CN-121999240-A - Diffusion-driven channel self-adaptive point cloud semantic communication method

CN121999240ACN 121999240 ACN121999240 ACN 121999240ACN-121999240-A

Abstract

The invention provides a diffusion-driven channel self-adaptive point cloud semantic communication method, which belongs to the technical field of wireless communication and comprises the following steps of acquiring a training point cloud data set, constructing a point cloud feature extraction network based on a hypergraph convolutional neural network as a semantic encoder, constructing a channel self-adaptive enhancement module as a channel decoder, constructing a channel self-adaptive recovery module at a receiving end as the channel decoder, constructing a diffusion reconstruction network as the semantic decoder, performing end-to-end joint training on the point cloud feature extraction network, the channel self-adaptive enhancement module, the channel self-adaptive recovery module and the diffusion reconstruction network, predicting original point cloud distribution by the diffusion reconstruction network, and constructing a loss function based on a chamfer distance between the predicted point cloud distribution and the original point cloud. In the communication process, the reconstructed semantic features are input into a diffusion reconstruction network to reconstruct an original point cloud structure. The high-order semantic features of the point cloud can be effectively extracted to improve the adaptability of the method to the incomplete point cloud.

Inventors

NONG LIPING
WANG JUNYI
ZHU GUIFA
PENG JIE
LIU JINLING
HUANG YIPING

Assignees

南宁桂电电子科技研究院有限公司
桂林电子科技大学

Dates

Publication Date: 20260508
Application Date: 20251225

Claims (8)

1. A diffusion-driven channel self-adaptive point cloud semantic communication method is characterized by comprising the following steps: Step 1, acquiring a training point cloud data set; Step 2, constructing a point cloud characteristic extraction network based on a hypergraph convolutional neural network at a transmitting end to serve as a semantic encoder, and constructing a channel self-adaptive enhancement module to serve as a channel decoder; step 3, constructing a channel self-adaptive recovery module at a receiving end as a channel decoder, and constructing a diffusion reconstruction network as a semantic decoder; step 4, performing end-to-end joint training on a point cloud feature extraction network, a channel self-adaptive enhancement module, a channel self-adaptive recovery module and a diffusion reconstruction network, wherein in the training process, the diffusion reconstruction network predicts original point cloud distribution, and builds a loss function based on the chamfer distance between the predicted point cloud distribution and the original point cloud so as to improve the point cloud reconstruction quality in different channel states; And 5, in the communication process, performing hypergraph modeling on point characteristics in point cloud data by a transmitting end to construct a hypergraph structure, then inputting the point cloud characteristics and the hypergraph structure into a semantic encoder together, extracting high-order semantic characteristics of point cloud, inputting the high-order semantic characteristics into the channel self-adaptive enhancement module, enhancing and clipping the semantic characteristics according to the signal-to-noise ratio and bandwidth of a current channel, generating code words through power normalization, and transmitting through a physical channel, after receiving the code words, firstly performing zero filling at the tail end of the code words to recover the dimensions, then reconstructing the semantic characteristics through the channel self-adaptive recovery module, and inputting the reconstructed semantic characteristics into a diffusion reconstruction network to reconstruct the original point cloud structure.
2. The diffusion-driven channel adaptive point cloud semantic communication method according to claim 1, wherein the specific process of step 1 is as follows: The point cloud data set is ShapeNet point cloud data set, comprises a plurality of point cloud models and covers a plurality of common object categories, wherein the sample number ratio of the training set to the verification set to the test set is 35:5:1.
3. The diffusion-driven channel adaptive point cloud semantic communication method according to claim 1, wherein the specific process of constructing the point cloud feature extraction network in the step 2 is as follows: The point cloud feature extraction network captures high-order features of the point cloud by utilizing a hypergraph convolution neural network, the point cloud feature extraction network consists of a plurality of feature extraction modules and a maximum pooling layer which are sequentially connected, each feature extraction module comprises a hypergraph convolution layer and a feature enhancement module, the hypergraph convolution layer extracts high-order semantic features of the point cloud through hypergraph convolution, and a hypergraph convolution formula is as follows: Wherein Representing the characteristics of the point cloud of the input, Representing the output characteristics of the layer, When (when) In the time-course of which the first and second contact surfaces, , A parameter representing the learning of the layer, Is hypergraph association matrix Diagonal matrix of vertex degrees, internal elements defined as , Is hypergraph association matrix Superside diagonal matrix, internal elements defined as To improve the expression capacity and nonlinear mapping capacity of the model, the traditional parameters are used for The method is replaced by a full connection layer, so that a hypergraph convolution formula is as follows: Wherein Subsequently, the first and second substrates were subjected to a vacuum, The feature enhancement module is further enhanced by the feature enhancement module, the feature enhancement module is composed of an SE layer and an FC layer, the SE layer adaptively adjusts the weight of a feature channel to highlight semantics related to tasks, global pooling is firstly applied to aggregate channel information, then two FC layers are used to capture inter-channel dependency relationships, finally sigmoid activation is used to generate feature weights, meanwhile, the FC layer keeps detailed information of a previous module, the risk of losing key characteristics in the transmission process is reduced, and the corresponding process is expressed as follows: In order to more effectively utilize the local and global features of the point clouds under different receptive fields, the outputs of the first four layers are input to the fifth layer together to adjust the feature representation and match the dimension of the target channel, the global features are aggregated by using a maximized pooling layer, and the cloud semantic representation is output Gao Jiedian 。
4. The diffusion-driven channel adaptive point cloud semantic communication method according to claim 1, wherein the channel adaptive enhancement module in step 2 is constructed by the following specific processes: channel adaptive enhancement module average signal-to-noise ratio SNR and available bandwidth according to channel feedback Point cloud semantic features Performing adaptive enhancement and dimension adjustment, firstly, SNR and bandwidth of channel feedback are calculated Feedback is input into two independent FC layers to obtain two L-dimensional feature vectors And Then, two feature vectors and high-order point cloud semantic features are combined Splicing to obtain vector Then enhancing the characteristic through residual structure to obtain Will (i) be Inputting to another FC layer, generating characteristic weights and biases through a Sigmoid activation function to realize nonlinear enhancement, wherein the corresponding process is expressed as follows: finally, adopting semantic clipping strategy to match the current channel bandwidth limit, only reserving Front part Data to obtain transmission data 。
5. The diffusion-driven channel adaptive point cloud semantic communication method according to claim 1, wherein the channel adaptive recovery module in step 3 is constructed by the following specific processes: Semantic information received by receiving end Is that the sender outputs semantic information To mitigate the effects of channel noise and to accommodate fluctuations in available bandwidth, a channel adaptive recovery module pair Feature reconstruction is performed to provide a more stable input to a subsequent flooding reconstruction network by first inferring the current available bandwidth based on the length of the received signal And receives semantic information Zero padding at the end of (2) from the original dimension Expanded to Obtaining Subsequently, the SNR and bandwidth of the channel feedback Respectively inputting two FC layers to obtain two feature vectors And And is connected with Splicing together to obtain splicing characteristics Finally, through a nonlinear transformation pair with a gating mechanism Processing to obtain reconstructed semantic features The corresponding calculation mode is that Wherein each is independent Are used to extract and map global semantic features, The activation function then generates dynamic weights for use in determining the current SNR and bandwidth The importance of the feature is adjusted to achieve adaptive recovery of the channel state.
6. The diffusion-driven channel adaptive point cloud semantic communication method according to claim 1, wherein the specific process of the diffusion reconstruction network construction in the step 3 is as follows: The diffusion reconstruction network utilizes semantic features Guiding the generation of point cloud, utilizing the strong generation capability of a diffusion model to ensure that the reconstructed point cloud is consistent with the original point cloud, thereby realizing point cloud reconstruction, wherein the diffusion reconstruction network comprises the principle of forward process and reverse process of the diffusion model and a point cloud prediction network, and the target of the diffusion reconstruction network is to control the reverse diffusion process to generate the point cloud distribution represented by the diffusion reconstruction network through semantic features, so that the semantic features are generated The condition information is combined into a back diffusion process, and the generation formula of the back diffusion process is as follows: Wherein , Is expressed as a diffusion time step The distribution of the noise point cloud in the time, Is a predefined time-dependent Scalar of (2), pair of Step-by-step denoising to obtain reconstruction point cloud When (when) When the number is larger than the set value, the sampling process can go through more sampling steps, so that the reconstruction process becomes slow, the practical applicability in the point cloud semantic communication is limited, and the back diffusion process is subjected to Markov, namely, the back diffusion process is removed For the following From an original sample sequence according to a corresponding principle Sequentially selecting subsequences Thereby reducing the reconstruction time, wherein The formula of the back diffusion process is updated as follows According to the following The next sampling point is obtained: Wherein , The original point cloud distribution predicted by the point cloud prediction network is obtained; point cloud prediction network utilizes point cloud semantic features Noisy point cloud Time step As input to predict origin cloud distribution The point cloud prediction network comprises two deconvolution layers, an FC layer and a plurality of point cloud prediction layers, wherein first, the point cloud semantic features are formed Gao Weidian cloud features obtained by two deconvolutions , Essentially a 1D transpose convolution with unit steps, followed by a time step Obtaining diffusion super-parameters And is connected with And Together spliced, input the FC layer to further obtain an L-dimensional time-step feature vector The corresponding formula is Finally, the point cloud is characterized Gao Weidian cloud features Feature vector of time step Semantic features Inputting a plurality of point cloud prediction layers connected in sequence together to obtain predicted original point cloud, wherein the point cloud prediction layers comprise a feature aggregation module, a feature enhancement module, a self-attention module and a semantic fusion module, the feature aggregation module comprises two FC layers, and the first FC layer is to be used for From the dimension of (2) Steering The second FC layer computes global statistics of features, capturing broader semantic context semantic features of the input data by aggregating features from the entire input space I.e. The feature enhancement module consists of an FC layer and uses time step feature vectors To semantic features Enhancing and filtering information to obtain time sequence enhancement characteristics Thereby more effectively retaining the most relevant characteristic information of the current time step, and the expression is as follows: the self-attention module consists of an FC layer and sigmoid activation function, and is controlled by a trainable gating mechanism and captured by capturing The dependency between features dynamically weights the feature channels to obtain While enhancing the information features while suppressing noise and uncorrelated features, the corresponding process is expressed as: when (when) In the time-course of which the first and second contact surfaces, The semantic fusion module strengthens time sequence Generating a gating signal sequentially through the FC layer and the sigmoid activation function I.e. And then to And FC (fiber channel) Applying corresponding gating signals and by Maintaining continuity of point cloud characteristics, obtaining output characteristics of corresponding layers, namely By superposing a plurality of point cloud prediction layers, the network can gradually filter noise in the point cloud characteristics, and simultaneously provide stronger semantic guidance for the characteristics of different dimensions, and each layer of point cloud prediction layer inherits and refines the characteristic representation of the previous layer to form a characteristic evolution chain with memory capacity, so that the point cloud prediction can predict the original point cloud from noise input.
7. The diffusion-driven channel adaptive point cloud semantic communication method according to claim 1, wherein the specific process of the step 4 is as follows: the point cloud characteristic extraction network, the channel self-adaptive enhancement module, the channel self-adaptive recovery module and the diffusion reconstruction network are trained in an end-to-end mode, the aim is to minimize the distance between the point cloud distribution predicted by the diffusion reconstruction network and the ground truth value distribution, in order to measure the difference between the two distributions, the chamfer distance CD is selected as a loss function, and the point cloud is predicted And ground real point cloud The CD loss function is defined as: 。
8. The diffusion-driven channel adaptive point cloud semantic communication method according to claim 1, wherein the specific process of the step 5 is as follows: Transmitting point cloud data over an additive white gaussian noise channel Wherein The representation space has N points in three-dimensional coordinates, Representation The feature set associated with each point only transmits the three-dimensional coordinates of the point cloud, thus the attributes Vector set to all ones, and At the transmitting end, for the input point cloud Modeling using hypergraph methods to obtain hypergraphs Wherein Representing a set of points cloud vertices, Is a set of hyperedges that are to be rounded, Is a set of weights for all supersides, for vertices in a point cloud Calculate affinity distances to all other vertices Wherein is the vertex And The euclidean distance between the two, Is the average distance between all vertices, followed by selection of the distance vertex Recently, the method of the present invention The vertexes form a superside, and according to the vertex set and the superside set, an incidence matrix of the point cloud supergraph can be constructed The elements within it are defined as: Then the point cloud characteristic extraction network correlates the hypergraph with the matrix Point cloud As input and encoded into semantic features Wherein For maximum available channel bandwidth, then, semantic features are added By the channel self-adaptive enhancement module, the channel self-adaptive enhancement module is based on SNR and bandwidth The method enhances the characteristics and cuts out the signals to obtain the transmission signals ; Transmitting through channel, signal received by receiving end Is that Noisy version of (1), channel adaptive enhancement module reconstruction Obtaining point cloud semantic features Diffusion reconstruction network uses point cloud semantic features As condition information, and then reconstructing the point cloud 。

Description

Diffusion-driven channel self-adaptive point cloud semantic communication method Technical Field The invention relates to the technical field of wireless communication, in particular to a diffusion-driven channel self-adaptive point cloud semantic communication method. Background The 3D point cloud is a three-dimensional data format generated by a laser radar and a depth sensor, can provide more accurate three-dimensional scene information, and is widely applied to a plurality of key fields such as automatic driving, geographic Information Systems (GIS), virtual and augmented reality (VR/AR) and the like. However, point cloud data is typically massive, as each scene may contain hundreds of thousands to millions of 3D points, each storing spatial coordinates and other attributes, such as color or reflectivity. The transmission of such large-scale point cloud data requires extremely high bandwidth and delay, exceeding the actual capacity of the current 5G network, thereby resulting in transmission bottlenecks. In order to realize efficient transmission of point cloud, a series of researches are carried out on a point cloud compression method by a learner. Point cloud compression methods can be broadly divided into three categories, octree-based methods, voxel-based methods, and projection-based methods. Octree-based methods divide and encode the geometry of a three-dimensional space to achieve point cloud compression, e.g., voxel-based methods attempt to voxel a 3D space so that the point cloud is distributed in an artificial regularized space. Projection-based methods map 3D point clouds into 2D space so that 2D convolution can be achieved, and then compress these images using conventional two-dimensional image coding techniques. These compression methods provide structured and compact representations of point clouds, enabling efficient storage, transmission, and compression of data to some extent. However, the compression ratio is relatively low, and the accurate bit level transmission is relied on, so that the cliff effect and the horizontal effect are caused, and the cliff effect and the horizontal effect together cause unstable communication performance, so that the overall reliability of the point cloud transmission is limited. Compared with the traditional compression method, the semantic communication has stronger robustness and adaptability. Semantic communication is a new communication paradigm that focuses on transmitting task intent-driven semantic information, rather than pursuing bit-level accuracy. On the one hand, the amount of transmission data is reduced by extracting the semantics of the message, and on the other hand, the reliability of communication is improved by deducing the context semantics. Unlike the extensive study of semantic communications in these euclidean data, the related study of semantic communications in point cloud data is still in an early stage. This is because unstructured, unordered, and spatially non-uniform point cloud data present serious challenges to semantic exchange of point clouds. The existing point cloud semantic communication method generally adopts a framework of an encoder and a decoder, the encoder extracts semantic features from input point clouds, and the decoder reconstructs the point clouds according to the received features. Such methods have inherent limitations in that all the information required for accurate reconstruction must be retained, which places stringent demands on the quality and accuracy of the input data. To overcome this limitation, and inspired by the powerful generation and modeling capabilities of diffusion models in image data, some researchers have proposed encoders and diffusion frameworks. These methods integrate semantic features as conditional guides into the diffusion process, enabling models to iteratively reconstruct structured 3D representations from noise, capturing finer geometric details and reducing reliance on precise input data. However, there is still considerable room for improvement in both encoder and decoder and encoder and diffusion methods, including 1) existing methods mostly employ conventional deep learning models, such as PointNet or PointNet ++, as semantic encoder trunks to effectively capture point cloud global shape information, but it is difficult to capture complex point-to-point dependencies and higher-order geometric relationships, resulting in loss of local structural cues, limiting the ability of the model to represent fine-granularity spatial semantics. 2) Existing methods can only be applied under fixed channel conditions, e.g. constant noise level or fixed bandwidth. However, little consideration is given to both dynamic noise and variable bandwidth adaptation. 3) Diffusion model-based methods generally follow a noise prediction paradigm, which not only increases training time, but also makes it difficult to correlate learned noise with reconstructed point clouds, resulting in reduced recon