A photographic composition assistance method based on eight degrees of freedom shooting perspective adjustment

By using an eight-degree-of-freedom shooting perspective adjustment method, an end-to-end camera parameter prediction process is constructed, which solves the problem of limited perspective adjustment range and degrees of freedom in existing technologies, realizes wide-range and precise shooting perspective adjustment, and improves the accuracy and reliability of composition assistance.

CN123118796APending Publication Date: 2026-11-17BEIJING INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611009385.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing AI composition assistance technologies have limited range and freedom of perspective adjustment in three-dimensional shooting space, making it difficult to achieve the large-scale and diverse perspective changes commonly used by professional photographers, resulting in insufficient composition optimization capabilities.

Method used

An eight-degree-of-freedom (DOF) shooting perspective adjustment method is adopted. Through steps such as feature extraction module, camera ray prediction module, shooting perspective adjustment module and optimization module, an end-to-end eight-DoF camera parameter prediction process is constructed, and a deep learning model is combined for perspective recommendation.

Benefits of technology

It enables a wide range of precise photographic perspective adjustments, enhances composition optimization capabilities, aligns with the perspective adjustment experience of professional photographers, improves the accuracy and reliability of composition assistance, and is suitable for actual photographic creation scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118796A_ABST
    Figure CN123118796A_ABST
Patent Text Reader

Abstract

This invention proposes an AI composition assistance method based on 8-DoF shooting perspective adjustment. The method includes: Step 1, the user uploads the image requiring composition assistance; Step 2, the input module preprocesses the image; Step 3, the feature extraction module extracts image features; Step 4, the camera ray prediction module predicts camera rays; Step 5, the shooting perspective adjustment module outputs the recommended shooting perspective, etc. The superior effect of this method is that it supports AI composition assistance based on 8-DoF shooting perspective adjustment, breaking the limitation of traditional two-dimensional cropping methods that can only be adjusted within the image plane, and alleviating the problem of limited freedom and range of perspective change in existing three-dimensional perspective recommendation methods. By constructing an end-to-end 8-DoF camera parameter prediction process, it achieves a wide range and accurate shooting perspective adjustment while basically maintaining the consistency of image content, thus improving the composition optimization capability and practical application value of AI composition assistance in complex real-world scenes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of computer vision and photographic aesthetics, and specifically proposes a method for adjusting the shooting perspective based on eight degrees of freedom (DoF) for AI composition assistance systems. Background Technology

[0002] AI composition assistance refers to a technique that uses artificial intelligence to automatically analyze the shooting scene, subject, and visual elements, and combines this with photographic composition theory to directly adjust the composition for users, or provide composition optimization suggestions, image evaluation, and other auxiliary information, thereby helping users obtain more aesthetically pleasing, expressive, and visually appealing images. This technology can lower the barrier to entry for photography, improve shooting efficiency, and help users obtain photographic works that better meet their aesthetic needs, showing broad application prospects in areas such as drone aerial photography and AI photography assistants. Existing AI composition assistance technologies are mainly divided into two categories. The first category is composition assistance technology based on the two-dimensional image plane, which adjusts the position of the subject and the proportion of the image through post-processing methods such as image cropping, thereby improving the composition effect of existing photos. The second category is composition assistance technology based on three-dimensional shooting space, which changes the projection relationship of objects in the scene into the image by adjusting the camera position and shooting direction, thereby achieving composition optimization during the shooting process. However, technologies based on the two-dimensional image plane can only make adjustments within the image space, and are difficult to effectively handle composition problems caused by unreasonable occlusion and unbalanced perspective relationships in three-dimensional space. In contrast, technologies based on three-dimensional shooting space can directly optimize composition at the shooting perspective level, which is more in line with the auxiliary needs in the actual photography process. However, existing methods are usually limited by the range and freedom of perspective adjustment, making it difficult to achieve the large-scale and diverse perspective changes often used by professional photographers in actual creation, and thus still have certain limitations in terms of composition optimization capabilities. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, this invention proposes an AI composition assistance method based on 8-DoF shooting angle adjustment.

[0004] The AI ​​composition assistance method based on 8-DoF shooting angle adjustment includes:

[0005] Step 1: The user uploads an image that requires composition assistance;

[0006] Step 2: The input module preprocesses the image;

[0007] Step 3: The feature extraction module extracts image features;

[0008] Step 4: The camera ray prediction module predicts the camera ray;

[0009] Step 5: The shooting angle adjustment module outputs the recommended shooting angle;

[0010] Step 6: The optimization module receives the prediction results and provides supervision for the deep learning model during training;

[0011] Step 7: The output module performs post-processing on the predicted shooting angle and outputs structured camera intrinsic and extrinsic parameter data;

[0012] Step 8: The preview module renders a preview image with adjusted viewing angle based on the output camera intrinsic and extrinsic parameters and the output image.

[0013] Furthermore, in step 1, the user uploads an image that requires composition assistance:

[0014] Users upload images for which they require photographic perspective recommendations to the input module. The input module supports images of any size in .png or .jpeg / jpg format.

[0015] Furthermore, the input module described in step 2 preprocesses the image:

[0016] The input module first normalizes the size of the original photographic image to eliminate the impact of differences in image resolution. Since the viewpoint adjustment task is sensitive to image content, the normalization process needs to convert the image into an input size that the feature extraction module can accept while preserving the original image content and composition information as much as possible. The input module adopts a scaling strategy that maintains the original aspect ratio: the image is scaled proportionally to a long side of 518 pixels, and the short side is filled with pure black areas to generate an input image with a size of 518×518 pixels.

[0017] Furthermore, the feature extraction module described in step 3 extracts image features:

[0018] The feature extraction module is used to extract geometrically perceptive visual features from the input image. It employs a 3D expert model backbone network as the feature extractor, processing the preprocessed input image... The encoding is a set of visual tags, including a camera tag and several image patch tags. The camera tag gathers the global semantic information and overall geometric information of the entire image, which can represent the scene layout, subject position and overall perspective relationship, providing a global feature basis for subsequent shooting perspective adjustment. The image patch tags retain the local detail information of the corresponding area, which can describe texture features, edge structure, local perspective relationship and spatial relationship between different areas, providing more fine-grained spatial information for subsequent geometric modeling.

[0019] Further, the camera ray prediction module described in step 4 predicts camera rays:

[0020] The camera ray prediction module establishes a direct link between the selected shooting angle and the features of local image regions, enabling the interpretation of image information. The camera ray prediction module takes the image patch markers output by the feature extraction module as input and predicts the corresponding camera ray map through a multilayer perceptron network, as shown in the following equation:

[0021] ,

[0022] in, Indicates the first The camera ray corresponding to each image patch The number of image patches is represented by each camera ray, which originates from the camera center and extends along the center of the corresponding image patch. This model demonstrates the imaging process of a pinhole camera, describing the shooting angle in a new way. With the help of camera rays, the selection of the shooting angle is directly linked to the local features of the image, improving the reliability and accuracy of the angle recommendation in a way that is more in line with the shooting practices of professional photographers.

[0023] Furthermore, the shooting angle adjustment module described in step 5 outputs a recommended shooting angle:

[0024] The shooting angle adjustment module predicts the composition direction and camera parameters corresponding to the recommended shooting angle, obtaining shooting angle adjustment results that can improve image composition quality and provide users with direct adjustment guidance. The shooting angle adjustment module takes the camera labels output by the feature extraction module as input and includes two parts: a composition direction prediction branch and a camera parameter prediction branch, which can be expressed as follows:

[0025] ,

[0026] in, ∈{0,1} indicates whether the current composition direction needs to be changed, and is a binary variable for the composition direction prediction branch prediction; This represents the camera parameters, which include intrinsic and extrinsic parameters. It is a vector output by the camera parameter prediction branch. The camera extrinsic parameters output by the shooting angle adjustment module use the COLMAP coordinate system. The axis is horizontal to the right. The axis is vertically downward. Axis horizontal forward, camera parameters From translation parameters Rotation parameters and the zoom ratio It consists of three parts, used to describe the transformation relationship from the world coordinate system to the camera coordinate system corresponding to the recommended viewpoint, under the condition that the world coordinate system coincides with the camera coordinate system corresponding to the original viewpoint. , which represents the translation distance between the camera center point corresponding to the recommended viewpoint and the current camera center point; The rotation angle from the current camera orientation to the camera orientation corresponding to the recommended viewpoint is represented by a rotation quaternion. , indicating the perspective is Axial direction and The zoom ratio in the axial direction; the viewing angle adjustment module uses the zoom ratio. To describe the changes in camera intrinsic parameters, it fully references the photographic practices of professional photographers who change lenses when selecting perspectives, and further enhances the freedom of perspective transformation.

[0027] Furthermore, the optimization module described in step 6 receives the prediction results and provides supervision for the deep learning model during training:

[0028] The optimization module receives the prediction results and provides supervision to the deep learning model during training. It only participates in computation during model training, providing supervision but not in the inference process. The optimization module receives the camera ray map output by the camera ray prediction module. Composition direction output by the shooting angle adjustment module and camera parameters And use real camera parameters to calculate the corresponding target values ​​for supervision, composition direction and camera parameters This belongs to sparse supervision, directly constraining the model's predicted composition direction, camera translation parameters, rotation parameters, and viewpoint reduction ratio to remain consistent with the true values, enabling the model to learn the overall spatial transformation relationship corresponding to the recommended viewpoint; camera ray map. This method belongs to dense supervision. It uses real camera parameters to calculate the target camera ray corresponding to each image patch and compares it with the predicted camera ray. It establishes supervision constraints at the image patch level. By combining these two supervision methods, it can improve the stability of model training and the accuracy of shooting angle prediction, and finally obtain a recommended shooting angle that is more in line with the rules of photographic composition.

[0029] Furthermore, the output module described in step 7 performs post-processing on the predicted shooting angle and outputs structured camera intrinsic and extrinsic parameter data:

[0030] The output module converts the output of the shooting angle adjustment module into a data format that can be used by downstream applications. The output module will convert the rotation quaternion. This is converted into specific rotation degrees in the three directions: pitch, roll, and yaw. Pitch represents the camera tilting to the right, roll (positive direction) represents clockwise rotation, and yaw (positive direction) represents the camera tilting upwards. The output module will then define the composition direction. Integrated into roll, the output module utilizes the view zoom ratio. Generate the predicted focal length based on the current camera's focal length. The output module converts this data into a standard camera intrinsic parameter matrix. and extrinsic parameter matrix It is supplied for direct use.

[0031] Furthermore, the preview module described in step 8 renders a preview image with adjusted viewing angle based on the output camera intrinsic and extrinsic parameters and the output image:

[0032] The preview module generates a preview image corresponding to the recommended shooting angle based on the camera parameters output by the output module and the original image input by the user. This visually demonstrates the compositional effect after the angle adjustment to the user. Specifically, the preview module reconstructs a 3D representation of the entire scene based on the input original image, renders the image under the recommended angle based on the camera parameters, repairs and optimizes the image under the new angle, and outputs the preview image corresponding to the recommended shooting angle to the user. This helps the user compare the compositional differences before and after the angle adjustment, providing an intuitive reference for shooting decisions.

[0033] Compared with the prior art in this field, the superior technical effect of the photographic composition assistance method based on eight degrees of freedom shooting angle adjustment described in this invention is as follows:

[0034] 1. The photographic composition assistance method based on eight-DoF shooting perspective adjustment described in this invention supports AI composition assistance based on eight-DoF shooting perspective adjustment. It breaks the limitation of traditional two-dimensional cropping methods that can only be adjusted within the image plane, and alleviates the problem of limited freedom and range of perspective change in existing three-dimensional perspective recommendation methods. By constructing an end-to-end eight-DoF camera parameter prediction process, it can achieve a wide range and accurate photographic perspective adjustment while basically maintaining the consistency of image content, thereby improving the composition optimization capability and practical application value of AI composition assistance in complex real-world scenes.

[0035] 2. The photographic composition assistance method based on eight-degree-of-freedom shooting perspective adjustment described in this invention has the ability to adjust perspective in accordance with professional photography experience: by using eight-DoF camera parameters to model the shooting perspective, it helps to simulate the perspective adjustment practice of professional photographers, thereby achieving multi-dimensional optimization adjustment of composition that is more in line with human experience and preferences.

[0036] 3. The photographic composition assistance method based on eight-degree-of-freedom shooting perspective adjustment described in this invention breaks through the limitations of traditional two-dimensional cropping methods by modeling the shooting perspective using eight-DoF camera parameters. It can not only focus on changes in the position of the subject, but also comprehensively consider factors such as the scale of the subject, spatial hierarchy, foreground and background relationship, and visual balance. This makes the recommended results more in line with the composition habits in the actual photographic creation process, improves the professionalism and practicality of the composition assistance results, and thus achieves the invention effect.

[0037] 4. The photographic composition assistance method based on eight-degree-of-freedom shooting perspective adjustment described in this invention significantly improves the accuracy and reliability of AI composition assistance: By supporting the system to predict the corresponding camera rays using image patch-level features, and establishing a dense geometric constraint mechanism through camera ray supervision, this mechanism further decomposes the overall shooting perspective into the spatial relationships of local regions, enabling the shooting perspective adjustment module to perceive the geometric relationships between compositional elements such as the subject, supporting elements, background, and visual guide lines in the image. Compared with methods that only rely on overall camera parameter supervision, this method can provide richer training signals, enhance the model's understanding of local perspective relationships and spatial structures, and improve the accuracy and stability of shooting perspective prediction, thereby achieving the invention's effect.

[0038] 5. The photographic composition assistance method based on eight degrees of freedom shooting perspective adjustment described in this invention has the application capability applicable to actual assistance scenarios: by outputting structured camera intrinsic and extrinsic parameter data, the recommendation result is represented as an executable camera intrinsic parameter transformation and pose adjustment, which has a good correspondence with the focusing and framing operations in photographic practice and can be directly used in downstream application scenarios.

[0039] 6. The photographic composition assistance method based on eight degrees of freedom shooting angle adjustment described in this invention has a preview module that can render a preview image after angle adjustment based on the output camera intrinsic and extrinsic parameters and the output image. It can also output a preview image under the recommended shooting angle, which further enhances the human-computer interaction attributes of the method and emphasizes its auxiliary role in the shooting process rather than making direct decisions, thereby better providing personalized recommendations based on user preferences.

[0040] 7. The photographic composition assistance method based on eight-degree-of-freedom shooting perspective adjustment described in this invention uses camera intrinsic and extrinsic parameters as a unified representation of the shooting perspective. It assists users in improving composition quality by optimizing the shooting perspective. To fully utilize the geometric information of the scene, the shooting perspective is modeled as multiple camera rays, linking the selection of the shooting perspective to local areas of the image. This enhances the perception of elements such as the main subject and supporting elements in the image, improving the reliability and accuracy of predictions in a way that better aligns with the perspective adjustment experience of professional photographers. Furthermore, considering the differences in visual expression between horizontal and vertical compositions, this invention also considers composition direction adjustment as an important dimension of composition optimization, thereby achieving more flexible and comprehensive AI composition assistance.

[0041] 8. The photographic composition assistance method based on eight degrees of freedom shooting angle adjustment described in this invention fully meets the needs of data format and human-computer interaction in practical applications, has strong practicality, and thus achieves the invention effect. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the photographic composition assistance method based on eight degrees of freedom shooting angle adjustment as described in this invention;

[0043] Figure 2 This is a detailed architecture diagram of the deep learning model for the photographic composition assistance method based on eight degrees of freedom shooting perspective adjustment as described in this invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0045] Example

[0046] Figure 1 The diagram illustrates the process of transforming the original image into camera intrinsic and extrinsic parameters corresponding to the recommended shooting angle in a specific embodiment of the present invention: The image is first preprocessed by the input module to be converted into a size that the feature extraction module can process; then, the feature extraction module extracts features and the shooting angle adjustment module outputs the composition direction and camera intrinsic and extrinsic parameters; then, the output module performs post-processing to convert it into a standard camera intrinsic and extrinsic parameter matrix format and outputs it; finally, the preview module can also output a preview image under the recommended shooting angle.

[0047] In some specific examples of the method described in this invention, step 1, which involves the user uploading an image requiring composition assistance, specifically includes:

[0048] Users upload images for which they require a photographic perspective recommendation to the 3D photographic perspective recommendation system (a photographic composition assistance method that outputs a recommended shooting angle for poorly composed images, with the expectation of obtaining images with better composition). The 3D photographic perspective recommendation system supports images of any size in .png or .jpeg / jpg format.

[0049] In a specific example of the method described in this invention, step 2, where the input module preprocesses the image, specifically includes:

[0050] The input module first normalizes the size of the original photographic image to eliminate the impact of differences in image resolution. Since the viewpoint adjustment task is sensitive to image content, the normalization process needs to convert the image into an input size that the feature extraction module can accept while preserving the original image content and composition information as much as possible. The input module adopts a scaling strategy that maintains the original aspect ratio: the image is scaled proportionally to a long side of 518 pixels, and the short side is filled with pure black areas to generate an input image with a size of 518×518 pixels.

[0051] In a specific example of the method described in this invention, step 3, the feature extraction module, extracts image features, specifically including:

[0052] The feature extraction module is used to extract geometrically perceptive visual features from the input image. It employs a 3D expert model backbone network as the feature extractor, processing the preprocessed input image... The encoding is a set of visual tags, including a camera tag and several image patch tags. The camera tag gathers the global semantic information and overall geometric information of the entire image, which can represent the scene layout, subject position and overall perspective relationship, providing a global feature basis for subsequent shooting perspective adjustment. The image patch tags retain the local detail information of the corresponding area, which can describe texture features, edge structure, local perspective relationship and spatial relationship between different areas, providing more fine-grained spatial information for subsequent geometric modeling.

[0053] In some specific examples of the method described in this invention, step 4, which involves the camera ray prediction module predicting camera rays, specifically includes:

[0054] The camera ray prediction module establishes a direct link between the selected shooting angle and the features of local image regions, enabling the interpretation of image information. The camera ray prediction module takes the image patch markers output by the feature extraction module as input and predicts the corresponding camera ray map through a multilayer perceptron network, as shown in the following equation:

[0055] ,

[0056] in, Indicates the first The camera ray corresponding to each image patch The number of image patches is represented by each camera ray, which originates from the camera center and extends along the center of the corresponding image patch. This explicitly models the imaging process of the pinhole camera, describing the shooting angle in a new way. With the help of camera rays, the selection of the shooting angle is directly linked to the local features of the image, improving the reliability and accuracy of the angle recommendation in a way that is more in line with the shooting practices of professional photographers.

[0057] In some specific examples of the method described in this invention, step 5, where the shooting angle adjustment module outputs a recommended shooting angle, specifically includes:

[0058] The shooting angle adjustment module predicts the composition direction and camera parameters corresponding to the recommended shooting angle, obtaining shooting angle adjustment results that can improve image composition quality and provide users with direct adjustment guidance. The shooting angle adjustment module takes the camera labels output by the feature extraction module as input and includes two parts: a composition direction prediction branch and a camera parameter prediction branch, expressed as follows:

[0059] ,

[0060] in, ∈{0,1} indicates whether the current composition direction needs to be changed, and is a binary variable for the composition direction prediction branch prediction; This represents the camera parameters, which include intrinsic and extrinsic parameters. It is a vector output by the camera parameter prediction branch. The camera extrinsic parameters output by the shooting angle adjustment module use the COLMAP coordinate system. The axis is horizontal to the right. The axis is vertically downward. Axis horizontal forward, camera parameters From translation parameters Rotation parameters and the zoom ratio It consists of three parts, used to describe the transformation relationship from the world coordinate system to the camera coordinate system corresponding to the recommended viewpoint, under the condition that the world coordinate system coincides with the camera coordinate system corresponding to the original viewpoint. , which represents the translation distance between the camera center point corresponding to the recommended viewpoint and the current camera center point; The rotation angle from the current camera orientation to the camera orientation corresponding to the recommended viewpoint is represented by a rotation quaternion. , indicating the perspective is Axial direction and The zoom ratio in the axial direction; the viewing angle adjustment module uses the zoom ratio. To describe the changes in camera intrinsic parameters, it fully references the photographic practices of professional photographers who change lenses when selecting perspectives, and further enhances the freedom of perspective transformation.

[0061] In some specific examples of the method described in this invention, step 6, the optimization module, receives the prediction results and provides supervision for the deep learning model during training, specifically including...

[0062] The optimization module receives the prediction results and provides supervision to the deep learning model during training. It only participates in computation during model training, providing supervision but not in the inference process. The optimization module receives the camera ray map output by the camera ray prediction module. Composition direction output by the shooting angle adjustment module and camera parameters And use real camera parameters to calculate the corresponding target values ​​for supervision, composition direction and camera parameters This belongs to sparse supervision, directly constraining the model's predicted composition direction, camera translation parameters, rotation parameters, and viewpoint reduction ratio to remain consistent with the true values, enabling the model to learn the overall spatial transformation relationship corresponding to the recommended viewpoint; camera ray map. This method belongs to dense supervision. It uses real camera parameters to calculate the target camera ray corresponding to each image patch and compares it with the predicted camera ray. It establishes supervision constraints at the image patch level. By combining these two supervision methods, it can improve the stability of model training and the accuracy of shooting angle prediction, and finally obtain a recommended shooting angle that is more in line with the rules of photographic composition.

[0063] In some specific examples of the method described in this invention, step 7, the output module, performs post-processing on the predicted shooting angle and outputs structured camera intrinsic and extrinsic parameter data, specifically including:

[0064] The output module converts the output of the shooting angle adjustment module into a data format that can be used by downstream applications. The output module will convert the rotation quaternion. This is converted into specific rotation degrees in the three directions: pitch, roll, and yaw. Pitch represents the camera tilting to the right, roll (positive direction) represents clockwise rotation, and yaw (positive direction) represents the camera tilting upwards. The output module will then define the composition direction. Integrated into roll, the output module utilizes the view zoom ratio. Generate the predicted focal length based on the current camera's focal length. The output module converts this data into a standard camera intrinsic parameter matrix. and extrinsic parameter matrix It is supplied for direct use.

[0065] In some specific examples of the method described in this invention, step 8, the preview module, renders a preview image with adjusted viewing angle based on the output camera intrinsic and extrinsic parameters and the output image, specifically including:

[0066] The preview module generates a preview image corresponding to the recommended shooting angle based on the camera parameters output by the output module and the original image input by the user. This visually demonstrates the compositional effect after the angle adjustment to the user. Specifically, the preview module reconstructs a 3D representation of the entire scene based on the input original image, renders the image under the recommended angle based on the camera parameters, repairs and optimizes the image under the new angle, and outputs the preview image corresponding to the recommended shooting angle to the user. This helps the user compare the compositional differences before and after the angle adjustment, providing an intuitive reference for shooting decisions.

[0067] Figure 2 The specific architecture of the deep learning model of the present invention is shown: First, the features of the input image are transformed into a camera label and several image patch labels through the feature extraction module. Then, the camera's composition direction and camera intrinsic and extrinsic parameters are output through the shooting view prediction module. The camera ray is predicted through the camera ray prediction module and used for dense supervision during training.

[0068] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An AI composition assistance method based on 8-DoF shooting perspective adjustment, comprising: Step 1: The user uploads an image that requires composition assistance; Step 2: The input module preprocesses the image; Step 3: The feature extraction module extracts image features; Step 4: The camera ray prediction module predicts the camera ray; Step 5: The shooting angle adjustment module outputs the recommended shooting angle; Step 6: The optimization module receives the prediction results and provides supervision for the deep learning model during training; Step 7: The output module performs post-processing on the predicted shooting angle and outputs structured camera intrinsic and extrinsic parameter data; Step 8: The preview module renders a preview image with adjusted viewing angle based on the output camera intrinsic and extrinsic parameters and the output image.

2. The AI ​​composition assistance method based on 8-DoF shooting angle adjustment as described in claim 1, characterized in that, Step 1, which involves the user uploading an image that requires composition assistance, specifically includes: Users upload images for which they require photographic perspective recommendations to the input module. The input module supports images of any size in .png or .jpeg / jpg format.

3. The AI ​​composition assistance method based on 8-DoF shooting angle adjustment as described in claim 1, characterized in that, Step 2 involves the input module preprocessing the image, specifically including: The input module first normalizes the size of the original photographic image to eliminate the impact of differences in image resolution. Since the viewpoint adjustment task is sensitive to image content, the normalization process needs to convert the image into an input size that the feature extraction module can accept while preserving the original image content and composition information as much as possible. The input module adopts a scaling strategy that maintains the original aspect ratio: the image is scaled proportionally to a long side of 518 pixels, and the short side is filled with pure black areas to generate an input image with a size of 518×518 pixels.

4. The AI ​​composition assistance method based on 8-DoF shooting angle adjustment as described in claim 1, characterized in that, Step 3, the feature extraction module, extracts image features, specifically including: The feature extraction module is used to extract geometrically perceptive visual features from the input image. It employs a 3D expert model backbone network as the feature extractor, processing the preprocessed input image... The encoding is a set of visual tags, including a camera tag and several image patch tags. The camera tag gathers the global semantic information and overall geometric information of the entire image, which can represent the scene layout, subject position and overall perspective relationship, providing a global feature basis for subsequent shooting perspective adjustment. The image patch tags retain the local detail information of the corresponding area, which can describe texture features, edge structure, local perspective relationship and spatial relationship between different areas, providing more fine-grained spatial information for subsequent geometric modeling.

5. The AI ​​composition assistance method based on 8-DoF shooting angle adjustment as described in claim 1, characterized in that, Step 4, the camera ray prediction module, predicts camera rays, specifically including: The camera ray prediction module establishes a direct link between the selected shooting angle and the features of local image regions, enabling the interpretation of image information. The camera ray prediction module takes the image patch markers output by the feature extraction module as input and predicts the corresponding camera ray map through a multilayer perceptron network, as shown in the following equation: , in, Indicates the first The camera ray corresponding to each image patch The number of image patches is indicated. Each camera ray originates from the center of the camera and extends along the center of the corresponding image patch, demonstrating the imaging process of the pinhole camera.

6. The AI ​​composition assistance method based on 8-DoF shooting angle adjustment as described in claim 1, characterized in that, Step 5, the shooting angle adjustment module, outputs a recommended shooting angle, specifically including: The shooting angle adjustment module predicts the composition direction and camera parameters corresponding to the recommended shooting angle, obtaining shooting angle adjustment results that can improve image composition quality and provide users with direct adjustment guidance. The shooting angle adjustment module takes the camera labels output by the feature extraction module as input and includes two parts: a composition direction prediction branch and a camera parameter prediction branch, which can be expressed as follows: , in, ∈{0,1} indicates whether the current composition direction needs to be changed, and is a binary variable for the composition direction prediction branch prediction; This represents the camera parameters, which include intrinsic and extrinsic parameters. It is a vector output by the camera parameter prediction branch. The camera extrinsic parameters output by the shooting angle adjustment module use the COLMAP coordinate system. The axis is horizontal to the right. The axis is vertically downward. Axis horizontal forward, camera parameters From translation parameters Rotation parameters and the zoom ratio It consists of three parts, used to describe the transformation relationship from the world coordinate system to the camera coordinate system corresponding to the recommended viewpoint, under the condition that the world coordinate system coincides with the camera coordinate system corresponding to the original viewpoint. , which represents the translation distance between the camera center point corresponding to the recommended viewpoint and the current camera center point; The rotation angle from the current camera orientation to the camera orientation corresponding to the recommended viewpoint is represented by a rotation quaternion. , indicating the perspective is Axial direction and The zoom ratio in the axial direction; the viewing angle adjustment module uses the zoom ratio. To describe the changes in camera intrinsic parameters.

7. The AI ​​composition assistance method based on eight-DoF shooting angle adjustment as described in claim 1, characterized in that... Furthermore, the optimization module described in step 6 receives the prediction results and provides supervision for the deep learning model during training: The optimization module receives the prediction results and provides supervision to the deep learning model during training. It only participates in computation during model training, providing supervision but not in the inference process. The optimization module receives the camera ray map output by the camera ray prediction module. Composition direction output by the shooting angle adjustment module and camera parameters And use real camera parameters to calculate the corresponding target values ​​for supervision, composition direction and camera parameters This belongs to sparse supervision, directly constraining the model's predicted composition direction, camera translation parameters, rotation parameters, and viewpoint reduction ratio to remain consistent with the true values, enabling the model to learn the overall spatial transformation relationship corresponding to the recommended viewpoint; camera ray map. This method belongs to dense supervision. It uses real camera parameters to calculate the target camera ray corresponding to each image patch and compares it with the predicted camera ray. It establishes supervision constraints at the image patch level. By combining these two supervision methods, it can improve the stability of model training and the accuracy of shooting angle prediction, and finally obtain a recommended shooting angle that is more in line with the rules of photographic composition.

8. The AI ​​composition assistance method based on 8-DoF shooting angle adjustment as described in claim 1, characterized in that, Step 7 describes the output module that performs post-processing on the predicted shooting angle and outputs structured camera intrinsic and extrinsic parameter data, specifically including: The output module converts the output of the shooting angle adjustment module into a data format that can be used by downstream applications. The output module will convert the rotation quaternion. This is converted into specific rotation degrees in the three directions: pitch, roll, and yaw. Pitch represents the camera tilting to the right, roll (positive direction) represents clockwise rotation, and yaw (positive direction) represents the camera tilting upwards. The output module will then define the composition direction. Integrated into roll, the output module utilizes the view zoom ratio. Generate the predicted focal length based on the current camera's focal length. The output module converts this data into a standard camera intrinsic parameter matrix. and extrinsic parameter matrix .

9. The AI ​​composition assistance method based on 8-DoF shooting angle adjustment as described in claim 1, characterized in that, Step 8, the preview module, renders a preview image with adjusted viewing angle based on the output camera intrinsic and extrinsic parameters and the output image, specifically including: The preview module generates a preview image corresponding to the recommended shooting angle based on the camera parameters output by the output module and the original image input by the user. This visually demonstrates the compositional effect after the angle adjustment to the user. Specifically, the preview module reconstructs a 3D representation of the entire scene based on the input original image, renders the image under the recommended angle based on the camera parameters, repairs and optimizes the image under the new angle, and outputs the preview image corresponding to the recommended shooting angle to the user. This helps the user compare the compositional differences before and after the angle adjustment, providing an intuitive reference for shooting decisions.