Football training posture analysis method and system based on image enhancement

Abstract

The application discloses a football training posture analysis method and system based on image enhancement, and belongs to the field of image processing. The method comprises image acquisition, low-light training image processing, motion blur detail reconstruction, posture image transformation, training posture joint node estimation model design, and football training posture analysis. The scheme ensures the visibility of the player's jersey number and facial features in the shadow by local gain improvement and histogram equalization enhancement. Through motion blur detail reconstruction, the blurred texture and sharp outline are restored. The design of the fidelity loss includes a modulus deviation term and a structure correlation error term, which makes the predicted joint heat map more accurate in the direction of the toe and the degree of finger opening. Through phased timing fine adjustment, the details of early frames are avoided from being smoothed, and the details of the football training posture are protected. The analysis effect of the football training posture is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method and system for analyzing football training postures based on image enhancement. Background Technology

[0002] Football training posture analysis methods typically refer to the use of computer vision technology to automatically identify and quantify key points of a player's body and their movement trajectories from training images or videos to evaluate movement standardization and training effectiveness. However, general football training posture analysis methods suffer from problems such as motion blur and failure to specifically restore masked texture details, leading to incorrect joint positioning and consequently poor posture analysis accuracy. Furthermore, these methods often lose details in complex football movements, resulting in blurred predicted positions of toes and fingertips, and excessive smoothing of detailed features in early frames, further contributing to poor posture analysis performance. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a method and system for football training posture analysis based on image enhancement. Addressing the problem that general football training posture analysis methods suffer from motion blur and lack of targeted restoration of masked texture details, leading to incorrect joint positioning and consequently poor posture analysis accuracy, this solution enhances the visibility of player jersey numbers and facial features in shadows through local gain enhancement and histogram equalization. Motion blur detail reconstruction restores blurred textures and sharp contours, reducing noise amplification and providing clear limb contour input for posture estimation, thereby improving the final accuracy of football training posture analysis. Furthermore, addressing the issues of lost details in complex football movements, resulting in blurred predicted positions of toes and fingertips, and excessive smoothing of details in early frames leading to poor posture analysis, this solution incorporates a fidelity loss mechanism, including a modulus bias term and a structural correlation error term, making the predicted joint heatmap more accurate in terms of toe direction and finger spread. Through phased temporal fine-tuning, the problem of early frame details being smoothed out during long-video training is avoided, achieving detail preservation in football training postures and thus improving the overall performance of football training posture analysis.

[0004] The technical solution adopted by this invention is as follows: The image enhancement-based soccer training posture analysis method provided by this invention includes the following steps:

[0005] Step S1: Image acquisition;

[0006] Step S2: Low-light training image processing;

[0007] Step S3: Motion blur detail reconstruction;

[0008] Step S4: Pose image transformation;

[0009] Step S5: Design of the pose joint estimation model;

[0010] Step S6: Football training posture analysis.

[0011] Further, in step S1, the image acquisition involves acquiring historical football training videos and decoding them into a continuous sequence of football training image frames at a fixed frame rate; after preprocessing, two-dimensional annotation is performed to construct an initial football training image set; the two-dimensional annotation consists of scene type annotation and pose ground truth annotation; the initial football training image set is pre-classified, a lightweight convolutional network is trained, a single frame image is used as input, and the output is a category probability distribution to determine the scene type of the initial football training image.

[0012] Further, in step S2, the low-light training image processing involves performing local gain enhancement on the low-light training image in each of the RGB three channels, and constraining the dynamic range expansion amplitude; specifically including:

[0013] Calculate the local block sampling basis;

[0014] Calculate the upper limit of the gain constraint;

[0015] The grayscale distribution of each local block in the low-light training image is cropped, and the cropped part is evenly distributed to the low-frequency range. Then, histogram equalization is performed, and bilinear interpolation is used to stitch back the complete image.

[0016] Further, in step S3, the motion blur detail reconstruction uses local statistical methods to estimate the texture obscured by the blur on the motion blur training image and reconstructs the contour; specifically, this includes:

[0017] Construct an imaging degradation model;

[0018] Detail retention factor estimation;

[0019] Perform clear image reconstruction;

[0020] The construction of a football training posture image set involves first enhancing low-light images for composite training images, then reconstructing details of the results, and performing color consistency correction between the two steps to finally obtain the football training posture image set.

[0021] Furthermore, in step S4, the posture image transformation is to perform a two-dimensional Fourier transform on the football training posture image and convert it to the frequency domain.

[0022] Further, in step S5, the training posture key point estimation model is designed based on the transformed football training posture image set, combining frequency domain loss and a phased fine-tuning strategy to construct the training posture key point estimation model, specifically including:

[0023] Step S51: Model architecture design. Based on the football training posture image set, a backbone network is selected to extract multi-band features. A hierarchical window attention mechanism is used to efficiently capture local details and global dependencies. Joint heatmap sequences are generated in the output layer. Multi-band feature pyramid fusion is introduced in the training stage to connect shallow high-resolution features and deep semantic features from top to bottom and horizontally.

[0024] Step S52: Fidelity loss design. The spatial domain loss is converted into the frequency domain and decomposed into a modulus deviation term and a structural phase mismatch term. The modulus deviation term measures the difference in energy distribution at each frequency, and the structural correlation error term measures the degree of phase matching. Fidelity loss is then established by weighting and summing the fidelity loss with the spatial domain heatmap fidelity loss to form the total training loss.

[0025] Step S53: Phased temporal fine-tuning, the training process is divided into two phases: first, fine-tuning the model on short action training segments, freezing the first two layers of the backbone network to retain the learned low-level texture features; after warm-up, gradually increasing the length of the input sequence to long action sequences, enhancing the overall action coherence while retaining the repaired details;

[0026] Step S54: Perturbation set attitude prediction. A perturbation set prediction mechanism is introduced, which generates N by injecting small random noise into the same input sequence. e Different key point position prediction sequences are used to form a prediction perturbation set. Then, the quality of the perturbation set is evaluated using the integral squared error.

[0027] Further, in step S6, the football training posture analysis involves acquiring real-time football training videos, decoding them into a sequence of football training image frames, preprocessing them, performing pre-classification, and then performing corresponding enhancements. The results are then input into the trained training posture key point estimation model, and football training posture analysis is performed based on the model output.

[0028] The football training posture analysis system based on image enhancement provided by this invention includes an image acquisition module, a low-light training image processing module, a motion blur detail reconstruction module, a posture image transformation module, a training posture joint estimation model design module, and a football training posture analysis module.

[0029] The image acquisition module collects historical football training videos, and after decoding, preprocessing and image annotation, constructs an initial football training image set and performs pre-classification.

[0030] The low-light training image processing module calculates the local block sampling basis and gain constraint upper limit for the pre-classified low-light training images and performs image enhancement.

[0031] The motion blur detail reconstruction module estimates the detail preservation coefficients of the pre-classified motion blur training images, reconstructs clear images, achieves image enhancement, and then constructs a set of football training posture images.

[0032] The posture image transformation module transforms the set of football training posture images into the frequency domain to distinguish the overall posture and detailed motion information corresponding to different spatial frequencies.

[0033] The training posture joint estimation model design module establishes a training posture joint estimation model based on the set of football training posture images after conversion to the frequency domain.

[0034] The football training posture analysis module is based on the training posture joint estimation model. It decodes and processes real-time football training videos and inputs them into the training posture joint estimation model. Based on the joint probability heatmap output by the model, it realizes football training posture analysis.

[0035] The beneficial effects achieved by the present invention using the above solution are as follows:

[0036] (1) In view of the problem that the general football training posture analysis method has the problem that joint positioning is incorrect due to motion blur and failure to restore the masked texture details, which leads to poor posture analysis accuracy, this solution improves the visibility of the player's jersey number and facial features in the shadow by enhancing local gain and histogram equalization; by reconstructing motion blur details, the blurred texture and sharp contour are restored, noise amplification is reduced, and clear limb contour input is provided for posture estimation, thereby improving the final accuracy of football training posture analysis.

[0037] (2) To address the issues of loss of detail in complex football movements in general football training posture analysis methods, resulting in blurred predicted positions of toes and fingertips, and poor posture analysis results due to excessive smoothing of detailed features in early frames, this solution designs a fidelity loss, including a modulus deviation term and a structural correlation error term, to make the predicted joint point heatmap more accurate in terms of toe direction and finger spread. Through phased temporal fine-tuning, the problem of early frame details being smoothed out due to one-time long video training is avoided, thus achieving detail protection of football training postures and improving the football training posture analysis effect. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the image enhancement-based soccer training posture analysis method provided by the present invention.

[0039] Figure 2 A schematic diagram of the image-enhanced soccer training posture analysis system provided by the present invention.

[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0042] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0043] Example 1, see Figure 1 The present invention provides a method for analyzing football training postures based on image enhancement, which includes the following steps:

[0044] Step S1: Image acquisition. Collect historical football training videos, and after decoding, preprocessing and image annotation, construct an initial football training image set and perform pre-classification.

[0045] Step S2: Low-light training image processing. For the pre-classified low-light training images, calculate the local block sampling basis and gain constraint upper limit, and perform image enhancement.

[0046] Step S3: Motion blur detail reconstruction. For the pre-classified motion blur training images, estimate the detail preservation coefficients, reconstruct clear images, achieve image enhancement, and then construct a set of football training posture images.

[0047] Step S4: Posture image transformation, transforming the soccer training posture image set into the frequency domain to distinguish the overall posture and detailed motion information corresponding to different spatial frequencies;

[0048] Step S5: Design of training posture joint estimation model. Based on the set of football training posture images after conversion to the frequency domain, establish a training posture joint estimation model.

[0049] Step S6: Football training posture analysis. Based on the training posture joint estimation model, the real-time football training video is decoded and processed and then input into the training posture joint estimation model. Football training posture analysis is realized based on the joint probability heatmap output by the model.

[0050] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, historical football training videos are acquired and decoded into a continuous sequence of football training image frames at a fixed frame rate (30fps). After preprocessing, two-dimensional annotation is performed to construct an initial football training image set. The preprocessing includes denoising, size normalization, and color space conversion. The extracted football training image frames are organized into segments in chronological order. The two-dimensional annotation consists of scene type annotation and pose ground truth annotation. Scene type annotation is based on the visual quality of the image, labeling it with category tags, including clear training images under normal lighting, training images under low lighting, and training images with motion blur. Composite training images (simultaneously exhibiting low lighting and motion blur); pose ground truth annotation involves mapping the coordinates of key human joints (ankle, knee, hip, shoulder, elbow, wrist) in each frame and generating corresponding Gaussian heatmaps; pre-classification of the initial football training image set is performed. During football training, different time periods, weather conditions, and field conditions can lead to significant differences in the acquired images. Therefore, a lightweight convolutional network (ShuffleNetV2) is trained, taking a single frame image as input and outputting a class probability distribution to determine the scene type of the initial football training images; a multi-class cross-entropy loss function is used for optimization.

[0051] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the low-light training image processing is necessary because when the image brightness is insufficient, the player's jersey color, skin color, and background are easily mixed together, affecting human body region segmentation and joint detection. Therefore, for low-light training images, local gain enhancement is performed on the RGB three channels respectively, and the dynamic range expansion is constrained to prevent dark area noise from being excessively amplified, while preserving color authenticity. The specific operation is as follows:

[0052] Calculate the local block sampling basis , is represented as: ;in, and These are the width and height of the local block, respectively, with values ​​ranging from 16×16 to 32×32; It represents the total number of grid cells in the image.

[0053] Calculate the upper limit of the gain constraint , represented as: ;in, It is a gain adjustment factor used to define the limit value of local enhancement intensity, with a value of 0.3 to 0.7;

[0054] The grayscale distribution of each local block in the low-light training image is cropped to the interval [0, N]. c The cropped portion is evenly distributed across the low-frequency range, followed by histogram equalization and bilinear interpolation to stitch the image back together. This improves the visibility of details in dark areas and color differentiation, making it easier to separate the human figure from the background and providing clearer input for pose detection. After histogram equalization, the gain amplitude of the RGB three channels is normalized to avoid oversaturation of a single channel, ensuring that the jersey and skin tone remain distinguishable after enhancement.

[0055] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, motion blur detail reconstruction is necessary because motion blur often occurs in images of players in high-speed motion, especially during long exposures or low frame rate acquisition, where limb edges may appear blurred, leading to joint positioning errors. Therefore, for motion blur training images, blur is considered as a form of information attenuation. Local statistical methods are used to estimate the textures obscured by blur and reconstruct sharper contours. The specific operation is as follows:

[0056] Construct an imaging degradation model, represented as: ;in, is the pixel value of the motion-blurred training image at the corresponding position, where x is the image position; These are the pixel values ​​of a potentially sharp image; B is the detail retention coefficient, indicating the degree of information retention; B is the background reference brightness, which is the average value of the image.

[0057] The detail retention factor estimate is expressed as: ;in, It is the deblurring intensity coefficient, with a value ranging from 0.6 to 0.9; y is the local neighborhood, where y is the neighborhood position of x; c is the color channel index. and These are the pixel value in the c channel and the background reference brightness, respectively.

[0058] Achieving a sharp image reconstruction is represented as: ; It is the lower limit of the detail preservation coefficient, with a value of 0.1~0.2; it reduces the breakage of joint edges caused by motion blur, so that the pose estimation algorithm can capture the limb position more accurately; after reconstruction, edge protection filtering is required to avoid amplifying noise during the deblurring process, especially for grass texture and reflective areas of protective gear;

[0059] The construction of the football training posture image set involves first enhancing low-light images, then reconstructing details of the results, and performing a color consistency correction between the two steps to prevent color cast accumulation caused by the two enhancement steps. The final result is a football training posture image set (normal lighting clear training images, enhanced low-light training images, enhanced motion blur training images, and enhanced composite training images).

[0060] By performing the above operations, this solution addresses the problem of poor posture analysis accuracy caused by motion blur and failure to specifically restore masked texture details in general football training posture analysis methods, which leads to incorrect joint positioning. This is achieved by enhancing local gain and histogram equalization to ensure improved visibility of player jersey numbers and facial features in shadows. Furthermore, by reconstructing motion blur details, the solution restores blurred textures and sharp contours, reduces noise amplification, and provides clear limb contour input for posture estimation, thereby improving the final accuracy of football training posture analysis.

[0061] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the posture image transformation is that in the football training image, different spatial frequencies correspond to different frequency bands of motion information: low frequencies mainly represent the overall posture and body contour, while high frequencies represent details (such as finger bending, shoe texture, and shin guard edge). Therefore, a two-dimensional Fourier transform is used on the football training posture image to transform it to the frequency domain in order to distinguish the overall posture and detailed motion information corresponding to different spatial frequencies.

[0062] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the training posture key point estimation model is designed based on the transformed football training posture image set. Combining frequency domain loss and a phased fine-tuning strategy, the training posture key point estimation model is constructed, specifically including:

[0063] Step S51: Model architecture design. Based on the football training posture image set, a backbone network (SwinTransformer) is selected to extract multi-band features. A hierarchical window attention mechanism is used to efficiently capture local details and global dependencies. Joint heatmap sequences are generated in the output layer. During the training phase, multi-band feature pyramid fusion is introduced to connect shallow high-resolution features and deep semantic features from top to bottom and laterally to enhance the model's ability to simultaneously preserve high-frequency details (such as toes and wrists) and low-frequency overall posture. The specific architecture includes three sub-units: a spatiotemporal feature extraction backbone, a multi-band feature pyramid fusion, and a heatmap regression output head. The spatiotemporal feature extraction backbone stacks the continuous football training image frame sequences from the football training posture image set and inputs them into the network. After each stage of the SwinTransformer, a lightweight temporal self-attention layer is introduced; a multi-band feature pyramid fusion network (FPN) is constructed to fuse and laterally connect the multi-band features output from different stages of the SwinTransformer, explicitly enhancing the network's ability to simultaneously preserve high-frequency details (toe pointing, wrist bending) and low-frequency overall posture (torso orientation, center of gravity position), ensuring that both high- and low-frequency information are fully expressed in the final output layer; the feature map of the heatmap regression output head after feature pyramid fusion is upsampled to sub-pixel level through a series of deconvolution layers, and finally reconstructed to the same spatial resolution as the input image, generating a sequence of joint probability heatmaps with the same size as the input image;

[0064] Step S52: Fidelity Loss Design. In pose estimation, if the joint heatmap predicted by the model lacks sufficient correlation with the real heatmap in the high-frequency range, the network tends to reduce the high-frequency amplitude to minimize the overall error. This results in details (such as toe direction and wrist flexion) being dulled. Therefore, the spatial domain loss is converted to the frequency domain and decomposed into a magnitude bias term and a structural phase mismatch term. The magnitude bias term measures the difference in energy distribution at each frequency (corresponding to the intensity matching between the overall and detailed aspects), while the structural correlation error term measures the degree of phase matching (corresponding to the consistency of the joint contour shape). This establishes a fidelity loss, ensuring that even if the high-frequency detail energy is low, it will not be suppressed by the overall error. Represented as: ;in, It is the frequency domain amplitude of the predicted heatmap (after two-dimensional Fourier transform); It is the frequency domain amplitude of the actual heatmap; These are frequency domain coordinates; It refers to structural coherence, which corresponds to frequency domain phase matching. It is an adaptive energy benchmark; It is the real part; F p (·) is the two-dimensional discrete Fourier transform for predicting the heatmap of key points; It is a complex conjugate; to avoid high-frequency details being dulled by the overall error of cross-entropy, the model retains key details while maintaining the correct overall pose; to prevent frequency domain loss from dominating training and causing a shift in the spatial domain probability distribution, the fidelity loss is combined with the spatial domain heatmap fidelity loss. The weighted summation constitutes the total training loss. , represented as: ; ; It is a heatmap of key point probabilities predicted by the model. It is a Gaussian heatmap of the joint points generated from actual annotations; This is the balance coefficient, ranging from 0.1 to 0.5;

[0065] Step S53: Phased temporal fine-tuning. Training a long video in one go can dull the detail of early frames over multiple iterations. Therefore, the training process is divided into two phases: First, fine-tune the model on a short action training segment (0.5 seconds), freezing the first two layers of the backbone network to preserve the learned low-level texture features. At this stage, detail correlation is high, and the loss can effectively reconstruct high-frequency amplitudes (first 1 / 3 of training). After warm-up, gradually increase the input sequence length to a long action sequence (6 seconds), enhancing overall action coherence while preserving the repaired details (last 2 / 3 of training). In each training phase, the loss still uses the frequency domain separation formula, but the input is the predicted heatmap and the actual heatmap of the current segment. When switching phases, the learning rate decays by a factor of ten to stabilize the long sequence training process.

[0066] Step S54: Perturbation set pose prediction. In football training pose analysis, a single deterministic prediction cannot reflect the model's confidence in complex scenes (severe motion blur, multiple people occlusion). To distinguish whether the prediction error stems from the inherent ambiguity of the data or the model's insufficient understanding, a perturbation set prediction mechanism is introduced. By injecting small random noise into the same input sequence, N values ​​are generated. e A set of predicted keypoint positions is generated by creating a set of predicted perturbations. The quality of this perturbation set is then evaluated using the integral squared error to ensure that the pose changes output by the model originate from the physical variability of the actual action, rather than non-physical random noise from the network. The metric for evaluating the quality of the perturbation set is the integral squared error, expressed as: ; where CRPS is integral squared error, used to evaluate model performance (the football training posture image set needs to be divided into training set, validation set and test set. The validation set is used for model training and monitoring the training process. The test set is the same as the final performance evaluation. The overall CRPS is obtained by averaging all samples, joints and time steps in the test set. If the overall CRPS is lower than the evaluation threshold, the initial parameters of the model are adjusted and retrained. Otherwise, the model training is completed). K is the total number of video segments, n is the video segment index; K is the number of key points, j is the key index; It is the joint weight; is the number of perturbation instances in the perturbation set, the number of prediction result groups generated by the perturbation, and m is the perturbation instance index in the perturbation set; It is the predicted coordinate of the position of the j-th joint by the perturbation instance of the m-th perturbation set in the n-th video segment; It is the predicted coordinate of the j-th joint for the l-th perturbation instance in the n-th video segment; These are the actual coordinates corresponding to the joints;

[0067] The first term (internal error term) The first term is the mean absolute error of the perturbation set, which measures how close the predicted center of the perturbation set is to the true value, penalizing the system's bias; the second term (discretion term) is... It is the average distance between all pairs of perturbation instances within the perturbation set, measuring the degree of divergence in the prediction results and penalizing predictions that are too concentrated (underdiscrete).

[0068] By performing the above operations, this solution addresses the problems of general football training posture analysis methods, such as loss of detail in complex football movements, resulting in blurred predicted positions of toes and fingertips, and poor posture analysis performance due to excessive smoothing of detailed features in early frames. This solution incorporates a fidelity loss term, including a modulus bias term and a structural correlation error term, to make the predicted joint heatmaps more accurate in terms of toe direction and finger spread. Through phased temporal fine-tuning, it avoids the problem of early frame details being smoothed out due to one-time long video training, thus preserving the details of football training postures and improving the overall performance of football training posture analysis.

[0069] Example 7, see Figure 1 This embodiment is based on the above embodiment. In step S6, the football training posture analysis involves acquiring real-time football training videos, decoding them into a sequence of football training image frames, preprocessing them, performing pre-classification, and then performing corresponding enhancements. These images are then input into the trained training posture joint estimation model. Based on the model output, football training posture analysis is performed: the model outputs a joint probability heatmap, and Argmax is applied to the heatmap to convert the probability distribution into specific two-dimensional pixel coordinates. All calculated joints are connected sequentially to form a human skeleton topology, which is then compared with a standard action template. The relative deviation percentage score is calculated. If the score is not lower than 0.9, the action is considered standard; if the score is between 0.75 and 0.9, it is considered good with minor flaws; if the score is between 0.6 and 0.75, it is considered to have significant deviations and needs correction; if the score is lower than 0.6, it is considered to have serious errors and high risks, and an early warning is issued.

[0070] Example 8, see Figure 2 Based on the above embodiments, the image-enhanced football training posture analysis system provided by the present invention includes an image acquisition module, a low-light training image processing module, a motion blur detail reconstruction module, a posture image transformation module, a training posture joint estimation model design module, and a football training posture analysis module.

[0071] The image acquisition module collects historical football training videos, and after decoding, preprocessing and image annotation, constructs an initial football training image set and performs pre-classification.

[0072] The low-light training image processing module calculates the local block sampling basis and gain constraint upper limit for the pre-classified low-light training images and performs image enhancement.

[0073] The motion blur detail reconstruction module estimates the detail preservation coefficients of the pre-classified motion blur training images, reconstructs clear images, achieves image enhancement, and then constructs a set of football training posture images.

[0074] The posture image transformation module transforms the set of football training posture images into the frequency domain to distinguish the overall posture and detailed motion information corresponding to different spatial frequencies.

[0075] The training posture joint estimation model design module establishes a training posture joint estimation model based on the set of football training posture images after conversion to the frequency domain.

[0076] The football training posture analysis module is based on the training posture joint estimation model. It decodes and processes real-time football training videos and inputs them into the training posture joint estimation model. Based on the joint probability heatmap output by the model, it realizes football training posture analysis.

[0077] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0079] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for analyzing soccer training postures based on image enhancement, characterized in that: The method includes the following steps: Step S1: Image acquisition. Collect historical football training videos, and after decoding, preprocessing and image annotation, construct an initial football training image set and perform pre-classification. Step S2: Low-light training image processing. For the pre-classified low-light training images, calculate the local block sampling basis and gain constraint upper limit, and perform image enhancement. Step S3: Motion blur detail reconstruction. For the pre-classified motion blur training images, estimate the detail preservation coefficients, reconstruct clear images, achieve image enhancement, and then construct a set of football training posture images. Step S4: Posture image transformation, transforming the soccer training posture image set into the frequency domain to distinguish the overall posture and detailed motion information corresponding to different spatial frequencies; Step S5: Design of training posture joint estimation model. Based on the set of football training posture images after conversion to the frequency domain, establish a training posture joint estimation model. Step S6: Football training posture analysis. Based on the training posture joint estimation model, the real-time football training video is decoded and processed and then input into the training posture joint estimation model. Football training posture analysis is realized based on the joint probability heatmap output by the model. In step S5, the training posture joint estimation model is designed based on the transformed football training posture image set, combining frequency domain loss and a phased fine-tuning strategy to construct the training posture joint estimation model, specifically including: Step S51: Model architecture design. Based on the football training posture image set, a backbone network is selected to extract multi-band features. A hierarchical window attention mechanism is used to efficiently capture local details and global dependencies. A joint heatmap sequence is generated in the output layer. Multi-band feature pyramid fusion is introduced during the training phase to connect shallow high-resolution features and deep semantic features from top to bottom and laterally. The specific architecture includes three sub-units: spatiotemporal feature extraction backbone, multi-band feature pyramid fusion, and heatmap regression output head. The spatiotemporal feature extraction backbone stacks the continuous football training image frame sequence from the football training posture image set into the network. A lightweight temporal self-attention layer is introduced after each stage of SwinTransformer. Multi-band feature pyramid fusion constructs a feature pyramid network to fuse and laterally connect the multi-band features output from different stages of SwinTransformer from top to bottom. The feature map after feature pyramid fusion in the heatmap regression output head is upsampled to sub-pixel level through a deconvolution layer and finally reconstructed to the same spatial resolution as the input image to generate a joint probability heatmap sequence with the same size as the input image. Step S52: Fidelity loss design. The spatial domain loss is converted to the frequency domain and decomposed into a magnitude deviation term and a structural phase mismatch term. The magnitude deviation term measures the energy distribution difference at each frequency, and the structural correlation error term measures the phase matching degree, thereby establishing the fidelity loss. Represented as: ;in, It predicts the frequency domain amplitude of the heatmap; It is the frequency domain amplitude of the actual heatmap; These are frequency domain coordinates; It is structural coherence; It is an adaptive energy benchmark; It is the real part; F p (·) is the two-dimensional discrete Fourier transform for predicting the heatmap of key points; It is a complex conjugate; combining the fidelity loss with the spatial domain heatmap fidelity loss. The weighted summation constitutes the total training loss. , represented as: ; ; It is a heatmap of key point probabilities predicted by the model. It is a Gaussian heatmap of the joint points generated from actual annotations; It is the balance coefficient; Step S53: Phased temporal fine-tuning, dividing the training process into two phases: First, fine-tuning the model on short action training segments, freezing the first two layers of the backbone network to retain the learned low-level texture features; after warm-up, gradually increasing the length of the input sequence to long action sequences, enhancing the overall action coherence while retaining the repaired details; in each training phase, the loss still uses the frequency domain separation formula, but the input is the predicted heatmap and the real heatmap of the current segment; Step S54: Perturbation set attitude prediction. A perturbation set prediction mechanism is introduced, which generates N by injecting small random noise into the same input sequence. e Different keypoint position prediction sequences are used to form a prediction perturbation set. The quality of the perturbation set is then evaluated using the integral squared error; the quality evaluation metric is the integral squared error, expressed as: Where CRPS is the integral squared error; K is the total number of video segments, n is the video segment index; K is the number of key points, j is the key index; It is the joint weight; is the number of perturbation instances in the perturbation set, and m is the index of the perturbation instance in the perturbation set; It is the predicted coordinate of the position of the j-th joint by the perturbation instance of the m-th perturbation set in the n-th video segment; It is the predicted coordinate of the j-th joint for the l-th perturbation instance in the n-th video segment; These are the actual coordinates corresponding to the joints.

2. The image enhancement-based soccer training posture analysis method according to claim 1, characterized in that: In step S2, the low-light training image processing involves performing local gain enhancement on the RGB three channels of the low-light training image and constraining the dynamic range expansion amplitude; specifically, it includes: Calculate the local block sampling basis; Calculate the upper limit of the gain constraint; The grayscale distribution of each local block in the low-light training image is cropped, and the cropped part is evenly distributed to the low-frequency range. Then, histogram equalization is performed, and bilinear interpolation is used to stitch back the complete image.

3. The image enhancement-based soccer training posture analysis method according to claim 1, characterized in that: In step S3, the motion blur detail reconstruction uses local statistical methods to estimate the texture obscured by the blur on the motion blur training image and reconstructs the contours; specifically, this includes: Construct an imaging degradation model; Detail retention factor estimation; Perform clear image reconstruction; The construction of a football training posture image set involves first enhancing low-light images for composite training images, then reconstructing details of the results, and performing color consistency correction between the two steps to finally obtain the football training posture image set.

4. The image enhancement-based soccer training posture analysis method according to claim 3, characterized in that: In step S5, the fidelity loss design involves converting the spatial domain loss into the frequency domain and breaking it down into a modulus deviation term and a structural phase mismatch term. The modulus deviation term measures the difference in energy distribution at each frequency, and the structural correlation error term measures the degree of phase matching. This establishes the fidelity loss. The fidelity loss is then weighted and summed with the spatial domain heatmap fidelity loss to form the total training loss.

5. The image enhancement-based soccer training posture analysis method according to claim 4, characterized in that: In step S5, the perturbation set attitude prediction is achieved by introducing a perturbation set prediction mechanism, which generates N by injecting small random noise into the same input sequence. e Different key point position prediction sequences are used to form a prediction perturbation set. Then, the quality of the perturbation set is evaluated using the integral squared error.

6. The image enhancement-based soccer training posture analysis method according to claim 1, characterized in that: In step S1, the image acquisition involves capturing historical football training videos and decoding them into a continuous sequence of football training image frames at a fixed frame rate. After preprocessing, two-dimensional annotation is performed to construct an initial football training image set; the two-dimensional annotation consists of scene type annotation and pose ground truth annotation; the initial football training image set is pre-classified, and a lightweight convolutional network is trained, which takes a single frame image as input and outputs a category probability distribution to determine the scene type of the initial football training image.

7. The image enhancement-based soccer training posture analysis method according to claim 1, characterized in that: In step S6, the football training posture analysis involves acquiring real-time football training videos, decoding them into a sequence of football training image frames, preprocessing them, performing pre-classification, and then performing corresponding enhancements. The results are then input into the trained training posture joint estimation model, and football training posture analysis is performed based on the model output.

8. A football training posture analysis system based on image enhancement, used to implement the football training posture analysis method based on image enhancement as described in any one of claims 1-7, characterized in that: It includes an image acquisition module, a low-light training image processing module, a motion blur detail reconstruction module, a posture image transformation module, a training posture joint estimation model design module, and a football training posture analysis module. The image acquisition module collects historical football training videos, and after decoding, preprocessing and image annotation, constructs an initial football training image set and performs pre-classification. The low-light training image processing module calculates the local block sampling basis and gain constraint upper limit for the pre-classified low-light training images and performs image enhancement. The motion blur detail reconstruction module estimates the detail preservation coefficients of the pre-classified motion blur training images, reconstructs clear images, achieves image enhancement, and then constructs a set of football training posture images. The posture image transformation module transforms the set of football training posture images into the frequency domain to distinguish the overall posture and detailed motion information corresponding to different spatial frequencies. The training posture joint estimation model design module establishes a training posture joint estimation model based on the set of football training posture images after conversion to the frequency domain. The football training posture analysis module is based on the training posture joint estimation model. It decodes and processes real-time football training videos and inputs them into the training posture joint estimation model. Based on the joint probability heatmap output by the model, it realizes football training posture analysis.

Images