Photographing posture correction system and method

Abstract

The invention belongs to the field of photographing systems, and provides a photographing posture correction system and method, and the method comprises the steps: collecting a current scene image through a camera, classifying a scene through a scene classification model, screening a photographing posture matched with the scene from a posture database according to a classification result, and recommending the photographing posture to a user; performing key point detection on the current posture of the user by using the key point detection model, extracting key point coordinates of the user and generating skeleton representation; the current posture of the user is compared with the recommended posture, the posture matching degree is calculated, key error points are marked, and the user is prompted to conduct posture adjustment.

Description

Technical Field

[0001] The present invention belongs to the field of photographing systems, and in particular relates to a photographing posture correction system and method. Background Art

[0002] With the development of the tourism industry, photography has become an important way for tourists to record their travel experiences. However, most current travel photography assistance tools fail to fully consider the diversity of scenarios and the personalized needs of users, resulting in a limited range of pose options, insufficient recognition accuracy, and a lack of real-time correction capabilities. In existing technologies, most photography assistance tools only provide static pose suggestions and lack the ability to recognize and correct poses during the photo shoot. Therefore, there is an urgent need for an intelligent pose recommendation and correction technology that can provide personalized recommendations based on different scenarios and recognize and correct the user's pose in real time, thereby improving the photo quality and user experience. Summary of the Invention

[0003] In order to solve the problems in the prior art, the present invention provides a photographing posture correction system, comprising the following modules:

[0004] The pose recommendation module is used to capture the current scene image through the camera, classify the scene using the scene classification model, filter the photo poses that match the scene from the pose database based on the classification results, and recommend them to the user;

[0005] A posture recognition module is used to detect key points of the user's current posture using a key point detection model, extract the user's key point coordinates and generate a skeleton representation;

[0006] The posture correction module is used to compare the user's current posture with the recommended posture, calculate the posture matching degree, mark the key error points and prompt the user to adjust the posture.

[0007] Furthermore, the posture recommendation module includes:

[0008] A scene classification unit is used to receive images captured by the camera, preprocess and classify the images using a scene classification model, and output the scene category;

[0009] A posture screening unit is used to extract a set of photographic postures related to the current scene from a posture database according to the scene category;

[0010] The posture ranking unit is used to sort the filtered posture set according to user preferences or historical data and output the top k recommended postures.

[0011] Furthermore, the scene classification model includes:

[0012] An image preprocessing module is used to receive the original input image and perform preprocessing;

[0013] The multi-scale convolutional feature extraction module uses a lightweight convolutional network to introduce a multi-scale receptive field, uses 3×3, 5×5, and 7×7 convolution kernels for parallel feature extraction, and concatenates and fuses the feature maps of the three scales;

[0014] The global attention fusion module is used to introduce an attention mechanism to automatically weight the areas in the feature map that are more sensitive to classification;

[0015] Hierarchical label decoding module, which uses a two-level label system to output labels layer by layer with a two-stage classifier

[0016] Output classification label module is used to finally output the classification results.

[0017] Furthermore, the gesture recognition module includes:

[0018] A key point detection unit is used to receive the user's real-time photo image and extract the coordinates and confidence values of the key points of the human body using a key point detection model;

[0019] A skeleton generation unit, used to generate a skeleton representation of the user's posture according to the human body key point connection rules;

[0020] The key point screening unit is used to remove key point sets with confidence levels lower than a threshold.

[0021] Furthermore, the key point detection model includes:

[0022] The image input and feature extraction module is responsible for encoding the input image into a multi-level spatial feature map;

[0023] A pose prior-aware encoder is used to introduce high-frequency pose priors in the photo scene as shape guidance to improve key point positioning accuracy;

[0024] The adjustable deformable convolution part detection module is used to flexibly model deformable parts using dynamic convolution to improve the positioning robustness of occluded parts;

[0025] The key point heat map decoder and confidence estimation module is used to output the spatial heat map and confidence prediction results of each key point.

[0026] Furthermore, the implementation of the pose prior-aware encoder includes:

[0027] Construct posture prior atlas library P = {P k ∈R h×w×d |k=1,...,K}, each P k A heatmap encoding of a skeleton image representing a typical photo pose;

[0028] Using the scene label c∈C of the current input image, a small pose classifier f pose-prior Predict the most relevant prior pose template:

[0029]

[0030] use Perform channel attention-guided fusion with the current feature map:

[0031]

[0032] Where α is the guide weight tensor, which enhances the high-frequency area features of the posture, F0 is the multi-level spatial feature map, MLP is the multi-layer perceptron, and Sigmoid is the activation function.

[0033] Furthermore, the implementation of the adjustable deformable convolution part detection module, key point heat map decoder and confidence estimation module includes:

[0034] Use dynamic offset convolution to extract slightly changing structural features:

[0035]

[0036] in:

[0037] p is the standard convolution sampling point; R represents the sampling area of the convolution kernel; Δp i is the learnable offset corresponding to the i-th key point; w(p) represents the convolution weight of the convolution kernel at the standard position p; F i is the response feature map of the i-th key point;

[0038] For each F i Generate keypoint heatmap H using 1×1 convolution i ∈R h×w ;

[0039] For H i Take the maximum response point as the prediction coordinate:

[0040] (x i ,y i )=argmax (x,y) H i (x,y)

[0041] At the same time, the confidence scoring module is introduced and regional maximum normalization is adopted:

[0042] c i =max (x,y) H i (x,y)

[0043] The final output of the human key point prediction result is:

[0044] P={(x i ,y i ,c i )|i=1,...,N}.

[0045] Furthermore, the posture correction module includes:

[0046] A polygon construction unit, configured to construct polygons according to key point sets of the user's current posture and the target posture respectively;

[0047] A similarity calculation unit is used to calculate the similarity between the current polygon and the target polygon, and obtain the overall posture matching degree by averaging;

[0048] The error positioning unit is used to mark the key error points of the current posture according to the polygon with the lowest matching degree and prompt the user to adjust the posture.

[0049] Furthermore, the system also includes a posture correction prompt module, which includes:

[0050] An image overlay prompt generating unit, used to overlay prompt graphics on the original image, so that the user can intuitively identify the body part that needs to be adjusted;

[0051] A voice guidance generation unit is used to generate voiced human-computer interaction instructions based on the location, type and direction of the key error points;

[0052] Dynamic action reference demonstration unit, used to provide users with standard action demonstration animations of target postures to assist understanding and imitation;

[0053] The correction feedback confirmation and re-judgment unit is used to re-identify the posture after the user attempts to adjust it and determine whether the matching criteria are met.

[0054] Another aspect of the present invention provides a method for correcting photographic posture, which utilizes the aforementioned system to correct photographic posture.

[0055] The present invention provides a system and method for correcting photographic postures, which perform a recommendation, recognition and correction algorithm for tourist photographic postures. It utilizes scene classification, human posture key point detection and geometric similarity algorithm to improve photographic effects and user experience, and solves the problems of posture selection, posture recognition accuracy and lack of real-time correction function in the prior art.

[0056] Furthermore, this invention employs a hierarchical scene classification model that combines multi-scale fusion with an attention mechanism. This model can accurately distinguish between different tourist scenes (such as beaches, snow-capped mountains, and ancient buildings), and selects highly matching pose sets from a pose database based on scene semantics. By introducing a two-layer labeling system and a scene prior modeling mechanism, the semantic consistency between scenes and poses and the contextual adaptability of recommendations are improved, significantly enhancing the personalization and professionalism of the photo-taking experience.

[0057] Furthermore, this invention utilizes an innovative keypoint detection model that incorporates pose priors and an adjustable deformable convolutional architecture. This model can detect common photographic pose structures and demonstrates excellent recovery from occluded areas (such as sideways arms and back legs). During the skeleton construction process, a confidence screening mechanism is incorporated to effectively remove low-confidence keypoints, ensuring the integrity and stability of the skeleton graph and providing highly reliable input for subsequent pose assessment and correction.

[0058] Furthermore, by grouping key points into structured polygonal regions and using geometric metrics such as angle similarity, side length ratio, and area, the system enables multi-dimensional local posture structure comparison. Compared to traditional whole-body skeleton vector matching methods, this invention supports regional posture matching, calculating the similarity of each limb region at a fine-grained level, effectively identifying subtle but significant posture deviations. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0060] Figure 1 It is a posture correction schematic diagram of the present invention. DETAILED DESCRIPTION

[0061] Below, the invention is preferably described with reference to the accompanying drawings and specific embodiments.

[0062] This embodiment solves the above problem through the following steps:

[0063] In one embodiment, the photographing posture correction system includes the following modules:

[0064] The posture recommendation module is used to capture the current scene image through the camera, classify the scene using the scene classification model, filter the photo poses that match the scene from the posture database based on the classification results, and recommend them to the user.

[0065] The pose recommendation module is used to obtain the current photo scene of the user and perform semantic recognition of the scene based on a pre-trained deep learning classification model. The module first collects scene images in real time through the camera and performs pre-processing operations such as size normalization and brightness adjustment on the image. Subsequently, a lightweight convolutional neural network (such as MobileNet or EfficientNet or ResNet-18, etc.) is used as a scene classification model to perform feature extraction and multi-class discrimination on the input image, and output scene category labels such as "beach", "ancient building", "mountain", "street", etc. Based on the classification results, the system retrieves a pose set that matches the scene from a pre-built pose database. The pose set can be associated with a specific scene through label or vector similarity. Finally, the system prioritizes the elements in the pose set based on the user's historical selection preferences or shooting style characteristics, and recommends the top several most suitable photo pose templates to the user.

[0066] The posture recognition module is used to detect key points of the user's current posture using a key point detection model, extract the user's key point coordinates and generate a skeleton representation.

[0067] The posture recognition module accurately extracts keypoint coordinates from the user's current captured image to represent the user's current body posture. This module receives user image frames captured by the camera and feeds them into a keypoint detection model. This model can be a mainstream human pose estimation network, such as MoveNet, OpenPose, HRNet, or BlazePose. The model output consists of a set of two-dimensional coordinate points, each representing the spatial position of a specific body part (such as the head, shoulder, elbow, wrist, waist, knee, ankle, etc.), along with a confidence score to assess the reliability of the keypoint detection results. Keypoints with a confidence score below a set threshold (e.g., 0.3) are removed to ensure the stability and accuracy of the skeleton recognition. Based on the retained keypoints, a skeleton diagram of the user's current posture is constructed according to human anatomical connection rules (e.g., the standard connection order in COCO or MPII datasets) for subsequent comparison and correction.

[0068] The posture correction module is used to compare the user's current posture with the recommended posture, calculate the posture matching degree, mark the key error points and prompt the user to adjust the posture.

[0069] like Figure 1As shown, the posture correction module compares and analyzes the differences between the user's current posture and the recommended target posture, identifying posture deviations and providing personalized feedback to guide the user in posture adjustment. This module first obtains the user's skeletal key points output by the posture recognition module and the standard posture skeleton provided by the recommendation module. The system then groups these key points into multiple geometric polygonal regions (for example, the left arm region is formed by the left shoulder, left elbow, and left wrist, and the torso region is constructed by the shoulder and waist key points). The system uses geometric metrics such as angle similarity, side length ratio, or area ratio to measure the similarity between the current posture and the recommended posture. A similarity score is calculated for each region and the average is taken as the overall posture match. If the similarity of a region falls below a set threshold (e.g., 0.8), the system deems it to be inaccurate and further identifies the specific key points that contribute to the poor match, such as an elbow not raised or a leg not extended. Finally, the system provides feedback to the user on the areas that need adjustment and recommended directions through visual markers, voice prompts, or on-screen guidance, helping them optimize their posture for better photography.

[0070] Based on the foregoing embodiment, in a further embodiment, the posture recommendation module includes:

[0071] The scene classification unit is used to receive images captured by the camera, preprocess and classify the images using a scene classification model, and output the scene category.

[0072] When assisting with travel photography, different photo scenes require different compositional requirements and action styles. For example, mountain scenes are suitable for leaping with arms extended, while scenes of ancient architecture are suitable for standing or sideways gestures. Accurately identifying the user's current photo environment allows for targeted recommendations of matching pose templates, thereby improving the quality of the photo. Therefore, the scene classification unit is designed to automatically semantically classify captured images and generate structured tags to drive subsequent pose selection logic.

[0073] The scene classification unit includes an image acquisition module and a classification processing module. The image acquisition module uses a camera integrated in the smart terminal to acquire the current scene image I in real time and inputs the image into a pre-trained scene classification model for processing. The classification model is a convolutional neural network structure, preferably a lightweight classification architecture such as MobileNet, EfficientNet, or ResNet. The classification process includes the following steps:

[0074] The original image I is normalized in size, normalized in brightness, and denoised to generate a standard input tensor.

[0075] Use the scene classification model for forward reasoning to obtain the scene category vector: C = f scene(I),C∈{c1,c2,...,c N}, where C represents the predicted category label corresponding to the current image, {c1,c2,…,c N} is a set of predefined scene categories, such as: beach, street scene, ancient building, forest, snow mountain, etc.

[0076] Through the above method, the system can efficiently and accurately convert complex and changeable image inputs into standardized scene labels, realize semantic mapping between photo poses and environmental content, and improve the accuracy and scene adaptability of personalized recommendations.

[0077] For example, a user takes a photo in front of a temple. The system camera captures the image and recognizes its architectural features. The MobileNet classification model identifies the scene category as "ancient building", thereby guiding the subsequent recommendation of a solemn standing posture rather than a leaping posture.

[0078] The posture screening unit is used to extract a set of photographing postures related to the current scene from the posture database according to the scene category.

[0079] Because the chosen pose should be consistent with the scene style, after completing scene classification, we need to further filter out poses from the pose database that match the current scene. For example, for a "beach" scene, we might recommend sitting or jumping poses, while for a "snowy mountain" scene, we might recommend poses like looking up with arms outstretched. The pose screening unit achieves this precise conversion from "semantic scene" to "action template" and is the core of the entire recommendation process.

[0080] Posture database S={s1,s2,…,s M Each photo pose s in i are associated with a set of adaptation labels T(s i ), indicating the scene type set that the posture is applicable to. Based on the scene classification result C, the system screens the posture set and obtains the candidate posture set: S C ={s i ∈S|C∈T(s i )} That is, among all postures, the postures whose label set contains the current scene category C are filtered out as the actions to be recommended.

[0081] By introducing a labeled semantic structure into the database construction process, the posture screening unit implements a semantically driven posture pre-selection mechanism. This has the advantage of reducing the probability of inappropriate posture recommendations while also establishing a scalable and clearly structured matching mechanism, which can be further enhanced with the subsequent expansion of the posture library and labeling system.

[0082] For example, the current scene is identified as "beach", and the system filters out postures with the label "beach" from the posture library, including "facing the beach with arms spread out", "sitting and looking back", "standing on tiptoe with arms outstretched", etc., to form a candidate set.

[0083] The posture ranking unit is used to sort the filtered posture set according to user preferences or historical data and output the top k recommended postures.

[0084] After completing the initial pose screening, multiple pose options are often available for the same scene. Different users have individual preferences for photo styles (e.g., lively, dignified, natural). Therefore, it's necessary to further prioritize the candidate poses based on user historical behavior or preferences to improve the personalization and acceptability of the recommended results. The pose ranking unit is used to accomplish this task.

[0085] Let the pose candidate set be S C ={s1,s2,…,s k}, the user preference vector is U=[u1,u2,…,u k ], which can be constructed in one of the following ways:

[0086] If there is user historical selection data, record the frequency of each posture being selected in the corresponding scenario and perform probability normalization:

[0087]

[0088] If personalized modeling is used, user preferences can be modeled as a multidimensional feature space (such as activity, conservatism, innovation, etc.), and each posture s i Embedded into the same space, sorted by cosine similarity:

[0089]

[0090] After the sorting is completed, the top k postures with the highest sorting scores are selected as the final recommendation results.

[0091] The posture sorting unit realizes the transformation of the static posture library into a dynamic preference model by introducing user personality modeling and behavior feedback mechanism, which not only improves user acceptance, but also enhances the system's long-term adaptability and user stickiness.

[0092] For example, for a user who prefers to take photos with arms outstretched, the "jumping with arms outstretched" pose is selected many times in their historical data. In the "lawn" scene, the system will rank this pose at the top as a key recommendation option.

[0093] Based on the foregoing embodiments or any combination thereof, an implementation of the scene classification model includes:

[0094] In tourist photography scenarios, traditional image classification models often rely on global image features for single-label classification, ignoring fine-grained differences between local semantic structure and environmental context. This leads to frequent misidentification of scenes with similar backgrounds but different functions, such as misidentifying "ancient buildings and alleys" as "city streets." To address this, this embodiment proposes a hierarchical tourism scene classification model that combines a global attention mechanism with multi-scale feature fusion to improve fine-grained scene differentiation in tourism images, ensuring the accuracy and semantic consistency of subsequent photo pose recommendations.

[0095] The scene classification model consists of the following five modules, each of which has the following functions:

[0096] Image preprocessing module:

[0097] Receive the original input image I and perform the following processing:

[0098] Scale to a uniform size (e.g. 224×224);

[0099] Convert to standard RGB space;

[0100] Apply normalization: Where μ is the mean of the images in the training set and σ is the standard deviation.

[0101] Multi-scale convolutional feature extraction module:

[0102] An extended lightweight convolutional network (such as MobileNetV3) is used, multi-scale receptive fields are introduced, and 3×3, 5×5, and 7×7 convolution kernels are used for parallel feature extraction, which can be expressed as:

[0103] F3=Conv 3×3 (I norm )

[0104] F5=Conv 5×5 (I norm )

[0105] F7=Conv 7×7 (I norm )

[0106] Finally, the feature maps of the three scales are spliced and fused: F concat =[F3,F5,F7]

[0107] Global attention fusion module:

[0108] To improve the perception of high semantic regions in the image, an attention mechanism is introduced to automatically weight regions in the feature map that are more sensitive to classification. The implementation is as follows:

[0109] F concatPerform global average pooling and maximum pooling to generate channel-level statistical vector g avg ,g max ;

[0110] Input to shared MLP: α = σ(MLP(g avg +g max ))where σ represents the Sigmoid activation function and α is the normalized channel attention weight vector;

[0111] The final attention weighted output is: F att =α⊙F concat , where ⊙ represents channel-by-channel multiplication.

[0112] Hierarchical label decoding module:

[0113] A two-level labeling system is used (e.g., "nature / humanities" → "beach / ancient town / forest / snow mountain"), and a two-stage classifier is used to output labels layer by layer:

[0114] The first-level classifier outputs the coarse-grained label L1;

[0115] The secondary classifier outputs fine-grained labels based on L1 in Represents a set of secondary subcategories under a primary label.

[0116] Output classification label module:

[0117] The final output classification result is: C = (L1, L2), which represents the complete scene category combination label of the current image, such as ("humanities", "ancient architecture").

[0118] Compared with the traditional single-stage classification model, this scenario classification model has significant advantages in the following aspects:

[0119] Multi-scale perception and attention mechanisms improve the recognition of key image regions and reduce misjudgments. A dual-layer labeling system enhances semantic expression and accurately distinguishes between different detailed environments within the same general category. The hierarchical structure facilitates the addition of new categories and subcategories, facilitating model maintenance and scenario expansion. A one-to-one correspondence with the scene labeling structure of the pose database enhances the end-to-end adaptability of the overall recommendation system.

[0120] For example, a user takes a photo at an "ancient building" scene during a tour:

[0121] After the camera captures the scene image, it inputs it into the model;

[0122] The model first outputs the first-level label as "Humanities";

[0123] The scene is then subdivided into “ancient buildings” under this first-level label;

[0124] The final output label combination is ("humanities", "ancient architecture");

[0125] The posture screening module retrieves photo poses that match "ancient buildings", such as "standing with hands clasped together" and "looking back and gazing at the side face", to improve the style consistency with the scene.

[0126] Based on the foregoing embodiment or any combination thereof, the gesture recognition module includes:

[0127] The key point detection unit is used to receive the user's real-time photo image and use the key point detection model to extract the coordinates and confidence values of the key points of the human body.

[0128] When recognizing poses for tourist photos, accurately extracting the locations of key body parts is fundamental to building a posture model. Key points typically include the head, shoulders, elbows, wrists, waist, knees, and ankles. Identifying the spatial coordinates of these points can be used to reconstruct the user's current pose and facilitate subsequent matching and correction. Therefore, the key point detection unit receives real-time images of the user and, based on a deep learning model, automatically outputs the coordinates of the key body joints and a confidence score to indicate the reliability of each point's recognition.

[0129] The key point detection unit includes an image receiving module and a posture estimation module. The image receiving module obtains the user's current photo frame I, which is collected in real time by the camera. The image is input to the key point detection model f pose The model is preferably a lightweight but high-precision structure, such as MoveNet, BlazePose, OpenPose or HRNet. The model performs forward reasoning on the image and outputs a set of human key points.

[0130] P, in the form:

[0131] P={(x i ,y i ,c i )|i=1,2,...,N}

[0132] in:

[0133] (x i ,y i ) represents the two-dimensional coordinates of the i-th key point;

[0134] c i ∈[0,1] represents the confidence of the i-th key point;

[0135] N is the number of key points output by the model, and its typical value is 17 (as in the COCO dataset standard).

[0136] The model supports target filtering through non-maximum suppression when recognizing multiple people, retaining only the single person instance with the highest score in the image as the current photographed object.

[0137] By using advanced key point detection models, high-precision positioning of various parts of the human body at the moment of taking a photo can be achieved. Combined with the confidence output mechanism, it can effectively suppress recognition errors caused by factors such as occlusion and lighting, providing a reliable data basis for subsequent posture judgment and correction.

[0138] For example, the user stands facing the camera, the system captures the current image frame, and uses the MoveNet model to identify 17 points, including the head, shoulders, elbows, wrists, hips, knees, and ankles. The system then outputs the corresponding coordinates and confidence levels. For example, if the left shoulder point is (210, 340, 0.95), it means that the left shoulder is located at the image coordinates (210, 340) with a confidence level of 95%.

[0139] The skeleton generation unit is used to generate a skeleton representation of the user's posture according to the human body key point connection rules.

[0140] After extracting the coordinates of each key point, a human skeleton must be constructed through regularized structure for pose structure analysis and shape matching. A single coordinate point cannot easily reflect pose information, but combining key points into a topological structure based on their connection relationships can more intuitively represent human motion states. Therefore, the skeleton generation unit is used to convert a set of two-dimensional key points into a connected skeleton graph to achieve pose structure modeling.

[0141] The skeleton generation unit includes a connection rule configuration module and a skeleton construction module. The connection rule configuration module presets a set of key point connection rules L, for example, the connection rules include the following edge set:

[0142] L={(1,2),(2,3),(3,4),...,(15,16)}

[0143] Each pair (i, j) represents a key point P i With P j The skeleton construction module generates the skeleton graph structure based on the connection set and the coordinate set P:

[0144]

[0145] Among them, line(P i ,P j ) represents the point P i to P j The straight line segment of , the skeleton graph S is output as the posture structure representation.

[0146] Through this module, the original key point set can be structured and represented as a posture skeleton graph, which not only improves the human readability of the posture expression, but also provides a shape basis for subsequent geometric similarity calculation, posture alignment and correction, which is conducive to the realization of spatial structure consistency analysis.

[0147] For example:

[0148] In the user's photo, the left shoulder key point (210, 340) and the left elbow key point (180, 410) are identified. The system represents the line connecting the two points as a skeleton edge based on the connection rules. Similarly, multiple points are connected to form a torso and limb chain such as "shoulder-elbow-wrist" and "hip-knee-ankle" to form a complete skeleton.

[0149] The key point screening unit is used to remove key point sets with confidence levels lower than a threshold.

[0150] Due to issues such as occlusion, uneven lighting, and misalignment of body parts in the camera's image capture environment, the confidence level of certain keypoints may be low. Directly using these low-confidence keypoints for pose analysis can introduce errors and reduce pose matching accuracy. Therefore, a confidence threshold is set and points below this threshold are discarded to ensure the reliability of subsequent skeleton construction and pose analysis.

[0151] The key point screening unit selects the key point set P = {(x i ,y i ,c i )} extract all c i Points ≥τ form a high confidence subset:

[0152] P * ={(x i ,y i )|c i ≥τ}

[0153] Where: τ represents the preset confidence threshold, the preferred value is 0.3 to 0.5;

[0154] The removed key points no longer participate in skeleton generation and matching analysis;

[0155] If one of the connection points of a skeleton edge is culled, the edge is not constructed.

[0156] Furthermore, this module can be executed before or after skeleton generation, and can also adjust the threshold in conjunction with a dynamic weight mechanism.

[0157] By quantitatively screening the confidence levels, we effectively avoid the interference of low-confidence key points in posture matching, improve the accuracy and stability of posture analysis, enhance the system's robustness to camera interference factors, and improve the reliability of the final correction suggestions.

[0158] For example, in the user image, the system detects the left ankle key point as (230,680), with a confidence level of only 0.25, which is lower than the threshold of 0.3 and is eliminated by the system. This point does not participate in the skeleton construction and subsequent angle calculation to avoid posture misjudgment due to occlusion.

[0159] Based on the foregoing embodiments or any combination thereof, an implementation of the key point detection model includes:

[0160] In the scenario of tourist photo pose recognition, there are the following challenges:

[0161] When facing the camera, the user's posture presents a typical "frontal or half-side" structure; the lighting at the scene is complex and partial occlusions (such as hats, hair, clothing, and handheld objects) are prone to occur; the goal of posture recognition is not general action recognition, but to serve the accuracy assessment and posture correction of photo postures, and it is necessary to enhance the recognition capabilities of features such as limb opening and closing, standing stability, and facial orientation; system resources may be limited by the terminal, and the model should have the capabilities of high deployment efficiency and fast response speed.

[0162] To address the above issues, this embodiment further proposes an adjustable constrained human key point detection model based on motion structure perception. Its innovation lies in the introduction of a posture prior encoder and a part-level adjustable deformation convolution module to optimize the target of common human postures for photography, improve the structural recognition and occlusion completion capabilities of key parts, and form a lightweight posture estimation model specifically for travel photography scenarios.

[0163] The key point detection model includes the following four main modules:

[0164] The image input and feature extraction module is responsible for encoding the input image into a multi-level spatial feature map. Specifically:

[0165] Input image I∈R H×W×3 , perform standard normalization and size scaling;

[0166] Input to a shallow convolutional backbone network (such as TourPoseNet-Backbone, or other similar feature extraction models), outputting the primary spatial feature F0∈R h×w×d ;

[0167] The network structure uses a multi-path residual module to fuse multi-scale information:

[0168] F0=Concat(ResBlock3(I),ResBlock5(I),ResBlock7(I))

[0169] Among them, ResBlock k Represents a residual convolution module with a k×k receptive field.

[0170] The pose prior-aware encoder is used to introduce high-frequency pose priors in the photo scene as shape guidance to improve key point positioning accuracy. Specifically:

[0171] Construct posture prior atlas library P = {P k ∈R h×w×d |k=1,...,K}, each P k Skeleton heatmap encoding of a typical photo pose (e.g., standing with arms outstretched, sitting and gazing);

[0172] Using the scene label c∈C of the current input image, a small pose classifier f pose-prior Predict the most relevant prior pose template:

[0173]

[0174] Among them, f pose-prior It can be obtained by using common image classification models in the prior art, such as VGGNet, ResNet, etc. or their improved training.

[0175] use Perform channel attention-guided fusion with the current feature map:

[0176]

[0177] Where α is the guide weight tensor, which enhances the high-frequency area features of the posture, MLP is a multi-layer perceptron, and Sigmoid is the activation function.

[0178] The adjustable deformable convolution part detection module is used to flexibly model deformable parts (such as elbows, wrists, and knees) using dynamic convolution to improve the positioning robustness of occluded parts. Specifically:

[0179] Divide the regional branch network into key parts, and each branch learns the spatial offset of its deformation area;

[0180] Use dynamic offset convolution to extract slightly changing structural features:

[0181]

[0182] in:

[0183] p is the standard convolution sampling point; Δp i is the learnable offset corresponding to the i-th key point; F i is the response feature map of the i-th key point.

[0184] The key point heat map decoder and confidence estimation module are used to output the spatial heat map and confidence prediction results of each key point. Specifically:

[0185] For each F i Generate keypoint heatmap H using 1×1 convolution i ∈R h×w ;

[0186] For H i Take the maximum response point as the prediction coordinate:

[0187] (x i ,y i )=argmax (x,y) H i (x,y)

[0188] At the same time, the confidence scoring module is introduced and regional maximum normalization is adopted:

[0189] c i =max (x,y) H i (x,y)

[0190] The final output of the human key point prediction result is:

[0191] P={(x i ,y i ,c i )|i=1,...,N}

[0192] In this embodiment, by combining scene labels with photo pose template guidance, the model can perceive the "key points that should appear" and improve the recognition effect of difficult parts; the multi-scale residual structure and branch deformation detection module collaboratively model the body structure relationship to improve spatial consistency; through deformation modeling and regional heat map normalization mechanism, the implicit posture under occlusion is inferred; the backbone network compression and guided fusion modules in the model structure can be independently tailored, suitable for mobile terminals or edge devices; the model is trained with the characteristics of tourist photo actions as the target, and its performance is better than the general posture recognition model.

[0193] For example:

[0194] A user takes a photo while standing in front of an ancient building. In the image, their right arm is partially obscured by the wall. The model identifies the scene as an "ancient building" using scene labels and loads the "front-facing, palms-clasped" pose as a priori template to enhance the features of the key points in the right arm. It then uses deformation convolution to predict the offset and confidence of the right elbow and wrist. Ultimately, even with the right hand partially obscured, the system accurately determines the position and confidence score of the user's right elbow and wrist, achieving complete skeleton modeling and providing a basis for subsequent posture correction.

[0195] Based on the foregoing embodiment or any combination thereof, the posture correction module includes:

[0196] The polygon construction unit is used to construct polygons according to the key point sets of the user's current posture and the target posture respectively.

[0197] Human posture is composed of several key points, but using only the position of the points makes it difficult to reflect the spatial structure and relative posture of the various parts of the body. Therefore, organizing the key points into a polygonal structure is an effective "local configuration expression method." By demarcating parts such as "shoulder-elbow-wrist" and "hip-knee-ankle" into regional polygons, geometric features such as the opening and closing angles and extension directions of the limbs can be quantified. In this way, the system can perform structured modeling of local postures, thereby supporting fine-grained evaluation of posture matching.

[0198] This unit relies on the pose key point set P = {(x i ,y i )|i=1,…,N}, according to the predefined polygon construction rules, polygon construction is performed for the user's current posture and the recommended target posture respectively. Let the polygon set constructed by the current posture be A={A1,A2,…,A m}, the target posture is constructed as B={B1,B2,...,B m}

[0199] Each polygon A i It is constructed from several key points in sequence, for example:

[0200] Left arm area: A1 = Polygon (P 肩左 ,P 肘左 ,P 腕左 )

[0201] Right leg area: A4 = Polygon (P 髋右 ,P 膝右 ,P 踝右 )

[0202] These regions represent the geometric topology of the movable parts of the body, and the system can automatically generate or manually configure construction rules.

[0203] By abstracting the local structure of the posture into a polygonal area, the subsequent similarity calculation has a stable structural benchmark, which helps to alleviate the recognition offset problem caused by overall position changes and image scaling, thereby more accurately reflecting the true structural differences of the posture.

[0204] For example, in the user's posture, the left arm triangle area is formed by key points (210, 340), (180, 410), and (150, 480), which forms a comparison polygon with the left arm area in the target posture, serving as the basis for subsequent matching.

[0205] The similarity calculation unit is used to calculate the similarity between the current polygon and the target polygon, and obtain the overall posture matching degree by taking the average value.

[0206] The core of posture matching evaluation lies in quantifying the structural differences between the current user's posture and the recommended posture. By evaluating the geometric similarity between each set of corresponding polygons, local posture matching accuracy can be analyzed. Polygonal similarity can be normalized using geometric parameters such as angle, side length, and area to calculate an overall matching score.

[0207] For each pair of corresponding polygons (A i ,B i ) calculate its similarity score s i . It is preferred to use the angle similarity function:

[0208] Polygon A i The internal angles of the angle vector Corresponding B i for

[0209] The similarity function is defined as:

[0210]

[0211] in:

[0212] represents the jth corner in the i-th current pose polygon;

[0213] is the polygon angle corresponding to the target posture;

[0214] s i ∈[0,1], the larger the value, the more similar it is.

[0215] Global pose similarity S total Average all polygons:

[0216]

[0217] Where m represents the number of matching polygons.

[0218] The above calculation method not only takes into account the local geometric features of each polygon, but is also scale-independent and rotation-tolerant. It can adapt to deviations caused by angles or slight movements of the user during the shooting process, and output a more stable and reliable posture matching evaluation.

[0219] For example:

[0220] The system's polygon matching results for the left arm area are s1 = 0.92, and for the right leg area are s4 = 0.75, and the overall matching degree S is calculated. total =0.85, indicating that the overall posture is well similar to the recommended posture, but there are still some deviations in the local area.

[0221] The error positioning unit is used to mark the key error points of the current posture according to the polygon with the lowest matching degree and prompt the user to adjust the posture.

[0222] To effectively provide posture correction prompts, it's necessary to pinpoint the specific locations where the overall match is compromised. By detecting the polygonal region with the lowest similarity score and further tracing the degree of deviation of key points within that region, the "critical error points" affecting posture accuracy can be identified. Based on this information, the system can provide targeted feedback, improving correction efficiency and user experience.

[0223] The error location unit includes the following steps:

[0224] Find s i Polygon A with τ i , where τ is the error judgment threshold (such as 0.8);

[0225] For each low matching polygon A i , take each key point P in the polygon j and the corresponding target key point P j ′’s Euclidean distance:

[0226]

[0227] If d j >δ(error judgment distance threshold), then mark P j is the error key point;

[0228] All error key points are integrated into a set for subsequent prompt presentation.

[0229] By introducing a local structure-driven error analysis approach, the positioning accuracy and interactive guidance of posture correction can be significantly improved. Compared to methods that rely solely on overall scoring, this approach provides prompts by region and action, avoiding information overload and correction failures for users.

[0230] For example:

[0231] In the example, the system found that the matching degree of the left leg area was only 0.68. Further analysis showed that the left knee position deviated from the target knee by more than 15 pixels. The left knee was marked as the key error point, and a voice or image prompt was given: "Please adjust the left leg angle."

[0232] Based on the aforementioned embodiment or any combination thereof, the system further includes a posture correction prompt module.

[0233] After posture recognition and error localization are complete, providing intuitive and perceptible posture feedback is essential to improve user understanding and response to correction instructions during photography. Since most users lack professional training in travel photography, simply outputting a "matching score" or "text prompt" is often insufficient to guide effective adjustments.

[0234] To this end, this embodiment constructs a multimodal posture correction prompt module based on error key point drive. The module integrates image overlay enhancement prompts, real-time voice guidance and dynamic posture animation demonstration, and dynamically updates the prompt content in combination with real-time user feedback, thereby constructing a "recognition-prompt-correction-confirmation" closed-loop control logic.

[0235] The posture correction prompt module includes the following submodules:

[0236] The image overlay prompt generation unit is used to overlay prompt graphics on the original image to enable the user to intuitively identify the body part that needs to be adjusted. Specifically:

[0237] From the error key point set ε={P j}Get the key point coordinates;

[0238] In the captured image, j =(x j ,y j ) is used as the center to draw the prompt element:

[0239] This can be indicated by a red circle, flashing arrow, or dynamic line;

[0240] According to the target key point P′ j =(x′ j ,y′ j ) Draw vector direction lines

[0241] The image enhancement prompt layer is superimposed on the preview interface, and users can observe the direction that needs to be adjusted in real time in the shooting picture.

[0242] The voice guidance generation unit is used to generate voice-based human-computer interaction instructions based on the location, type and direction of the key error points. Specifically:

[0243] Create a template library of part-action-direction prompts, for example:

[0244] "Please raise your left elbow."

[0245] Please move your right knee slightly forward.

[0246] System from Pj Generate text prompt sentences based on the three elements of part label, offset direction and error amplitude;

[0247] Use a text-to-speech (TTS) engine to convert sentences into speech and broadcast them through a speaker on the user side.

[0248] The dynamic action reference demonstration unit is used to provide users with standard action demonstration animations of target postures to assist in understanding and imitation. Specifically,

[0249] The target pose key point set P target ={(x j′ ,y j′ )}Convert to skeletal animation;

[0250] Generate smooth human motion animation (e.g. transition from current pose P to P frame by frame) using skeleton interpolation and action library templates target );

[0251] A "virtual instructor" or "skeleton guide map" is presented in the interface and compared with the user's real-time posture on the same screen.

[0252] The correction feedback confirmation and re-judgment unit is used to re-identify the posture after the user attempts to adjust it and determine whether it meets the matching criteria. Specifically:

[0253] The system monitors the new image frames in the shooting preview screen;

[0254] Automatically trigger key point re-identification, skeleton reconstruction and matching update processes;

[0255] If the matching degree S total >τ, trigger the prompt: "The current posture is good, you can start taking pictures"; otherwise, continue to prompt for fine-tuning.

[0256] For example:

[0257] The system identified that the user's left knee position was 15 pixels off the target position and determined it to be a critical error point.

[0258] The image overlay module displays a red circle at the left knee and draws an arrow pointing forward and downward;

[0259] Voice prompt: "Please move your left knee slightly forward";

[0260] Play standard standing posture animation in the interface to guide users to imitate;

[0261] After the user adjusted their posture, the system re-identified the user's left knee, which was now closer to the target. The match improved to 0.91, and the system announced, "Good posture. Please hold still. Prepare to take the photo."

[0262] In another embodiment, the present invention further provides a method for correcting photographic posture, which uses the aforementioned embodiment to perform photographic posture correction.

[0263] It should be noted that the explanation of the aforementioned embodiment of the photographing posture correction system is also applicable to the method of the embodiment of the present application and will not be repeated here.

[0264] Those skilled in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented using a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0265] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0266] In the several embodiments provided in this application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), magnetic disk or optical disk, and other media that can store program code.

[0267] The above is only a specific embodiment of the present application. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in this application, which should be included in the scope of protection of this application. For some module structures that are not particularly clear in the present invention, the content recorded in the prior art shall prevail. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be regarded as part of the present invention and is used to understand the meaning of some technical features or parameters.

Claims

1. A photographic posture correction system, characterized in that: The system includes the following modules: The pose recommendation module is used to capture the current scene image through the camera, classify the scene using the scene classification model, filter the photo poses that match the scene from the pose database based on the classification results, and recommend them to the user; A posture recognition module is used to detect key points of the user's current posture using a key point detection model, extract the user's key point coordinates and generate a skeleton representation; The posture correction module is used to compare the user's current posture with the recommended posture, calculate the posture matching degree, mark the key error points and prompt the user to adjust the posture.

2. The photographing posture correction system according to claim 1, characterized in that: The posture recommendation module includes: A scene classification unit is used to receive images captured by the camera, preprocess and classify the images using a scene classification model, and output the scene category; A posture screening unit is used to extract a set of photographic postures related to the current scene from a posture database according to the scene category; The posture ranking unit is used to sort the filtered posture set according to user preferences or historical data and output the top k recommended postures.

3. The photographing posture correction system according to claim 1, characterized in that: The scene classification model includes: An image preprocessing module is used to receive the original input image and perform preprocessing; The multi-scale convolutional feature extraction module uses a lightweight convolutional network to introduce a multi-scale receptive field, uses 3×3, 5×5, and 7×7 convolution kernels for parallel feature extraction, and concatenates and fuses the feature maps of the three scales; The global attention fusion module is used to introduce an attention mechanism to automatically weight the areas in the feature map that are more sensitive to classification; Hierarchical label decoding module, which uses a two-level label system to output labels layer by layer using a two-stage classifier; Output classification label module is used to finally output the classification results.

4. The photographing posture correction system according to claim 1, characterized in that: The gesture recognition module includes: A key point detection unit is used to receive the user's real-time photo image and extract the coordinates and confidence values of the key points of the human body using a key point detection model; A skeleton generation unit, used to generate a skeleton representation of the user's posture according to the human body key point connection rules; The key point screening unit is used to remove key point sets with confidence levels lower than a threshold.

5. The photographing posture correction system according to claim 1, characterized in that: The key point detection model includes: The image input and feature extraction module is responsible for encoding the input image into a multi-level spatial feature map; A pose prior-aware encoder is used to introduce high-frequency pose priors in the photo scene as shape guidance to improve key point positioning accuracy; The adjustable deformable convolution part detection module is used to flexibly model deformable parts using dynamic convolution to improve the positioning robustness of occluded parts; The key point heat map decoder and confidence estimation module is used to output the spatial heat map and confidence prediction results of each key point.

6. The photographing posture correction system according to claim 5, characterized in that: The implementation of the pose prior-aware encoder includes: Construct posture prior atlas library P = {P k ∈R h×w×d |k=1,...,K}, each P k A heatmap encoding of a skeleton image representing a typical photo pose; Using the scene label c∈C of the current input image, a small pose classifier f pose-prior Predict the most relevant prior pose template: use Perform channel attention-guided fusion with the current feature map: Where α is the guide weight tensor, which enhances the high-frequency area features of the posture, F0 is the multi-level spatial feature map, MLP is the multi-layer perceptron, and Sigmoid is the activation function.

7. The photographing posture correction system according to claim 6, characterized in that: The implementation of the adjustable deformable convolution part detection module, key point heat map decoder and confidence estimation module includes: Use dynamic offset convolution to extract slightly changing structural features: in: p is the standard convolution sampling point; R represents the sampling area of the convolution kernel; Δp i is the learnable offset corresponding to the i-th key point; w(p) represents the convolution weight of the convolution kernel at the standard position p; F i is the response feature map of the i-th key point; For each F i Generate keypoint heatmap H using 1×1 convolution i ∈R h×w ; For H i Take the maximum response point as the prediction coordinate: (x i ,y i )=argmax (x,y) H i (x,y) At the same time, the confidence scoring module is introduced and regional maximum normalization is adopted: c i =max (x,y) H i (x,y) The final output of the human key point prediction result is: P={(x i ,y i ,c i )∣i=1,…,N}。 8. The photographing posture correction system according to claim 1, characterized in that: The posture correction module includes: A polygon construction unit, configured to construct polygons according to key point sets of the user's current posture and the target posture respectively; A similarity calculation unit is used to calculate the similarity between the current polygon and the target polygon, and obtain the overall posture matching degree by averaging; The error positioning unit is used to mark the key error points of the current posture according to the polygon with the lowest matching degree and prompt the user to adjust the posture.

9. The photographing posture correction system according to claim 1, characterized in that: The system further includes a posture correction prompt module, which includes: An image overlay prompt generating unit, used to overlay prompt graphics on the original image, so that the user can intuitively identify the body part that needs to be adjusted; A voice guidance generation unit is used to generate voiced human-computer interaction instructions based on the location, type and direction of the key error points; Dynamic action reference demonstration unit, used to provide users with standard action demonstration animations of target postures to assist understanding and imitation; The correction feedback confirmation and re-judgment unit is used to re-identify the posture after the user attempts to adjust it and determine whether the matching criteria are met.

10. A method for correcting a photographic posture, characterized in that: The method uses the system described in any one of claims 1-9 to correct photographic posture.

Images