An intelligent composition guiding method and device based on multi-view image analysis

Abstract

This invention discloses an intelligent composition guidance method and device based on multi-view image analysis, relating to the fields of computer vision and intelligent photography technology. The method includes: acquiring two or more original images from different perspectives captured by the user; performing subject extraction, scene classification, interference element marking, spatial information construction, and light parameter acquisition on the original images; comparing camera position parameters of multiple original images to determine perspective differences and image defects; scoring the composition of multiple original images to derive the optimal shooting angle and composition scheme; generating visual composition guidelines, layered verbal guidance, and real-time calibration prompts to guide the user in locating the optimal shooting angle and completing the shot. This invention upgrades AI analysis from "post-process composition optimization" to "pre-process scenario-based guidance," achieving low-cost implementation based on open-source models and algorithms, significantly lowering the barrier to photographic composition, increasing the success rate of ordinary users' shots, and is applicable to various imaging devices such as mobile phones and cameras.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, artificial intelligence and intelligent photography technology, specifically an intelligent composition guidance method and device based on multi-view image analysis, applicable to scenarios such as consumer cameras, smartphones, and portable imaging devices. Background Technology

[0002] With the widespread adoption of smart imaging devices, more and more ordinary users are relying on the built-in intelligent composition functions of these devices to improve their shooting results. Existing intelligent composition technologies are mainly divided into two categories: one is real-time framing and composition assistance, which uses guidelines such as the rule of thirds and the golden ratio to overlay on the viewfinder, allowing users to adjust the camera position manually. This method relies on the user's understanding of composition rules and has a relatively high learning curve. The other is post-processing composition optimization, which uses algorithms to crop and rotate the captured image to optimize the composition. However, this method suffers from issues such as image quality loss, inability to adapt to the scene (e.g., obstruction, changes in lighting), and limited generalization of suggestions.

[0003] In existing technologies, intelligent composition relies on real-time recognition or post-processing of a single image. AI algorithms can only provide suggestions based on a single perspective and cannot deduce the optimal camera position by combining information from multiple perspectives on site. This easily leads to problems such as misjudgment of the subject, deviation in horizon correction, and untimely avoidance of interfering elements. It is difficult to meet the core needs of ordinary users to "quickly find the best shooting angle without professional knowledge," which is specifically reflected in the following aspects: (1) Limitations of perspective: Existing technologies are based on single-perspective image analysis, which cannot capture multi-dimensional scene information on site. The composition suggestions are easily detached from the actual shooting environment and have poor adaptability to irregular scenes and complex backgrounds. (2) High user threshold: Real-time auxiliary lines only provide rule references. Users need to judge the camera position and adjust the direction independently. Post-shooting optimization cannot make up for the angle deviation during shooting. All of these depend on the user having basic composition knowledge. (3) Poor balance between image quality and effect: Post-cropping optimization will result in loss of image pixels and cannot fundamentally avoid problems such as cluttered background and subject offset during shooting, resulting in limited composition effect; (4) Insufficient algorithm adaptability: The existing model is not very targeted to different scenes (portraits, landscapes, buildings), and it is difficult to dynamically adjust the composition rules according to the actual situation on site, resulting in a high misjudgment rate; (5) Lack of light adaptation: Ignoring the influence of ambient light on the composition effect, and failing to optimize the shooting angle by combining parameters such as the position and intensity of the light source, resulting in insufficient image quality. Summary of the Invention

[0004] The purpose of this invention is to overcome the defects and shortcomings of existing technologies and provide an intelligent composition guidance method and device based on multi-view image analysis. By collecting two or more multi-view images of the user's scene, and combining AI multi-dimensional analysis, 3D spatial modeling, lighting adaptation optimization and visualization guidance, a closed loop of "image acquisition and analysis → angle derivation → precise guidance" is realized, which reduces the user's operation threshold, improves the accuracy and on-site adaptability of composition guidance, avoids the loss of image quality in post-processing, and takes into account both composition aesthetics and lighting texture.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A smart composition guidance method based on multi-view image analysis, characterized by the following steps: S1. Through the visual guidance interface of the imaging device, guide the user to collect two or more original images from different perspectives around a single subject. The different perspectives cover at least one of the orientation, height, and distance dimensions. The original shooting parameters are completely preserved during the acquisition process, and no image preprocessing is performed. S2. Call the target detection model, scene classification model, 3D composition analysis unit and light source acquisition unit to synchronously analyze each of the original images and output core data. The core data includes the subject contour coordinates, subject center position, subject pixel ratio, subject key feature points, scene type label, interference element information (interference element position and area, etc.), spatial data composed of spatial depth parameters based on the 3D point cloud model and the spatial position relationship between the subject and the environment, and light data composed of light source light data parameters and light defect area markings. S3. Based on the geometric analysis algorithm, compare the perspective transformation relationship of multiple original images, derive the camera position parameters corresponding to each original image, detect the screen tilt angle and quantify the posture deviation, and combine the interference element information, spatial data and light data in the core data to complete the identification of screen defect type and defect level assessment, and output the screen defect assessment result. S4. Construct a two-dimensional composition scoring system. Combining the scene type label, spatial depth parameters, light source parameters, and image defect evaluation results output in step S2, perform comprehensive scoring and ranking on multiple original images. Select the image with the highest basic score as the benchmark viewpoint. Combine the camera position parameters to derive 1-2 optimal shooting viewpoints and corresponding composition schemes. The composition schemes include optimal composition ratio, subject positioning point, spatial depth optimization suggestions, and optimal light source adaptation suggestions. S5. Generate visual composition guide lines, layered conversational camera position adjustment guidance, and real-time calibration prompts. Overlay the visual composition guide lines onto the viewfinder of the imaging device to guide the user to adjust the camera position to the optimal shooting angle. Once the camera position is matched to the standard, trigger the calibration prompt signal to complete the shooting.

[0006] Furthermore, in step S1, the original shooting parameters include pixel resolution, lens focal length, shooting timestamp, and imaging device attitude parameters.

[0007] Furthermore, in step S2, the target detection model uses the open-source model of YOLOv8 or above, which has been fine-tuned and optimized for the subject recognition task; the scene classification model is based on the ResNet series network pre-trained model, and is fine-tuned again by combining the MITPlaces365 dataset and the photography scene-specific dataset; the 3D composition analysis unit is built based on the PointPillars lightweight model; the light source light acquisition unit is implemented based on OpenCV brightness analysis and color temperature detection algorithms; the light source light parameters include light source type, light source direction, light intensity and color temperature range; the light defect areas include overexposed highlight areas, underexposed shadow areas and backlit subject blur areas.

[0008] Furthermore, in step S3, the geometric analysis algorithm includes a homography transformation algorithm and a Hough transformation algorithm. The homography transformation algorithm is used to calculate the perspective transformation matrix between images, and then derive the camera position parameters. The camera position parameters include azimuth offset angle, height difference, and distance difference, with accuracy controlled within ±1cm and ±1°. The image defect types include subject offset, background clutter, horizon tilt, subject occlusion, light interference, and spatial layer blurring.

[0009] Furthermore, in step S4, the dual-dimensional composition scoring system includes a rule-based scoring dimension and a deep learning scoring dimension. The rule-based scoring dimension is based on classic composition rules, dynamically adjusting the weight ratio of each composition rule in conjunction with scene type labels, while also incorporating lighting adaptability scoring. The deep learning scoring dimension uses the PhotoScore open-source pre-trained model, which is fine-tuned and optimized through a composition aesthetics scoring task, outputting a picture composition quality score of 0-10. The optimal composition ratios include 16:9, 4:3, 1:1, and 9:16.

[0010] Furthermore, in step S5, the visual composition guide lines include the rule of thirds, golden ratio point markers, and a level, along with a dynamic dotted frame to mark the position, boundary outline, and size proportion of the subject in the optimal frame. The dynamic dotted frame adaptively scales with the composition ratio. Both the visual composition guide lines and the dynamic dotted frame are dynamically adjusted in real time by moving randomly, with color changes indicating the degree of deviation. The layered, conversational camera position adjustment guidance includes basic adjustment prompts, spatial adaptation prompts, obstacle avoidance prompts, and lighting optimization prompts. The real-time calibration prompts are triggered by at least one of sound signals, icon flashing signals, or vibration signals, including split-screen preview comparison, directional guidance icons, and lighting status indicators.

[0011] A smart composition guidance device based on multi-view image analysis, characterized in that it comprises an image acquisition module, an image parsing module, a camera position analysis module, a composition derivation module, a visualization guidance module, and a storage module, which are sequentially and communicatively connected. Each module works collaboratively to implement the smart composition guidance method based on multi-view image analysis, wherein: Image acquisition module: Provides a minimalist visual guide interface to guide users in acquiring two or more original images from different perspectives, fully preserving the original shooting parameters, and prohibiting image preprocessing operations during the acquisition process; Image analysis module: Built-in finely tuned and optimized target detection model, scene classification model, 3D composition analysis unit and light source acquisition unit, used to synchronously analyze the original image and output standardized core data including subject contour coordinates, subject center position, subject pixel ratio, subject key feature points, scene type label, interference element information (interference element position and area, etc.), spatial data and light data. Camera position analysis module: Built-in homography transformation algorithm and Hough transform algorithm, used to calculate the transformation relationship of multiple image viewpoints, derive camera position parameters, detect the tilt angle of the screen, and complete the image defect assessment by combining interference element information (position and area of ​​interference elements, etc.); Composition derivation module: used to construct a two-dimensional composition scoring system, realize the comprehensive scoring and ranking of original images, select the benchmark viewpoint and derive the optimal shooting viewpoint and composition scheme, and output the optimal composition parameters; Visualization guidance module: Used to generate visual composition guide lines, dynamic dashed frames, layered conversational camera position adjustment guidance and calibration prompts based on optimal composition parameters. It provides split-screen preview comparison and dynamic guide icon interaction functions, and improves the accuracy of subject positioning through dual guidance. Storage module: Used to store network pre-trained models, original images acquired by the image acquisition module, core data parsed by the image parsing module, optimal camera position data analyzed by the camera position analysis module, optimal composition parameters derived by the composition derivation module, and layered, conversational camera position adjustment guidance and dynamic guidance data output by the visualization guidance module.

[0012] Furthermore, the device can be deployed in smartphones, consumer-grade mirrorless cameras, and portable imaging devices. The core models of the device are all exported in ONNX format, adapted to local ISP image processors or CPUs, and support millisecond-level analysis response without relying on network transmission.

[0013] Furthermore, the visualization guidance module also has a customizable visualization composition guide line function and an interactive mode switching function, which can turn on or off the corresponding type of visualization composition guide line according to user needs, adapting to the shooting habits of different users.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention relies on two or more original images taken by the user from different perspectives on-site (with a very simple operation threshold). Through AI algorithms, it completes "scene information extraction → multi-view comparison analysis → aesthetic composition derivation → visualization guidance generation", transforming the traditional "passive optimization" into "active scene-based guidance". It can combine on-site multi-view data, 3D spatial information and lighting parameters to make the composition suggestions more in line with the actual shooting environment. At the same time, through layered verbal guidance and dynamic visual composition auxiliary lines, users without professional knowledge can directly follow the operation, balancing accuracy, ease of use and a very simple operation experience.

[0015] Specifically: 1. High composition accuracy: It integrates multi-view, 3D space and lighting data to solve the problems of limited viewpoint and lack of lighting adaptation, dynamically adapts to scene rules, and reduces the misjudgment rate; 2. Extremely low barrier to entry: 2 or more images + minimalist guidance, layered conversational instructions and dynamic interaction, no professional knowledge required, ordinary users can get started quickly; 3. Lossless image quality optimization: Pre-angle guidance replaces post-cropping, preserving the original image quality and balancing composition effects with high-definition requirements; 4. Low implementation cost: Based on open source models and algorithms, it supports local offline operation, is easy to integrate into existing devices, and has broad market application prospects; 5. Wide scene adaptability: Covers a variety of photography scenes, dynamically adjusts composition and lighting weights, and balances compositional aesthetics and image quality. Attached Figure Description

[0016] Figure 1 This is a flowchart of an intelligent composition guidance method based on multi-view image analysis according to the present invention.

[0017] Figure 2 This is a structural principle block diagram of an intelligent composition guidance device based on multi-view image analysis according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 A smart composition guidance method based on multi-view image analysis is proposed, with the following specific steps: S1. Acquisition of raw images from multiple perspectives: The imaging device's acquisition module guides users to capture two or more original images from different perspectives around the subject using a minimalist visual interface. The guidance prompts users to focus on key differences (such as "shooting from the front + side, or eye level + slightly upward angle"), reducing operational costs. During the acquisition process, the original image parameters (pixel resolution, lens focal length, shooting timestamp, device posture parameters) are fully preserved without any preprocessing, ensuring the integrity of scene and lighting information and simplifying user operation steps to the greatest extent.

[0020] S2. Image content analysis and multi-dimensional element extraction: The system utilizes object detection models, scene classification models, 3D composition analysis units, and light source acquisition units to synchronously and parallelly analyze each original image, outputting standardized core data containing dimensions of subject, scene, space, and lighting. This provides comprehensive data for subsequent analysis, specifically including: (1) Subject extraction: Using the YOLOv8 and above open source model (fine-tuned by the subject recognition task), the core subject of the picture (people, animals, buildings, still life, etc.) is identified, and the subject outline coordinates, center position, pixel ratio and key feature points are output to accurately locate the position of the subject in the picture; (2) Scene classification: Based on the ResNet series pre-trained models (combined with the MIT Places365 dataset and photography-specific dataset for secondary fine-tuning), determine the image scene type (portrait, landscape, architecture, food, street photography, etc.), and simultaneously match the composition rule weights of the corresponding scene (such as portraits emphasizing the rule of thirds and negative space, and landscapes emphasizing the horizon ratio). (3) 3D spatial information construction: The 3D composition analysis unit relies on the open source PointPillars lightweight model, and constructs a lightweight 3D point cloud model based on feature point matching of two or more images. It restores the spatial position relationship and depth of the subject and the surrounding environment, outputs spatial depth parameters, and adapts to shooting scenes that need to highlight the sense of space. (4) Light Parameter Acquisition: The light acquisition unit uses OpenCV brightness analysis and color temperature detection algorithms to analyze the image pixel brightness distribution, light and shadow gradient and highlight / shadow areas, and acquire core light parameters (light source type: natural light / artificial light; light source direction: front light / back light / side light / top light; light intensity, color temperature range), evaluate the illumination effect of light on the subject, and mark the light defect areas (overexposed highlight areas, underexposed shadow areas, backlit subject blur areas).

[0021] S3. Multi-view camera position analysis and image defect identification: Based on geometric analysis algorithms, the multidimensional data output from step S2 is integrated and analyzed to locate camera position differences and image defects, and to clarify optimization directions, specifically including: (1) Derivation of camera position parameters: The perspective transformation matrix between images is calculated by homography transformation algorithm, and the camera position parameters (azimuth offset angle, height difference, distance difference) of each original image are derived. The accuracy is controlled within ±1cm and ±1° to clarify the positional differences between multiple perspectives. (2) Image posture correction analysis: The Hough transform algorithm is used to detect the horizon and horizontal lines of the image, quantify the tilt angle and posture deviation of the image, and provide data support for subsequent composition correction; (3) Comprehensive assessment of image defects: Combining interference element marking, 3D spatial data and light parameters, comprehensively identify the types of image defects (subject offset, background clutter, horizon tilt, subject occlusion, spatial layer blur, overexposure / underexposure, color temperature imbalance, etc.), output the defect level, and simultaneously mark the optimization direction of spatial and light dimensions to provide targeted basis for composition derivation.

[0022] S4. Composition scheme derivation and scoring ranking: A two-dimensional composition scoring system is constructed, integrating spatial information and lighting parameters to derive the optimal solution, ensuring consistency with the two-dimensional scoring and lighting adaptation rules in the claims. Specifically, this includes: (1) Two-dimensional scoring: ① Regular scoring: Based on classic rules such as the rule of thirds, golden ratio, symmetrical composition, and leading lines, the weights are dynamically adjusted according to the scene type, while incorporating the light adaptability score (light uniformity, proportion of defect-free light, and the shaping effect of light on the subject); ② Deep learning scoring: The PhotoScore open-source pre-trained model (fine-tuned by the composition aesthetics task) is used to output a picture aesthetic quality score of 0-10. (2) Derivation of the optimal solution: The two scoring results are weighted and fused, multiple original images are sorted and a reference angle is selected. Combined with camera position parameters, 3D spatial depth and light adaptation requirements, 1-2 optimal shooting angles are derived, the optimal composition ratio (16:9, 4:3, 1:1, 9:16) and the subject positioning point are determined, and spatial depth optimization suggestions and best light adaptation suggestions are output simultaneously (such as portraits with priority front lighting / soft side lighting, and landscapes with priority side lighting to enhance the layers).

[0023] S5. Visual guidance and shooting instructions: Multi-dimensional guidance content is generated based on the optimal graph composition scheme to complete the "derivation-guidance" closed loop, specifically including: (1) Visualized composition guide lines and dynamic dotted frame guidance: The guide lines (the rule of thirds, the golden ratio mark, and the level) corresponding to the composition rules are superimposed on the viewfinder. At the same time, a dynamic dotted frame is generated to accurately mark the position, boundary outline and size ratio of the subject in the optimal picture. The dotted frame is adaptively scaled according to the composition ratio. Both the visualized composition guide lines and the dynamic dotted frame are moved randomly and dynamically adjusted in real time. The degree of deviation is indicated by the color change from red to yellow to green (red indicates too large deviation, yellow indicates small deviation, and green indicates perfect match), which intuitively guides the user to align the position of the subject. (2) Layered conversational guidance for camera position adjustment: The guidance is divided into layers according to “basic adjustment (orientation / height / distance values) → spatial adaptation → obstacle avoidance → lighting optimization”, which clarifies the key points of lighting adjustment and spatial adaptation; (3) Real-time calibration prompts: Provide split-screen preview comparison (current screen and optimal simulation screen, including dual comparison of lighting effects and subject dotted frame position), directional guidance icons (arrow flashing frequency corresponds to the deviation size), and lighting status indicators. After the camera position and dynamic dotted frame and visual composition auxiliary lines are fully matched and meet the standards, sound / flash / vibration prompts are triggered, and the "Shooting ready" indicator is displayed to guide the user to complete the shooting and ensure that the subject position, composition ratio, and lighting adaptation are all optimal.

[0024] See Figure 2 A smart composition guidance device based on multi-view image analysis includes an image acquisition module, an image parsing module, a camera position analysis module, a composition derivation module, a visualization guidance module, and a storage module, which are sequentially connected in communication. Each module works collaboratively to implement the aforementioned smart composition guidance method based on multi-view image analysis, wherein: (1) Image acquisition module: Provides a minimalist visual guide interface to guide users to acquire two or more original images from different perspectives, fully preserve the original shooting parameters, prohibit preprocessing, and simplify operation; (2) Image parsing module: Built-in finely tuned and optimized target detection, scene classification model, 3D composition analysis unit (PointPillars model), and light source acquisition unit (OpenCV algorithm), outputting multi-dimensional standardized core data; (3) Camera position analysis module: Built-in homography transformation algorithm and Hough transformation algorithm to complete camera position derivation, attitude correction and comprehensive defect assessment; (4) Composition derivation module: Construct a two-dimensional scoring system to derive the optimal composition scheme that takes into account aesthetics, spatial and lighting adaptation; (5) Visualization guidance module: Generates visual composition guide lines, dynamic dotted frames (marking the position and outline of the main subject), layered guidance and interactive prompts. Through the dual guidance of dynamic dotted frames and visual composition guide lines, it realizes multi-dimensional guidance for main subject positioning, composition alignment and real-time calibration; (6) Storage module: Stores the network pre-trained model, fine-tuning parameters, acquired original images and images with optimal composition, and supports local offline operation.

[0025] Hardware deployment environment of this invention: The device of this invention can be deployed in smartphones, consumer mirrorless cameras, and portable imaging devices. The hardware requirements include: an image sensor (to acquire raw images), a processor (ISP / CPU, supporting local model operation), a display module (to display auxiliary lines and guidance content), a storage module (to store models and images), and a vibration / sound module (to trigger calibration prompts). No additional hardware is required, and it can be adapted to existing devices at low cost.

[0026] Implementation process of this invention: (1) Model deployment: Export YOLOv8, ResNet fine-tuning model, PhotoScore model, and PointPillars lightweight model to ONNX format. The light source acquisition unit is integrated based on the OpenCV algorithm and deployed locally on the device to achieve millisecond-level analysis response without the need for network connection. (2) User operation process: ① The user turns on the intelligent composition guidance function, and the interface prompts "take 2 photos from different angles (such as front + side)"; ② The user takes the picture according to the prompt, and the device automatically uploads the original image to the analysis module; ③ The device completes the analysis, camera position analysis, and composition derivation within 1-2 seconds and generates guidance content; ④ The user adjusts the camera position according to the layered guidance, and the dynamic auxiliary lines of the viewfinder and interactive elements provide real-time feedback on the deviation; ⑤ After the target is met, a multi-dimensional prompt is triggered, the user presses the shutter to complete the shooting, and the device saves the original image and the optimal composition.

[0027] Example of scenario adaptation for this invention: Example 1: Portrait Shooting Scene The user takes two photos (a close-up shot with backlighting and a distant shot from the side at eye level). The image analysis module identifies the subject as a person and determines the portrait scene. The 3D unit constructs a point cloud model to analyze the depth, and the lighting unit detects underexposure due to backlighting on the front and soft side lighting on the side. After dual-dimensional scoring, the side view receives a higher overall score, leading to the optimal guidance: "Move one step to the right, raise the phone 5cm, align the person's face with the intersection of the right thirds line, and maintain the side lighting angle to enhance the outline." The interface simultaneously displays green auxiliary lines and a dotted outline of the person, with directional arrows guiding alignment. A vibration notification is given upon achieving the target, resulting in a portrait image with accurate subject positioning, clear outline, and even lighting.

[0028] Example 2: Landscape photography scene The user takes two photos (a level panorama and a shot of the sky from below). The analysis module determines the landscape scene, and the 3D unit analysis shows that the level panorama has better depth. The lighting unit detects that the level panorama is side-lit (with rich mountain layers) and the shot of the sky from below is overexposed. The derivation guides: "Keep the view level, rotate 2° to the left to correct the horizon, the vertical ratio of the image is 1:2 (sky: ground), and the side-lit angle preserves the details of the mountains." The interface overlays a level and the rule of thirds, provides a left-turn arrow prompt, and offers a split-screen preview of the correction effect. Once the target is met, an audio prompt is given, resulting in a landscape photo with a well-composed and appropriately lit image.

[0029] This invention features a minimalist operation design, requiring only the acquisition of two or more images to complete the analysis. Coupled with a minimalist guide, it significantly reduces the user's operating threshold and is suitable for fast-paced shooting scenarios. This invention employs multi-dimensional data fusion, integrating multi-view data, 3D spatial information, and lighting parameters for the first time, solving the problems of limited perspective and lack of lighting adaptation in existing technologies, and providing more accurate composition suggestions; This invention upgrades "post-event optimization" to "pre-event guidance" through proactive guidance logic, avoiding image quality loss. At the same time, it balances professionalism and ease of use through layered guidance and dynamic interaction. This invention is open source and low-cost to implement. The core model and algorithm are developed based on open source resources, can be exported to ONNX format for local operation, do not require network connection, and are easy to integrate into various imaging devices.

[0030] Although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0031] Therefore, the above description is only a preferred embodiment of this application and is not intended to limit the scope of this application; that is, all equivalent modifications made in accordance with the scope of the claims of this application shall be within the protection scope of the claims of this application.

Claims

1. A smart composition guidance method based on multi-view image analysis, characterized in that: Specifically, the following steps are included: S1. Through the visual guidance interface of the imaging device, guide the user to collect two or more original images from different perspectives around a single subject. The different perspectives cover at least one of the orientation, height, and distance dimensions. The original shooting parameters are completely preserved during the acquisition process, and no image preprocessing is performed. S2. Call the target detection model, scene classification model, 3D composition analysis unit and light source acquisition unit to synchronously analyze each of the original images and output core data. The core data includes the subject contour coordinates, subject center position, subject pixel ratio, subject key feature points, scene type label, interference element information, spatial data composed of spatial depth parameters based on the 3D point cloud model and the spatial position relationship between the subject and the environment, and light data composed of light source light data parameters and light defect area markings. S3. Based on the geometric analysis algorithm, compare the perspective transformation relationship of multiple original images, derive the camera position parameters corresponding to each original image, detect the screen tilt angle and quantify the posture deviation, and combine the interference element information, spatial data and light data in the core data to complete the identification of screen defect type and defect level assessment, and output the screen defect assessment result. S4. Construct a two-dimensional composition scoring system. Combining the scene type label, spatial depth parameters, light source parameters, and image defect evaluation results output in step S2, perform comprehensive scoring and ranking on multiple original images. Select the image with the highest basic score as the benchmark viewpoint. Combine the camera position parameters to derive 1-2 optimal shooting viewpoints and corresponding composition schemes. The composition schemes include optimal composition ratio, subject positioning point, spatial depth optimization suggestions, and optimal light source adaptation suggestions. S5. Generate visual composition guide lines, layered conversational camera position adjustment guidance, and real-time calibration prompts. Overlay the visual composition guide lines onto the viewfinder of the imaging device to guide the user to adjust the camera position to the optimal shooting angle. Once the camera position is matched to the standard, trigger the calibration prompt signal to complete the shooting.

2. The intelligent composition guidance method based on multi-view image analysis according to claim 1, characterized in that: In step S1, the original shooting parameters include pixel resolution, lens focal length, shooting timestamp, and imaging device attitude parameters.

3. The intelligent composition guidance method based on multi-view image analysis according to claim 1, characterized in that: In step S2, the target detection model uses the open-source model YOLOv8 or later, fine-tuned and optimized for the subject recognition task; the scene classification model is based on the ResNet series network pre-trained model, and fine-tuned again by combining the MIT Places365 dataset and a photography scene-specific dataset; the 3D composition analysis unit is built based on the PointPillars lightweight model; the light source acquisition unit is implemented based on OpenCV brightness analysis and color temperature detection algorithms; the light source parameters include light source type, light source direction, light intensity, and color temperature range; the light defect areas include overexposed highlight areas, underexposed shadow areas, and backlit subject blur areas.

4. The intelligent composition guidance method based on multi-view image analysis according to claim 1, characterized in that: In step S3, the geometric analysis algorithm includes a homography transformation algorithm and a Hough transformation algorithm. The homography transformation algorithm is used to calculate the perspective transformation matrix between images, and then derive the camera position parameters. The camera position parameters include azimuth offset angle, height difference, and distance difference, with accuracy controlled within ±1cm and ±1°. The image defect types include subject offset, background clutter, horizon tilt, subject occlusion, light interference, and spatial layer blur.

5. The intelligent composition guidance method based on multi-view image analysis according to claim 1, characterized in that: In step S4, the dual-dimensional composition scoring system includes a rule-based scoring dimension and a deep learning scoring dimension. The rule-based scoring dimension is based on classic composition rules, dynamically adjusting the weight ratio of each composition rule in conjunction with scene type labels, and incorporating lighting adaptability scoring. The deep learning scoring dimension uses the PhotoScore open-source pre-trained model, which is fine-tuned and optimized through a composition aesthetics scoring task to output a picture composition quality score of 0-10. The optimal composition ratios include 16:9, 4:3, 1:1, and 9:

16.

6. The intelligent composition guidance method based on multi-view image analysis according to claim 1, characterized in that: In step S5, the visual composition guide lines include the rule of thirds, golden ratio point markers, and a level, along with a dynamic dotted frame to mark the position, boundary outline, and size ratio of the subject in the optimal frame. The dynamic dotted frame adaptively scales with the composition ratio. Both the visual composition guide lines and the dynamic dotted frame are dynamically adjusted in real time by moving randomly, with color changes indicating the degree of deviation. The layered, conversational camera position adjustment guidance includes basic adjustment prompts, spatial adaptation prompts, obstacle avoidance prompts, and lighting optimization prompts. The real-time calibration prompts are triggered by at least one of the following: sound signal, icon flashing signal, or vibration signal, including split-screen preview comparison, directional guidance icon, and light status indicator.

7. A smart composition guidance device based on multi-view image analysis, characterized in that: The system includes an image acquisition module, an image parsing module, a camera position analysis module, a composition derivation module, a visualization guidance module, and a storage module, which are sequentially connected in communication. These modules work together to implement the intelligent composition guidance method based on multi-view image analysis as described in any one of claims 1-6, wherein: Image acquisition module: Provides a minimalist visual guide interface to guide users in acquiring two or more original images from different perspectives, fully preserving the original shooting parameters, and prohibiting image preprocessing operations during the acquisition process; Image analysis module: Built-in finely tuned and optimized target detection model, scene classification model, 3D composition analysis unit and light source acquisition unit, used to synchronously analyze the original image and output standardized core data including subject contour coordinates, subject center position, subject pixel ratio, subject key feature points, scene type label, interference element information, spatial data and light data; Camera position analysis module: Built-in homography transformation algorithm and Hough transform algorithm, used to calculate the transformation relationship of multiple image viewpoints, derive camera position parameters, detect the tilt angle of the screen, and complete the image defect assessment by combining interference element information; Composition derivation module: used to construct a two-dimensional composition scoring system, realize the comprehensive scoring and ranking of original images, select the benchmark viewpoint and derive the optimal shooting viewpoint and composition scheme, and output the optimal composition parameters; Visualization guidance module: Used to generate visual composition guide lines, dynamic dashed frames, layered conversational camera position adjustment guidance and calibration prompts based on optimal composition parameters. It provides split-screen preview comparison and dynamic guide icon interaction functions, and improves the accuracy of subject positioning through dual guidance. Storage module: Used to store network pre-trained models, original images acquired by the image acquisition module, core data parsed by the image parsing module, optimal camera position data analyzed by the camera position analysis module, optimal composition parameters derived by the composition derivation module, and layered, conversational camera position adjustment guidance and dynamic guidance data output by the visualization guidance module.

8. The intelligent composition guidance device based on multi-view image analysis according to claim 7, characterized in that: The device can be deployed on smartphones, consumer mirrorless cameras, and portable imaging devices. The core models of the device are all exported in ONNX format, adapted to local ISP image processors or CPUs, and support millisecond-level analysis response without relying on network transmission.

9. The intelligent composition guidance device based on multi-view image analysis according to claim 7, characterized in that: The visualization guidance module also features a customizable visualization composition guide line function and an interactive mode switching function. Users can turn on or off the corresponding type of visualization composition guide line according to their needs, adapting to the shooting habits of different users.

Images