length and different viewing angles, and performing polynomial fitting based on the camera internal parameter matrix to solve the polynomial to obtain the camera internal parameter value; based on the obtained scaling center and camera internal parameter value, completing the self-calibration of the zoom camera. The present invention not only reduces the computational complexity, but also expands the applicable scenario range of the method, effectively improving the calibration accuracy and robustness.

Description

Technical Field

[0001] The present invention belongs to the technical field of camera calibration, and particularly relates to a calibration plate-free zoom camera self-calibration method and system based on deep learning. Background Art

[0002] Camera calibration is a key step in obtaining three-dimensional spatial depth information from two-dimensional images and plays an important role in multiple fields such as computer vision, robotics, and autonomous driving. Through precise calibration, the internal and external parameters of the camera can be determined, thereby enabling high-precision three-dimensional reconstruction, target recognition, and tracking tasks. Camera calibration is not only the basis for achieving accurate depth perception but also a key link in ensuring the performance of various vision systems. Nowadays, the calibration methods for fixed-focus cameras are relatively mature and can meet the needs of most applications. However, in modern image processing and photography applications, multi-focus cameras are widely used in consumer devices (such as single-lens reflex cameras and mobile phones) and industrial devices (such as drones and surveillance systems). In the case of multiple focal lengths, due to the dynamic changes in the camera's internal parameters, distortion coefficients, and optical center displacements caused by the focal length change, it poses a great challenge to camera calibration and subsequent calculations.

[0003] Currently, although there are available parameter models for zoom camera calibration, its research is usually based on the high-precision camera calibration method using a calibration plate and is not applicable to daily or dynamic scenarios. Therefore, the self-calibration of zoom cameras is an important research direction. Among the current self-calibration methods, the techniques based on vanishing points or Aruco codes rely on specific geometric characteristics of the scene and are difficult to meet the requirements of dynamic or complex scenarios; the methods based on feature matching, although getting rid of the limitation of the target, mostly use traditional matching algorithms (such as SIFT), with high computational complexity, weak robustness, and the need for additional post-processing to eliminate mismatched point pairs, resulting in complex operations and low efficiency. The ASTR algorithm based on deep learning can handle complex or dynamic scenarios and can also adapt to large-scale changes, but the model is relatively complex and the computational amount is large.

[0004] Therefore, it is necessary to design a calibration plate-free zoom camera self-calibration method and system based on deep learning for the above problems. Summary of the Invention

[0005] The object of the present invention is to address the problems existing in the prior art and provide a calibration plate-free zoom camera self-calibration method and system based on deep learning. By proposing an improved lightweight ASTR algorithm and combining the prior knowledge of the change in the internal parameters of the zoom camera, the polynomial fitting method is optimized, which not only reduces the deviation caused by noise and non-linear errors in the traditional method but also effectively improves the calibration accuracy and robustness, expands the applicable scene range of the method, and introduces deep learning feature matching and the least squares method into the calibration of the image scaling center, effectively improving the calibration accuracy and robustness.

[0006] According to one aspect of this specification, a method for self-calibration of a zoom camera without a calibration board based on deep learning is provided, including:

[0007] Based on the captured images, group the images to obtain image groups with the same viewing angle and different focal lengths and image groups with the same focal length and different viewing angles;

[0008] Use the improved lightweight adaptive point-guided transformer model to perform feature point matching on the two image groups to obtain the first set of matching point pairs and the second set of matching point pairs;

[0009] Based on the first set of matching point pairs, use the least squares method to calculate the scaling center of the image groups with the same viewing angle and different focal lengths; based on the second set of matching point pairs, calculate the camera internal parameter matrix of the image groups with the same focal length and different viewing angles, and perform polynomial fitting based on the camera internal parameter matrix to solve the polynomial to obtain the camera internal parameter values;

[0010] Based on the scaling center of the image groups with the same viewing angle and different focal lengths and the camera internal parameter values of the image groups with the same focal length and different viewing angles, complete the self-calibration of the camera.

[0011] Furthermore, the improvement of the lightweight adaptive point-guided transformer model (ASTR) includes:

[0012] Based on the backbone network of the Mobile Convolutional Neural Network V4 (MobileNetV4), introduce the inverted bottleneck block (UIB) and the depthwise separable convolution module;

[0013] Replace the Feature Pyramid Network (FPN) module of the backbone network with the Mobile Convolutional Neural Network V4-Feature Pyramid Network module;

[0014] In the Spot-Guided Aggregation module of the Adaptive Point-Guided Transformer (ASTR), introduce the Dual Cross-Attention (DCA) mechanism, reduce the cross-sampling layer from four layers to two layers, and change the scale of the coarse-grained matching from to .

[0015] Furthermore, calculating the camera internal parameter matrix of the image groups with the same focal length and different viewing angles includes:

[0016] Based on the second set of matching point pairs, estimate the fundamental matrix of the zoom camera, and calculate the projection matrix according to the fundamental matrix;

[0017] The quasi-linear method is used to solve the absolute quadric constraint corresponding to the projection matrix, and the camera intrinsic matrix of the image groups with the same focal length and different perspectives is obtained.

[0018] Further, the camera intrinsic values are obtained by solving the polynomial, including:

[0019] Based on the camera intrinsic matrix, polynomial fitting is performed to construct a system of linear equations;

[0020] The system of linear equations is solved by the least squares method to obtain the camera intrinsic values of the zoom camera.

[0021] Further, the scaling centers of the image groups with the same perspective and different focal lengths are calculated, including:

[0022] Based on each pair of matching points in the first group of matching point pairs, a vector equation is formulated;

[0023] The vector equation is solved by the least squares method to obtain the scaling center values of the image groups with the same perspective and different focal lengths.

[0024] According to one aspect of the present specification, a deep learning-based calibration-free zoom camera self-calibration system is provided, including:

[0025] An image grouping module, configured to group the captured images to obtain image groups with the same perspective and different focal lengths and image groups with the same focal length and different perspectives;

[0026] A feature point matching module, configured to use an improved lightweight adaptive point-guided transformer model to perform feature point matching on the two image groups to obtain a first group of matching point pairs and a second group of matching point pairs;

[0027] A parameter calculation module, configured to calculate the scaling centers of the image groups with the same perspective and different focal lengths based on the first group of matching point pairs by using the least squares method; calculate the camera intrinsic matrix of the image groups with the same focal length and different perspectives based on the second group of matching point pairs, and perform polynomial fitting based on the camera intrinsic matrix to solve the polynomial to obtain the camera intrinsic values;

[0028] A camera self-calibration module, configured to complete the self-calibration of the camera based on the scaling centers of the image groups with the same perspective and different focal lengths and the camera intrinsic values of the image groups with the same focal length and different perspectives.

[0029] According to one aspect of the present specification, an electronic device is provided, including a memory and a processor, the memory stores a computer program, and it is characterized in that when the processor executes the computer program, the steps of the above-mentioned deep learning-based calibration-free zoom camera self-calibration method are implemented.

[0030] According to one aspect of the present specification, there is provided a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the above-mentioned deep learning-based zoom camera self-calibration method without a calibration board are implemented.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] 1. The present invention processes two different sets of image group data separately obtained by a zoom camera, improving the data usage efficiency and reducing the computational redundancy.

[0033] 2. In view of the characteristics of large-scale change images (such as image groups with different focal lengths or viewing angles) involved in the self-calibration of zoom cameras, the present invention proposes an improved lightweight ASTR algorithm to replace the feature matching part in the conventional camera self-calibration method, which not only reduces the computational complexity, significantly improves the calibration accuracy, but also expands the applicable scenario range of the method.

[0034] 3. The present invention optimizes the polynomial fitting method in combination with the prior knowledge of the change of the internal parameters of the zoom camera, simplifies the calculation process of the zoom internal parameters while ensuring the fitting accuracy and robustness, reduces the computational complexity, and better adapts to the actual application requirements of zoom camera calibration.

[0035] 4. The present invention introduces deep learning feature matching and the least squares method into the calibration of the image scaling center, effectively improving the calibration accuracy and robustness, and reducing the deviation caused by noise and non-linear errors in the conventional method. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0037] Figure 1 It is a flowchart of the zoom camera self-calibration method according to an embodiment of the present invention;

[0038] Figure 2 It is a schematic diagram of the improved ASTR algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] The present invention provides a calibration-free zoom camera self-calibration method based on deep learning, as Figure 1 shown, including: based on the captured images, grouping to obtain image groups with the same viewing angle and different focal lengths and image groups with the same focal length and different viewing angles; using an improved lightweight adaptive point-guided transformer model to perform feature point matching on the two image groups to obtain a first group of matching point pairs and a second group of matching point pairs; based on the first group of matching point pairs, using the least squares method to calculate the scaling center of the image groups with the same viewing angle and different focal lengths; based on the second group of matching point pairs, calculating the camera internal parameter matrix of the image groups with the same focal length and different viewing angles, and performing polynomial fitting based on the camera internal parameter matrix to solve the polynomial to obtain the camera internal parameter values; based on the scaling center of the image groups with the same viewing angle and different focal lengths and the camera internal parameter values of the image groups with the same focal length and different viewing angles, complete the self-calibration of the camera.

[0041] Specifically, the main purpose of the embodiments of the present invention is to calibrate the scaling center of the zoom camera and the internal parameter matrix of the zoom camera , the matrix has the following expression:

[0042] (1)

[0043] where, , are the focal lengths of the camera in the X and Y directions, is the pixel shear coefficient, , are the positions of the principal point in the image coordinate system. According to the characteristics of the zoom camera, each parameter in the internal parameter matrix changes with the focal length can be expressed as a polynomial function, and the expression is as follows:

[0044] (2)

[0045] where, is the polynomial coefficient corresponding to the parameter , and needs to be determined through the calibration process.

[0046] Specifically, the key task of calibration is to determine the polynomial coefficients of each parameter through feature point matching and mathematical fitting , so as to achieve the precise calculation and representation of the internal parameter matrix at any focal length .

[0047] Specifically, the embodiment of the present invention also provides data acquisition based on the variable data of the fixed interval of the focal length. At least 5 shooting perspectives are preset and the camera is fixed. Each perspective is set with no less than 5 focal lengths, and the focal lengths include the maximum and minimum focal lengths within the adjustable range of the zoom camera, and the interval is fixed. The captured images are used as experimental shooting data. At each preset angle, the focal length is adjusted in turn from the minimum to the maximum for shooting. After the image acquisition of all focal lengths is completed, switch to the next perspective and repeat the operation until the image acquisition of all perspective and focal length combinations is completed. The captured images are classified and sorted. Among them, the images with different focal lengths under the same perspective are used to calculate the image scaling center of the zoom camera, and the perspective and focal length parameters of each image are recorded to ensure the integrity and accuracy of the data. The images with different perspectives under the same focal length are used to calculate the camera internal parameters at a specific focal length, and the influence of the focal length change on the camera internal parameters is analyzed.

[0048] Specifically, the embodiment of the present invention also provides using the improved ASTR algorithm for feature point matching, preprocessing the collected images, and adjusting the images to the input format required by the algorithm. Use the improved ASTR algorithm to extract image features, and the improved ASTR algorithm is as Figure 2 shown. The specific improvement is as follows: referring to the backbone network of MobileNetV4, an inverted bottleneck block (UIB) and a depthwise separable convolution module are introduced, combined with the feature pyramid network (FPN) in the original CNN module, and the convolutional feature extraction module is replaced with a lighter MobileNetV4-FPN module. In the Spot-Guided Aggregation module, a dual cross-attention mechanism (DCA) is introduced, which includes two parts: channel cross-attention (CCA) and spatial cross-attention (SCA). The number of cross-sampling layers is reduced from four layers to two layers, and the scale of coarse-grained matching is changed from to . While retaining the original core structure and optimizing the attention mechanism, the model calculation amount is significantly reduced. According to the type of the collected images, an indoor model is trained on the ScanNet dataset, and an outdoor model is trained on the MegaDepth dataset. Prune the trained model and convert it to the ONNX format. The lightweight model can be transplanted to edge devices. Combine the denoising and robustness optimization algorithms to screen the matching point pairs and output the optimized feature matching results , and the expression is as follows:

[0049] (3)

[0050] Among them, and They are feature points in homogeneous coordinate form, , which are the feature points 's abscissa and ordinate on the image plane, , which are the abscissa and ordinate of the feature point on the image plane respectively.

[0051] Specifically, the embodiments of the present invention also provide the calculation of the internal parameters for a fixed focal length with interval changes. Using the feature matching point pairs at the same focal length and different perspectives to calculate the internal parameters of the zoom camera. First, use the matching point pairs and estimate the fundamental matrix by the 8-point method , which satisfies the following conditions:

[0052] (4)

[0053] where, is a 3×3 matrix that describes the epipolar geometry relationship between two images. Then, the epipoles of the two images need to be found , and the epipole satisfies the condition of Equation (5):

[0054] (5)

[0055] where, , and are all epipolar lines in the epipolar geometry. Ideally, the point passes through all the epipolar lines . In actual operation, a point will be calculated to make 's distance sum to each line the smallest, and solved by the least squares method, as shown in Equation (6):

[0056] (6)

[0057] After calculating , , solve for the projection matrix according to Equation (7):

[0058] (7)

[0059] where, is the identity matrix of , and correspond to the camera projection matrices at two same focal lengths and different perspectives.

[0060] Specifically, a class of surfaces in three-dimensional projective space that satisfy Equation (8) is called a quadric surface. Among them, is a matrix of

[0061] (8)

[0062] Among them, refers to the four-dimensional homogeneous coordinate form of a point in three-dimensional space, is the transpose matrix of

[0063] (9)

[0064] Among them, is the absolute conic, i.e., in Equation (7), is the transpose matrix of and thus

[0065] (10)

[0066] Among them, is the matrix element value of the absolute quadric surface The projection constraint of the absolute quadric surface is the basis of the self-calibration method, and its constraint is as Equation (11):

[0067] (11)

[0068] Among them, is the anti-symmetrization operation, is the absolute conic, is the absolute quadric surface.

[0069] Specifically, to solve the constraint (11) using the quasi-linear method, first represent the independent components of and as vectors and construct an outer product matrix. The dimension of this matrix should be . This matrix represents the projection constraint and is linearly dependent among these variables with a rank of 15. Then, using more than 5 image data, linearly recover this matrix. Next, use the singular value decomposition (SVD) of and to project the matrix into a rank-1 space and decompose it into vectors has a rank of 3 and is adjusted by setting its minimum eigenvalue to zero , thus completing the solution for and . After calculating , the camera intrinsic matrix is obtained using the Cholesky decomposition method, as shown in Equation (12):

[0070] (12)

[0071] Specifically, the embodiment of the present invention also provides a functional relationship between the polynomial fitting of the intrinsic parameters with respect to the focal length. Denote the camera intrinsic parameters at different focal lengths as , is the focal length of the camera. For the accuracy of polynomial fitting, should include the maximum and minimum focal lengths that the zoom camera can adjust. A linear equation system is constructed, as shown in Equation (13):

[0072] (13)

[0073] In practical applications, according to experience, when in the highest degree is 2, in most cases, the requirements can be met.

[0074] (14)

[0075] Among them, is the design matrix, are the coefficients to be solved for the polynomial, is the vector of observed intrinsic parameter values. After obtaining the linear equation system (13), the coefficients of the five degrees of freedom of the intrinsic matrix are solved by the least squares method:

[0076] (15)

[0077] The obtained polynomial is . Using this expression, the intrinsic parameter values at any focal length within the focal length range can be predicted:

[0078] (16)

[0079] Specifically, the embodiment of the present invention also provides a method for calibrating the image scaling center based on the least squares method. Using the feature matching point pairs at the same viewing angle and different focal lengths obtained by the improved ASTR algorithm in step two, the image scaling center of the zoom camera is calculated. If there are pairs of matching points, according to the fact that two points determine a straight line equation, pairs of matching points can determine A straight-line equation. The straight-line equation determined by each pair of matching points is:

[0080] (17)

[0081] Among them, Equation (17) is the basic form of the straight-line equation on the plane, that is .

[0082] Zoom center The coordinates of are , for an ideal zoom lens, its zoom center satisfies the straight-line equation determined by any matching pair. Expressed by a mathematical expression, such as Equation (18):

[0083] (18)

[0084] Writing Equation (18) in matrix form gives:

[0085] (19)

[0086] Writing Equation (19) in a one-to-one abbreviated form gives:

[0087] (20)

[0088] Using the least squares method to solve Equation (20) to obtain the estimator of the zoom center:

[0089] (21)

[0090] Among them, Represents the estimator obtained by solving using the least squares method.

[0091] Finally, the zoom center of the zoom camera can be calculated from Equation (21).

[0092] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functions. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this actual situation, on the basis of the above embodiments, the embodiments of the present invention provide a deep learning-based uncalibrated board zoom camera self-calibration system, which is used to execute a deep learning-based uncalibrated board zoom camera self-calibration method in the above method embodiments.

[0093] The system includes: an image grouping module for grouping the captured images to obtain image groups with the same viewing angle and different focal lengths and image groups with the same focal length and different viewing angles; a feature point matching module for using an improved lightweight adaptive point-guided transformer model to perform feature point matching on the two image groups to obtain a first set of matching point pairs and a second set of matching point pairs; a parameter calculation module for calculating the scaling center of the image groups with the same viewing angle and different focal lengths by using the least squares method based on the first set of matching point pairs; calculating the camera internal parameter matrix of the image groups with the same focal length and different viewing angles based on the second set of matching point pairs, and performing polynomial fitting based on the camera internal parameter matrix to solve the polynomial to obtain the camera internal parameter values; a camera self-calibration module for completing the self-calibration of the camera based on the scaling center of the image groups with the same viewing angle and different focal lengths and the camera internal parameter values of the image groups with the same focal length and different viewing angles.

[0094] An uncalibrated board zoom camera self-calibration system based on deep learning provided by an embodiment of the present invention addresses the problems of low accuracy and stability in feature matching and calibration methods in camera self-calibration methods. By using the above-mentioned several modules, two different sets of image group data are obtained through a zoom camera and processed separately. The improved lightweight ASTR algorithm and the least squares method are proposed, which not only reduce the computational complexity and significantly improve the calibration accuracy, but also expand the applicable scenario range of the method.

[0095] Based on the same inventive concept as the foregoing embodiment, an embodiment of the present invention also provides an electronic device, including a memory and a processor. The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the uncalibrated board zoom camera self-calibration method based on deep learning proposed in the above embodiment.

[0096] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it is used to reduce the deviation caused by noise and non-linear errors in the traditional method, improve the data usage efficiency, reduce computational redundancy, and improve the calibration accuracy. The storage medium can be any non-volatile storage device such as a hard disk, a solid-state drive, a flash drive, an optical disc, etc., for storing computer program code and necessary data files. The stored computer program includes: an image grouping module, a feature point matching module, a parameter calculation module, and a camera self-calibration module.

[0097] Finally, it should be noted that the above specific embodiments are only relatively representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and can have many variations. Any simple modification, equivalent change, and modification made to the above specific embodiments based on the technical essence of the present invention should be considered to fall within the protection scope of the present invention.

Claims

1. A self-calibration method for a zoom camera without a calibration board based on deep learning, characterized in that, Including: Based on the captured images, group the images to obtain image groups with the same viewing angle but different focal lengths and image groups with the same focal length but different viewing angles; Use the improved lightweight adaptive point-guided transformer model to perform feature point matching on the two image groups to obtain the first set of matching point pairs and the second set of matching point pairs; Improvements to the lightweight adaptive point-guided transformer model include: a backbone network based on the Mobile Convolutional Neural Network V4, introducing an inverted bottleneck block and a depthwise separable convolution module; replacing the feature pyramid network module of the backbone network with a Mobile Convolutional Neural Network V4 - Feature Pyramid Network module; in the point-guided aggregation module of the adaptive point-guided transformer, introducing a dual cross-attention mechanism, reducing the number of cross-sampling layers from four to two, and changing the scale of coarse-grained matching from to ; Based on the first set of matching point pairs, use the least squares method to calculate the scaling center of the image group with the same viewing angle but different focal lengths; Based on the second set of matching point pairs, calculate the camera internal parameter matrix of the image group with the same focal length but different viewing angles, and perform polynomial fitting based on the camera internal parameter matrix to solve the polynomial to obtain the camera internal parameter values; Based on the scaling center of the image group with the same viewing angle but different focal lengths and the camera internal parameter values of the image group with the same focal length but different viewing angles, complete the self-calibration of the camera.

2. The self-calibration method of a zoom camera without a calibration board based on deep learning according to claim 1, wherein Calculating the camera internal parameter matrix of the image group with the same focal length but different viewing angles includes: Based on the second set of matching point pairs, estimate the fundamental matrix of the zoom camera, and calculate the projection matrix according to the fundamental matrix; Use the quasi-linear method to solve the absolute quadric surface constraint corresponding to the projection matrix to obtain the camera internal parameter matrix of the image group with the same focal length but different viewing angles.

3. A self-calibration method for a zoom camera without a calibration board based on deep learning according to claim 1, characterized in that, Obtaining the camera internal parameter values by solving the polynomial includes: Based on the camera internal parameter matrix, perform polynomial fitting to construct a system of linear equations; Solve the system of linear equations by the least squares method to obtain the camera internal parameter values of the zoom camera.

4. A self-calibration method for a zoom camera without a calibration board based on deep learning according to claim 1, characterized in that, Calculating the scaling center of the image group with the same viewing angle but different focal lengths includes: Based on each matching point pair in the first set of matching point pairs, formulate a vector equation; Use the least squares method to solve the vector equation to obtain the scaling center values of the image group with the same viewing angle but different focal lengths.

5. A self-calibration system for a zoom camera without a calibration board based on deep learning, characterized in that, Including: An image grouping module for grouping the captured images to obtain image groups with the same viewing angle but different focal lengths and image groups with the same focal length but different viewing angles; A feature point matching module for using the improved lightweight adaptive point-guided transformer model to perform feature point matching on the two image groups to obtain the first set of matching point pairs and the second set of matching point pairs; Improvements to the lightweight adaptive point-guided transformer model include: a backbone network based on the Mobile Convolutional Neural Network V4, introducing an inverted bottleneck block and a depthwise separable convolution module; replacing the feature pyramid network module of the backbone network with a Mobile Convolutional Neural Network V4 - Feature Pyramid Network module; in the point-guided aggregation module of the adaptive point-guided transformer, introducing a dual cross-attention mechanism, reducing the number of cross-sampling layers from four to two, and changing the scale of coarse-grained matching from to ; A parameter calculation module for calculating the scaling center of the image group with the same viewing angle but different focal lengths based on the first set of matching point pairs by using the least squares method; calculating the camera internal parameter matrix of the image group with the same focal length but different viewing angles based on the second set of matching point pairs, and performing polynomial fitting based on the camera internal parameter matrix to solve the polynomial to obtain the camera internal parameter values; A camera self-calibration module for completing the self-calibration of the camera based on the scaling center of the image group with the same viewing angle but different focal lengths and the camera internal parameter values of the image group with the same focal length but different viewing angles.

6. An electronic device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based calibration-free zoom camera self-calibration method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based calibration-free zoom camera self-calibration method according to any one of claims 1 to 4.

Images