Single action generation deep learning method based on diffusion-generative adversarial framework

Abstract

The invention provides a single motion generation deep learning method based on a diffusion-generative adversarial framework, and the method comprises the steps: dividing a motion sequence, obtaining dynamic features and static features, carrying out the foot contact calculation through forward kinematics, and obtaining the calculated dynamic features; coding the calculated dynamic features into Gaussian random noise, inputting the Gaussian random noise into a generator, and denoising the Gaussian random noise through a diffusion model; performing action synthesis on close real actions by using measured action similarity and an average voting strategy; an input action sequence is given for the calculated dynamic features, and noise is injected in an iteration form through a Markov denoising process in a special local attention layer of a diffusion model for denoising; and independently evaluating the more accurate body action and the reconstructed action through a discriminator to generate a classification result. According to the method, through joint optimization of gradient penalty and a residual structure, the problem of mode collapse under the complex skeleton topology is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of computer graphics, and in particular to a single action generation deep learning method based on a diffusion-generation adversarial framework. Background Art

[0002] Generating a variety of realistic movements has been a long-standing goal in computer graphics, playing a vital role in enhancing the realism and visual appeal of animated content. In recent years, generative adversarial networks (GANs) have emerged as a powerful solution, demonstrating impressive capabilities in the synthesis of high-quality motion sequences. In particular, GANs have demonstrated amazing generative capabilities even in the case of sparse data, producing convincing results with limited training samples.

[0003] Single instance object generation methods in the motion domain mainly rely on GANs. Among them, Ganimator is a striking method that utilizes multi-level GANs to generate novel motions from a single motion sequence. However, its hierarchical cascade generator and discriminator structure consumes more computational resources. Similarly, LS-GAN leverages the power of GANs and the compact representation of motion sequences in the latent space. However, it is mainly tailored for human motion and is limited by simple text-to-motion cues.

[0004] While GANs focus on capturing global features of motion synthesis—emphasizing the distinction between real and generated samples, diffusion-based techniques excel at refining local details. For example, SinMDM

[15] employs a shallow diffusion-based network with a local attention layer, which effectively reduces the receptive field to enhance motion diversity. However, this approach encounters difficulties in synthesizing out-of-distribution motion. Summary of the invention

[0005] In view of the above situation, the main purpose of the present invention is to propose a single action generation deep learning method based on a diffusion-generation adversarial framework to solve the above technical problems.

[0006] The present invention proposes a single action generation deep learning method based on a diffusion-generation adversarial framework, the method comprising the following steps: Step 1: Construct the generator and discriminator based on the Diffusion-GAN architecture, and construct the matching and mixing modules based on the matching and mixing mechanism. The generator, discriminator, matching and mixing modules constitute the Diffusion-GAN model. The generator includes a diffusion model, the diffusion model is combined with a shallow U-Net architecture, the discriminator operates in a multi-scale manner, and the matching and mixing module includes a matching module and a mixing module; Step 2: Divide the motion sequence to obtain dynamic features and static features; The dynamic features are converted to obtain a general motion representation, a foot contact label is added to the general motion representation, and foot contact is calculated by forward kinematics to obtain the calculated dynamic features; Step 3: Add Gaussian random noise to the calculated dynamic features, and input them into the generator for denoising through the diffusion model to obtain the denoised real action features; Based on the matching and mixing module, the denoised real action features are synthesized using the measurement of action similarity and the average voting strategy to obtain the synthesized body action features. Based on the calculated dynamic features and the input action sequence, the Markov denoising mechanism in the special local attention layer of the diffusion model is used to perform denoising in an iterative manner, and the reconstructed action features are output when the maximum number of iterations is reached; Step 4: The synthesized body motion features and the reconstructed motion features are independently evaluated by the discriminator to generate classification results; Step 5: Construct adversarial loss based on the real motion features after denoising, construct reconstruction loss based on the classification results, construct diffusion loss based on the reconstructed motion features, use reconstruction loss and diffusion loss to get the total reconstruction loss, and construct matching loss based on the synthesized body motion features; The diffusion generative adversarial network model is optimized using adversarial loss, reconstruction loss, diffusion loss and matching loss to obtain the optimized diffusion generative adversarial network model; The final classification result is obtained based on the optimized diffusion generative adversarial network model.

[0007] Compared with the prior art, the present invention has the following beneficial effects: 1. This paper combines the global feature capture capability of GAN with the local detail optimization mechanism of the diffusion model to achieve a double breakthrough in the kinematic rationality and detail richness of the generated action, solving the problem that the traditional single model is difficult to balance global coherence and local refinement; 2. The present invention constructs a joint evaluation system in spatiotemporal dimensions through a multi-scale patch generative adversarial network (Patch-GAN) discriminator architecture, which enables the discriminator to simultaneously perceive the microscopic dynamic features and macroscopic motion patterns of action sequences, and exhibits higher generation quality than traditional single-scale discriminators; 3. The present invention uses a shallow U-Net and attention collaborative architecture to control the model receptive field within the range of 8-12 frames, while ensuring the continuity of the action, it supports the generation of sequences of any length, and successfully achieves the completion of 60-second long sequence synthesis in a single training; 4. The present invention realizes the adaptive fusion of local features of generated samples and reference samples in adversarial training through the matching-mixing dynamic adjustment mechanism, which improves the inter-diversity index of synthetic actions by 15.8% (please refer to Table 4 in the specification), while maintaining more than 97% action coverage (please refer to Table 2 in the specification); 5. The present invention effectively solves the mode collapse problem under complex bone topology through the joint optimization of gradient penalty and residual structure.

[0008] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flowchart of the steps of the single action generation deep learning method based on the diffusion-generation adversarial framework proposed in the present invention.

[0010] Figure 2 Schematic diagram of the SinMDGan framework of the single action generation deep learning method based on the diffusion-generation adversarial framework proposed in the present invention.

[0011] Figure 3 Motion representation graph of the single action generation deep learning method based on the diffusion-generative adversarial framework proposed in this invention.

[0012] Figure 4 This is a spatial action combination graph of the single action generation deep learning method based on the diffusion-generation adversarial framework proposed in this invention.

[0013] Figure 5 Motion expansion graph for the proposed single action generation deep learning method based on the diffusion-generative adversarial framework.

[0014] Figure 6 Spatial composition of the single action generation deep learning method based on the diffusion-generation adversarial framework proposed in this invention.

[0015] Figure 7 This is a motion sequence generation graph of the single action generation deep learning method based on the diffusion-generative adversarial framework proposed in this invention. DETAILED DESCRIPTION

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout are the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0017] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to provide some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0018] See also Figure 1 The embodiment of the present invention proposes a single action generation deep learning method based on a diffusion-generation adversarial framework, the method comprising the following steps: Step 1: Build the generator and discriminator based on the Diffusion-GAN architecture, build the matching and mixing modules based on the matching and mixing mechanism, and the generator, discriminator, matching and mixing modules constitute the Diffusion-GAN model; The generator includes a diffusion model, which is combined with a shallow U-Net architecture. The discriminator operates in a multi-scale manner. The matching and mixing modules include a matching module and a mixing module.

[0019] Step 2: Divide the motion sequence to obtain dynamic features and static features; The dynamic features are transformed to obtain a general motion representation, a foot contact label is added to the general motion representation, and the foot contact is calculated through forward kinematics to obtain the calculated dynamic features.

[0020] See also Figure 3 In step 2, the dynamic features are transformed to obtain a general motion representation, and the corresponding process relationship is as follows: ; in, Indicates general movement expression, The dimension is A real matrix of Indicates the total number of frames. Represents the feature dimension of each frame; In the steps of adding foot contact labels to the general motion representation and performing foot contact calculations through forward kinematics to obtain the calculated dynamic features, the corresponding process has the following relationship: ; in, Indicates foot contact label, Indicates the number of joints that contact the ground at high frequency during movement. Indicates the number of bone joints, Indicates foot No. Joints in frame speed, represents forward kinematics, Represents the bone parameters in the static feature, Indicates The ground height of the frame, represents the first Frame time joint The dynamic feature vector of .

[0021] It should be noted that adding foot contact tags To reduce the common foot sliding artifacts, general motion representation It is the motion data in skeletal motion (dynamic features), such as joint speed and position, Indicates the ground height or contact threshold at the nth frame. This value changes dynamically.

[0022] Further, in Figure 3 In , the 3D skeleton structure of the 6D rotation representation is presented as a cube, vividly showing the 3D skeleton representation at the intersection of time and space.

[0023] It should be noted that represents joint rotation, represents the local root joint displacement, represents the number of rotation features of each joint, Represents rotation feature + local contact feature (used to describe the motion state of the skeleton in space and time), Represents rotation feature + local contact feature + displacement feature (generally the displacement feature of the root node is set to 3). Indicates the total number of frames The subset frame in is a smaller unit, Represents a skeleton.

[0024] Step 3: Add Gaussian random noise to the calculated dynamic features, and input them into the generator for denoising through the diffusion model to obtain the denoised real action features; Based on the matching and mixing module, the denoised real action features are synthesized using the measurement of action similarity and the average voting strategy to obtain the synthesized body action features. Based on the calculated dynamic features, combined with the input action sequence, and input into the Markov denoising mechanism in the dedicated local attention layer of the diffusion model, denoising is performed in an iterative form, and the reconstructed action features are output when the maximum number of iterations is reached.

[0025] See also Figure 2 In step 3, based on the matching and mixing module, the denoised real action features are synthesized using the measured action similarity and average voting strategy to obtain the synthesized body action features. The corresponding process has the following relationship: ; in, represents the synthesized body movement characteristics, represents the matching and mixing modules, represents the generator network, Indicates time Represents a sequence.

[0026] It should be noted that The matching module in

[14] measures the similarity between example action patches and synthesized action patches, while the mixing module combines the collected action patches using an average voting strategy to construct synthesized part-body action features.

[0027] Based on the calculated dynamic features, combined with the input action sequence, the Markov denoising mechanism in the special local attention layer of the diffusion model is used to perform denoising in an iterative form. When the maximum number of iterations is reached, the reconstructed action features are output. The corresponding process has the following relationship: ; in, represents the probability distribution function in the Markov denoising process, represents a Gaussian distribution, represents the parameter that controls the intensity of noise addition during the diffusion process, Represents the identity matrix.

[0028] Furthermore, the diffusion model uses a probabilistic generation mechanism to effectively capture complex details in the data distribution. By gradually introducing and eliminating noise, the diversity of generated samples is enhanced. In contrast, traditional generative adversarial networks (GANs) rely on generators that directly synthesize samples, which sometimes leads to mode collapse and fails to capture complex motion changes. In the generator design of the present invention, the diffusion model is combined with a shallow U-Net architecture and a dedicated local attention layer. Experiments show that in the U-Net structure, the use of a narrower receptive field can significantly improve the realism and credibility of motion in the motion field.

[0029] It should be noted that in Figure 2 In the noise removal: the model is removed from the noise Starting from, gradually recover the data through parameterized network learning, going through the intermediate state → →...→ , and finally get the denoising result .by → Steps as an example: When the model is completed When denoising the state, the network calculates the deviation between the current state and the target distribution, and adjusts the noise prediction gradient to make Update to a purer one along the decreasing direction of the loss function state.

[0030] Step 4: Use the discriminator to independently evaluate the synthesized body motion features and the reconstructed motion features to generate classification results.

[0031] Furthermore, the discriminator of a traditional GAN ​​typically outputs a single scalar value to classify the input as real or fake. However, this overly simplistic approach may lead to mode collapse, where the generator overfits the limited set of training samples, thereby reducing the diversity of the generated output. To address this problem, the present invention adopts MultiScaleGAN, an enhanced version of PatchGAN, in which multiple discriminators operate at different scales, independently evaluate the input and produce real or fake classifications, and the total loss is calculated as the weighted sum of multi-scale evaluations, thereby improving the robustness and diversity of motion generation.

[0032] Step 5: Construct adversarial loss based on the real motion features after denoising, construct reconstruction loss based on the classification results, construct diffusion loss based on the reconstructed motion features, use reconstruction loss and diffusion loss to get the total reconstruction loss, and construct matching loss based on the synthesized body motion features; The diffusion generative adversarial network model is optimized using adversarial loss, reconstruction loss, diffusion loss and matching loss to obtain the optimized diffusion generative adversarial network model; The final classification result is obtained based on the optimized diffusion generative adversarial network model.

[0033] In step 5, the adversarial loss is constructed based on the real action features after denoising. The relationship between the corresponding process is as follows: ; in, Represents resistance to loss, Represents the distribution from real data Sample the first data The expected value of Represents the noise distribution The noise vector is sampled from The expected value of represents the gradient penalty term, Indicates the situation where the penalty gradient norm deviates from 1, represents the distribution from linear interpolation Sample the second data The expected value of represents the gradient operator, Represents the discriminator.

[0034] It should be noted that Indicates that the data distribution A data randomly extracted from The expected value of Represents the noise distribution A noise vector randomly drawn from The expected value of Indicates that the data distribution The expected value of the second data randomly selected from .

[0035] The reconstruction loss is constructed based on the classification results, and the relationship between the corresponding process is as follows: ; in, represents the reconstruction loss, represents the number of different states of the discriminator, Represents a generator Distribution Samples generated by sampling And calculate the expected value, Indicates The discriminator generates samples The discrimination score of .

[0036] Based on the reconstruction action, the diffusion loss is constructed, and the relationship between the corresponding process is as follows: ; in, represents the diffusion loss, represents the sampling time step in the discrete uniform distribution , from 1 to the total time step Randomly select with equal probability and generate expected value, Represents action sequence data before processing.

[0037] It should be noted that Indicates the total number of steps in the diffusion and denoising process. Specifically, in diffusion, it is represented by the number of steps in which the input sequence is denoised until it reaches a Gaussian distribution; in denoising, it is represented by the number of steps in which denoising is continued until the noise is removed.

[0038] Based on the synthesized body motion features, the matching loss is constructed, and there is also a patch distance matrix calculation for the skeleton part. The corresponding relationship in the process is as follows: ; in, Indicates The motion is generated from the patch distance matrix of the skeleton parts. Patch and reference motion The normalized distance of patches, represents the first The first part of the skeleton Sports patches, represents the first The first part of the skeleton Sports patches, represents the hyperparameter, Represents all patches in the generated motion Find the reference patch in The closest patch, i.e. the smallest L2 distance squared.

[0039] Furthermore, the total reconstruction loss is obtained by using the reconstruction loss and diffusion loss. The relationship between the corresponding process is as follows: ; in, represents the total reconstruction loss.

[0040] The complete training objectives are summarized below: ; represents the total loss, represents the hyperparameters of the balanced adversarial objective, Represents the hyperparameters of the reconstruction objective.

[0041] For benchmarking, the proposed model is compared with SinMDM and Ganimator, as the SinMDM model and Ganimator model represent the state-of-the-art baselines. To ensure that the evaluation is valid and fair, the same metrics used in the established benchmarks are adopted.

[0042] The evaluation framework of the present invention prioritizes models that exhibit a balanced score distribution on multiple metrics rather than models that excel in a single aspect. To maintain rigor, this example follows the established methods in previous studies and is consistent with the metric evaluation method introduced in SinMDM. Specifically, the standard harmonic mean metric recognized in machine learning is used to provide a comprehensive performance evaluation.

[0043] To ensure comparability between indicators, a standardization procedure was applied using the following steps: Normalized range: Scores are scaled between zero and the maximum value of each corresponding indicator; Maximum estimation: If there is no true maximum, the 90th percentile of the observed scores is used to maximize the approximation; Directional adjustment: For indicators where lower scores are preferred, the scores are transformed by subtracting the obtained value from the determined maximum value; allowing the inclusion of negative values ​​while preserving the interpretability of the indicator; The formula for calculating the harmonic mean value is: ; in, represents the harmonic mean value, Represents the total number of indicators, The harmonic mean formula Item value; Evaluation indicators: The present invention uses the following five indicators to evaluate performance: Metric a. Coverage - measures the input motion sequence The proportion of time windows that are successfully reproduced in the synthetic motion; Indicator b. Global diversity - calculation and The distance between represents the minimization of the input sequence Mosaic of distance. The function is equivalent to Mapped to an optimization space, it means minimizing the input sequence The mosaic of distance; Metric c. Local diversity - measures the average distance between a time window in the synthetic motion and the nearest corresponding window in the input sequence; Metric d. Mutual diversity - quantifies the differences between different synthetic motion samples; Metric e. Intra-diversity difference - calculates the difference between the intra-sequence diversity of the synthetic motion and the intra-sequence diversity of the original input motion; For indicators a to d, higher scores indicate better performance, while for indicator e, lower scores indicate better performance.

[0044] Table 1: Quantitative comparison of the model of the present invention with Ganimator and SinMDM

[0045] As can be seen from Table 1, each metric is calculated separately on the benchmark action and then averaged. The results show that, except for diversity, the proposed model outperforms Ganimator in all metrics. It is worth noting that SinMDM consistently outperforms Ganimator in almost all metrics, especially in the harmonic mean metric, highlighting its excellent overall performance.

[0046] Table 2: Comparison results of the model of the present invention with MotionTexture and acRNN on Salar-Dancing motion data

[0047] As can be seen from Table 2, quantitative analysis shows that the model of the present invention achieves state-of-the-art performance in terms of global diversity and harmonic mean, outperforming existing methods. In addition, the coverage and local diversity indicators show comparable performance levels to Ganimator and SinMDM, further verifying the effectiveness of the method of the present invention. The model of the present invention is in the leading position in terms of global diversity and harmonic mean scores, and achieves comparable scores in terms of coverage and local diversity.

[0048] Given a reference motion sequence and a region of interest (ROI) mask, the goal is to synthesize a new motion. The region of interest is generated by random noise, while the rest of the region is as close to the given motion as possible. The corresponding process has the following relationship: ; in, Indicates the new movement of the synthesis, represents element-wise multiplication, A mask representing the region of interest (a binary tensor marking the area to be synthesized), Represents the reference motion sequence.

[0049] Our model aims to achieve seamless transitions between original and synthesized segments. The reference motion can be an arbitrary sequence, independent of the training data. When using binary masks to define regions of interest, abrupt transitions can occur between fixed and generated segments due to discontinuities. To address this, we introduce a linear interpolation boundary; specifically, fixed motion segments remain unchanged, while sampled regions to be synthesized are gradually blended; during inference, we iteratively refine the motion, ensuring that the regions are indexed by the mask. The relevant parts are merged into In; expressed as: ; like , the ROI is not modified and the output remains the same as the reference example. , the specified area is synthesized.

[0050] Temporal composition involves filling or extending the frames within a motion sequence; this concept is demonstrated in two key cases: in-between frames and motion extension.

[0051] In-between frames play a vital role in animation by creating smooth transitions between keyframes. Our model is good at producing realistic in-between motions using only the initial and final poses. Figure 4 As shown in , the synthesized in-between frames create a smooth transition, preserving the realism of the motion and extending the duration of the sequence. Figure 4 In , the initial and final motions are kept constant, while the intermediate motions are synthesized to achieve realistic and diverse variations.

[0052] See also Figure 5 , motion expansion expands a given motion sequence, similar to image inpainting. Figure 5 Motion stretching is illustrated in Figure 1, where the reference motion is extended forward and backward. In this case, the ROI mask is set to zero for the center frame, while the outer regions are assigned zeros. Motion stretching not only extends the sequence, but also reorganizes the motion dynamics of each frame, resulting in diverse and high-quality continuous motion; the motion on the sides is newly generated, while the motion in the middle represents the reference motion; the extended sequence demonstrates diverse synthesis from a single input.

[0053] See also Figure 6 Motion synthesis can also be applied spatially by specifying specific joint indices in the ROI mask. Figure 6 In , editing of upper body motion using ROI masks is illustrated; the upper body motion is guided by the reference motion, while the lower body motion is synthesized for consistency.

[0054] See also Figure 7 , the model of the present invention is able to generate motion sequences of variable length, including extended animations, without the need for additional training. Since the model has a narrow receptive field, the model of the present invention can generate sequences of arbitrarily long lengths. Figure 7 In , an example of long action generation is shown; in it, a one-minute animation is synthesized based on the learned salad-dance sequence; the learned sequence is the salad dance, and the synthesized action is a 60-second sequence.

[0055] In order to evaluate the contributions of the two key components, the multi-scale Patch-GAN structure and the Match and Blend Mechanism, the present invention conducts an ablation study, separates each component, and discusses the impact of the ablation module on the results.

[0056] Table 3: Internal diversity difference results

[0057] As can be seen from Table 3, the lower the mean and variance values, the smaller the difference between the internal diversity of the synthesized motion and the internal diversity of the input motion, thus achieving better global consistency. In addition, adding multi-scale Patch-GAN can significantly reduce the internal diversity difference, thereby improving global motion consistency, highlighting its advantages in enhancing motion realism and structural coherence.

[0058] Table 4: Internal diversity results model

[0059] As can be seen from Table 4, the matching and mixing mechanism enhances the diversity between sub-windows in the motion sequence, thereby improving the diversity index. The introduction of the matching and mixing mechanism can improve the diversity metric, which measures the differences between sub-windows in the generated motion. The matching and mixing mechanism enhances the diversity of the synthesized motion while preserving the coherence of each sequence.

[0060] In this experiment, we replaced the traditional Patch-GAN used in Ganimator with a multi-scale Patch-GAN. The original Ganimator framework uses a Patch-GAN classifier to prevent overfitting by limiting the receptive field of the discriminator. The classifier assigns confidence values ​​at the patch level, and the final discriminator output is calculated as the average confidence of all patches; while this method is effective for evaluating local details, it may ignore global consistency in the generated motion.

[0061] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0062] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0063] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A single action generation deep learning method based on a diffusion-generation adversarial framework, characterized in that: The method comprises the following steps: Step 1: Construct a generator and a discriminator based on the diffusion generative adversarial network architecture, and construct a matching and mixing module based on the matching and mixing mechanism. The generator, discriminator, matching and mixing modules constitute a diffusion generative adversarial network model. The generator includes a diffusion model, the diffusion model is combined with a shallow U-Net architecture, the discriminator operates in a multi-scale manner, and the matching and mixing module includes a matching module and a mixing module; Step 2: Divide the motion sequence to obtain dynamic features and static features; The dynamic features are converted to obtain a general motion representation, a foot contact label is added to the general motion representation, and foot contact is calculated by forward kinematics to obtain the calculated dynamic features; Step 3: Add Gaussian random noise to the calculated dynamic features, and input them into the generator for denoising through the diffusion model to obtain the denoised real action features; Based on the matching and mixing module, the denoised real action features are synthesized using the measurement of action similarity and the average voting strategy to obtain the synthesized body action features. Based on the calculated dynamic features and the input action sequence, the Markov denoising mechanism in the special local attention layer of the diffusion model is used to perform denoising in an iterative manner, and the reconstructed action features are output when the maximum number of iterations is reached; Step 4: The synthesized body motion features and the reconstructed motion features are independently evaluated by the discriminator to generate classification results; Step 5: Construct adversarial loss based on the real motion features after denoising, construct reconstruction loss based on the classification results, construct diffusion loss based on the reconstructed motion features, use reconstruction loss and diffusion loss to get the total reconstruction loss, and construct matching loss based on the synthesized body motion features; The diffusion generative adversarial network model is optimized using adversarial loss, reconstruction loss, diffusion loss and matching loss to obtain the optimized diffusion generative adversarial network model; The final classification result is obtained based on the optimized diffusion generative adversarial network model.

2. The single action generation deep learning method based on the diffusion-generation adversarial framework according to claim 1 is characterized in that: In step 2, the dynamic features are transformed to obtain a general motion representation, and the relationship between the corresponding process is as follows: ； in, Indicates general movement expression, The dimension is A real matrix of Indicates the total number of frames. Represents the feature dimension of each frame; In the steps of adding foot contact labels to the general motion representation and performing foot contact calculations through forward kinematics to obtain the calculated dynamic features, the corresponding process has the following relationship: ； in, Indicates that the foot is touching the label, Indicates the number of joints that contact the ground at high frequency during movement. Indicates the number of bone joints, Indicates foot No. Joints in frame speed, represents forward kinematics, Represents the bone parameters in the static feature, Indicates The ground height of the frame, represents the first Frame time joint The dynamic feature vector of .

3. The single action generation deep learning method based on the diffusion-generation adversarial framework according to claim 2 is characterized in that: In step 3, based on the matching and mixing module, the denoised real action features are synthesized using the measured action similarity and average voting strategy to obtain the synthesized body action features. The corresponding process has the following relationship: ； in, represents the synthesized body movement characteristics, represents the matching and mixing modules, represents the generator network, Indicates time Represents a sequence.

4. The single action generation deep learning method based on the diffusion-generation adversarial framework according to claim 3 is characterized in that: Based on the calculated dynamic features, combined with the input action sequence, the Markov denoising mechanism in the special local attention layer of the diffusion model is used to perform denoising in an iterative form. When the maximum number of iterations is reached, the reconstructed action features are output. The corresponding process has the following relationship: ； in, represents the probability distribution function in the Markov denoising process, represents a Gaussian distribution, represents the parameter that controls the intensity of noise addition during the diffusion process, Represents the identity matrix.

5. The single action generation deep learning method based on the diffusion-generation adversarial framework according to claim 4 is characterized in that: In step 5, the adversarial loss is constructed based on the real action features after denoising, and the relationship between the corresponding process is as follows: ； in, Represents resistance to loss, Represents the distribution from real data Sample the first data The expected value of Represents the noise distribution The noise vector is sampled from The expected value of represents the gradient penalty term, Indicates the situation where the penalty gradient norm deviates from 1, represents the distribution from linear interpolation Sample the second data The expected value of represents the gradient operator, Represents the discriminator.

6. The single action generation deep learning method based on the diffusion-generation adversarial framework according to claim 5 is characterized in that: The reconstruction loss is constructed based on the classification results, and the relationship between the corresponding process is as follows: ； in, represents the reconstruction loss, represents the number of different states of the discriminator, Represents a generator Distribution Samples generated by sampling And calculate the expected value, Indicates The discriminator generates samples The discrimination score of .

7. The single action generation deep learning method based on the diffusion-generation adversarial framework according to claim 6 is characterized in that: Based on the reconstructed motion features, the diffusion loss is constructed, and the relationship between the corresponding process is as follows: ； in, represents the diffusion loss, represents the sampling time step in the discrete uniform distribution From 1 to the total time step Randomly select with equal probability and generate expected value, Represents action sequence data before processing.

8. The single action generation deep learning method based on the diffusion-generation adversarial framework according to claim 7 is characterized in that: Based on the synthesized body motion features, the matching loss is constructed, and there is also a patch distance matrix calculation for the skeleton part. The corresponding relationship in the process is as follows: ； in, Indicates The motion is generated from the patch distance matrix of the skeleton parts. Patch and reference motion The normalized distance of patches, represents the first The first part of the skeleton Sports patches, represents the first The first part of the skeleton Sports patches, represents the hyperparameter, Represents all patches in the generated motion Find the reference patch in The closest patch.

Images