A lightweight video block partitioning compression method for unstable satellite communication channels
By using an adaptive threshold based on pixel standard deviation, an adaptive number of gray levels, and a channel-aware dual-path residual block partitioning model, the video block partitioning under unstable satellite communication channels is optimized, solving the problem of reduced compression efficiency caused by channel noise interference and achieving efficient video coding.
Patent Information
- Application Number
- CN202610855962.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-15
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-06-15
AI Technical Summary
Under unstable satellite communication channel conditions, the accuracy of existing video coding block partitioning decisions decreases due to channel noise interference, leading to a deterioration in compression efficiency.
A dual-path residual block partitioning model based on pixel standard deviation adaptive threshold, adaptive gray level number, channel confidence weight, and channel awareness is adopted, combined with speculative execution mechanism and ping-pong buffer mechanism, to optimize video block partitioning decision.
Under unstable satellite communication channel conditions, the accuracy of video coding block partitioning decisions is improved, compression efficiency and coding throughput are increased, and the requirements for real-time video coding are met.
Smart Images

Figure CN122513558B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of satellite technology, and more specifically, relates to a lightweight video block partitioning and compression method for unstable satellite communication channels. Background Technology
[0002] Block partitioning technology in video coding is a core element in improving compression efficiency. Existing coding standards such as H.265 / HEVC and H.266 / VVC employ a recursive block partitioning framework based on rate-distortion optimization. This involves dividing the coding tree units layer by layer into quadtrees, binary trees, or ternary trees to adapt to the complex distribution of local textures in the image, thereby minimizing reconstruction distortion under a given bitrate constraint. In terrestrial high-definition video coding scenarios, the above methods rely on image statistical features such as Laplacian gradient and information entropy for partitioning decisions, achieving good compression results. However, when the coding system is deployed on a satellite communication link, the combined effects of orbital altitude, atmospheric refraction, and the Doppler effect cause the satellite channel to exhibit Nakagami-m fading characteristics. The time-varying fading of the channel introduces random noise during pixel acquisition and transmission, severely contaminating the pixel gradient response that originally reflects the image structure. The Laplacian operator, as a second-order differential operator, amplifies the noise a second time, leading to numerous misjudgments in partitioning decisions based on gradient statistics during deep fading. In existing technologies, block partitioning decisions rely solely on image domain features, failing to incorporate channel quality status into the partitioning decision logic. Under deep fading conditions in satellite channels, gradient statistics become dominated by noise and lose their structural representational significance, leading to a significant decrease in partitioning decision accuracy. In other words, existing technologies suffer from the technical problem of decreased accuracy in video coding block partitioning decisions due to channel noise interference under unstable satellite communication channel conditions, resulting in degraded compression efficiency. Summary of the Invention
[0003] In view of this, the present invention provides a lightweight video block partitioning compression method for unstable satellite communication channels, which can solve the technical problem in the prior art where the accuracy of video coding block partitioning decisions decreases due to channel noise interference, leading to degraded compression efficiency.
[0004] This invention is implemented as follows: This invention provides a lightweight video block partitioning and compression method for unstable satellite communication channels, comprising the following steps:
[0005] Calculate the pixel standard deviation for the coding block, compare the pixel standard deviation with the first adaptive threshold, and if the pixel standard deviation is less than the first adaptive threshold, terminate the division of the coding block; otherwise, proceed to the next step.
[0006] The coded block is divided into four sub-blocks. The sub-block entropy value of each sub-block is calculated based on the number of adaptive gray levels. The sum of horizontal entropy values and the sum of vertical entropy values are calculated. The absolute value of the entropy difference is obtained from the absolute value of the difference between the two. The absolute value of the entropy difference is compared with the second adaptive threshold to determine the priority division direction.
[0007] For each sub-block in the priority partitioning direction, scale-adaptive Gaussian pre-smoothing is introduced before the Laplacian gradient calculation. The Laplacian response value of each pixel in each sub-block is multiplied by the channel confidence weight, and the weighted pixel neighborhood difference mean of each sub-block is calculated. Then, the weighted center gradient difference and the weighted edge gradient difference are calculated. The weighted gradient difference ratio is obtained from the ratio of the two. The binary tree partitioning or ternary tree partitioning is determined according to the weighted gradient difference ratio.
[0008] The pixel standard deviation, the sum of horizontal entropy values, the sum of vertical entropy values, the weighted gradient difference ratio, and the channel quality parameters are input into the channel-aware dual-path residual block partitioning model. The model outputs the termination partitioning probability, the priority partitioning direction probability, the binary tree partitioning probability, and the ternary tree partitioning probability. The model also combines the speculative execution mechanism and the ping-pong buffering mechanism to eliminate pipeline bubbles.
[0009] A block boundary pixel copying strategy is adopted to pre-copy the protection band pixels of each coding block to an independent memory area. The pixel data in physical memory is rearranged in conjunction with the Z-order curve data layout. Rate distortion optimization calculation is performed on the partitioning modes that are not skipped, and the partitioning mode with the minimum rate distortion cost is selected for output.
[0010] After completing the processing of the current coding block, process the next coding block in the coding order, and repeat the above steps.
[0011] Specifically, the pixel standard deviation is calculated by taking the square root of the square of the difference between each pixel value and the mean, based on the mean of all pixel values within the coding block.
[0012] The first adaptive threshold is determined by the product of the quantization parameter and the first coefficient, with a reference value of 0.6 for the first coefficient.
[0013] The adaptive grayscale level number is determined by the quantization parameter and the block size, and a dynamic adjustment mechanism based on histogram flatness is introduced to dynamically correct the grayscale level number according to the uniformity of the probability of each grayscale level in the current sub-block histogram.
[0014] The second adaptive threshold is determined by the product of the quantization parameter and the second coefficient, with a reference value of 0.025 for the second coefficient. If the absolute value of the entropy difference is greater than the second adaptive threshold and the sum of the horizontal entropy values is less than the sum of the vertical entropy values, the vertical division mode is selected first. If the absolute value of the entropy difference is greater than the second adaptive threshold and the sum of the horizontal entropy values is greater than the sum of the vertical entropy values, the horizontal division mode is selected first. If the absolute value of the entropy difference is not greater than the second adaptive threshold, all candidates for division directions are retained.
[0015] The Gaussian kernel scale of the scale-adaptive Gaussian pre-smoothing is automatically selected based on the estimated noise variance of the current frame. The larger the estimated noise variance, the larger the Gaussian kernel scale. The value range of the Gaussian kernel scale is 0.5 to 3.0.
[0016] The channel confidence weight is determined by the fading severity parameter obtained by real-time estimation of the Nakagami-m parameter of the current frame signal. The smaller the fading severity parameter, the lower the channel confidence weight.
[0017] The main path of the channel-aware dual-path residual block partitioning model adopts a depthwise separable convolutional residual chain, which contains four depthwise separable convolutional residual blocks; the auxiliary path is a three-layer fully connected network, with input consisting of four-dimensional channel quality parameters composed of signal-to-noise ratio, bit error rate, Doppler estimation value and quantization parameters, and outputting a 64-dimensional channel state embedding vector.
[0018] In this process, the output feature map of each depth-separable convolutional residual block of the main path and the 64-dimensional channel state embedding vector corresponding to the auxiliary path are cross-fused through an adaptive instance normalization variant layer. After fusion, the features are compressed by global average pooling and then fed into a three-way parallel softmax classification head to output the classification probabilities of each item.
[0019] The channel-aware dual-path residual block partitioning model introduces a confidence-gated loop mechanism. If the maximum probability value output by any softmax classification head is lower than the confidence threshold, the current feature vector is sent to the shared parameter refining block for feature refining iteration. The confidence threshold ranges from 0.6 to 0.9.
[0020] Specifically, the speculative execution mechanism involves simultaneously speculatively initiating sub-block data prefetching and sub-block entropy value initialization operations during pixel standard deviation calculation. If the decision is to terminate the partitioning, the calculation results of the speculative initiation are discarded.
[0021] Specifically, the ping-pong buffering mechanism allocates independent buffers to two adjacent coding blocks, so that when the current coding block performs the gradient calculation step, the pixel standard deviation calculation step and the sub-block entropy value calculation step of the next coding block are executed in parallel in another independent buffer.
[0022] Specifically, the block boundary pixel copying strategy involves storing a 1-pixel-wide guard band pixel in an independent memory region for each coding block, enabling the Laplacian gradient calculation to access local independent memory throughout the process and eliminating cross-block pixel read / write contention during multi-threaded parallelism.
[0023] The Z-order curve data layout arranges the pixel data of spatially adjacent coding blocks continuously in physical memory according to the Z-order curve order. The mapping rule of the Z-order curve is to interleave the binary bits of the coordinate components to form a linear address.
[0024] The training of the channel-aware dual-path residual block partitioning model uses the cross-entropy loss function to calculate the classification loss of the three softmax classification heads respectively. The weighted sum of the three classification losses is used as the total loss. Adaptive learning rate decay is performed using a learning rate adjustment function that includes the validation set classification loss, the current training round and the current channel average signal-to-noise ratio.
[0025] This invention addresses the technical problem of decreased accuracy and reduced compression efficiency in video coding block partitioning decisions due to channel noise interference under unstable satellite communication channel conditions by introducing a Nakagami-m channel confidence weighting mechanism and a channel-aware dual-path residual block partitioning model. This is achieved by directly coupling real-time estimated channel fading parameters to the gradient calculation and partitioning decision process. The channel confidence weighting function uses the fading severity parameter as an independent variable, ensuring that the gradient contribution weight of each pixel decreases with deeper fading. The gradient mean automatically shrinks to zero during deep fading rather than fluctuating randomly, thus maintaining statistical stability of the weighted gradient difference ratio under deep fading conditions and avoiding erroneous partitioning decisions caused by random gradient fluctuations. The dual-path residual model injects channel quality parameters into the main path feature extraction process as side-channel conditions through the auxiliary path. An adaptive instance normalization variant layer corrects the feature offset introduced by channel noise in real time, ensuring that the partitioning probabilities output by the model are always aligned with the current channel conditions. In summary, this invention solves the technical problem mentioned in the background art of decreased accuracy and reduced compression efficiency in video coding block partitioning decisions due to channel noise interference under unstable satellite communication channel conditions. Attached Figure Description
[0026] Figure 1 This is a flowchart of the method of the present invention.
[0027] Figure 2 This is a pipeline timing diagram for speculative execution mechanism and ping-pong buffer mechanism.
[0028] Figure 3 This is a schematic diagram of memory mapping for Z-order curve data layout and block boundary pixel copying strategy.
[0029] Figure 4The graph shows the rate-distortion performance comparison between the method of this invention and the baseline method under different channel conditions. Detailed Implementation
[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.
[0031] like Figure 1 The diagram shows a flowchart of a lightweight video block partitioning and compression method for unstable satellite communication channels provided by this invention. The method includes the following steps:
[0032] S01. Calculate the pixel standard deviation for a 16×16 or 32×32 coding block, and compare the pixel standard deviation with a first adaptive threshold. If the pixel standard deviation is less than the first adaptive threshold, terminate the division of the coding block; otherwise, proceed to step S02.
[0033] S02. Divide the coding block into four sub-blocks, calculate the sub-block entropy value of each sub-block based on the adaptive gray level number, add the sub-block entropy values of the four horizontally arranged sub-blocks to obtain the total horizontal entropy value, add the sub-block entropy values of the four vertically arranged sub-blocks to obtain the total vertical entropy value, calculate the absolute value of the difference between the total horizontal entropy value and the total vertical entropy value to obtain the absolute value of the entropy difference, compare the absolute value of the entropy difference with the second adaptive threshold to determine the priority division direction;
[0034] S03. For each sub-block in the preferred partitioning direction, scale-adaptive Gaussian pre-smoothing is introduced before the Laplacian gradient calculation. The Laplacian response value of each pixel in each sub-block is multiplied by the channel confidence weight. The weighted pixel neighborhood difference mean of each sub-block is calculated. Then, the weighted center gradient difference and the weighted edge gradient difference are calculated. The weighted gradient difference ratio is obtained from the ratio of the weighted center gradient difference to the weighted edge gradient difference. The binary tree partitioning or ternary tree partitioning is determined based on the weighted gradient difference ratio.
[0035] S04. Input the pixel standard deviation, the sum of horizontal entropy values, the sum of vertical entropy values, the weighted gradient difference ratio, and the channel quality parameters into the channel-aware dual-path residual block partitioning model, and output the termination partitioning probability, the priority partitioning direction probability, the binary tree partitioning probability, and the ternary tree partitioning probability. Combine the speculative execution mechanism and the ping-pong buffering mechanism to eliminate pipeline bubbles.
[0036] S05. A block boundary pixel copying strategy is adopted to precopy the 1-pixel-wide guard band pixels around each coding block to an independent memory area. The pixel data in physical memory is rearranged in conjunction with the Z-order curve data layout. Rate distortion optimization calculation is performed on the partitioning modes that are not skipped, and the partitioning mode with the lowest rate distortion cost is selected for output.
[0037] S06. After completing the processing of the current coding block, process the next coding block according to the coding order, and repeat steps S01 to S05.
[0038] In step S01, the formula for calculating the pixel standard deviation is as follows:
[0039] ;
[0040] ;
[0041] in, This represents the average pixel value of the coded block. This represents the pixel standard deviation of the coded block. and These represent the width and height of the encoded block, respectively. Represents pixels The pixel value. First adaptive threshold. The calculation formula is expressed as follows:
[0042] ;
[0043] in, These are quantization parameters. The coefficient 0.6 was determined by testing multiple sets of standard test video sequences at different... The values were obtained through coding experiments. The experiments compared the coding efficiency loss and computational savings under different coefficient values, selecting the coefficient value that minimized both costs. This selection was verified through multiple iterative experiments. If... Then the division of the coded block is terminated, and the coded block is treated as a whole coded unit; if Then proceed to step S02.
[0044] In step S02, the number of adaptive gray levels From quantization parameters With block size The joint determination and segmentation rules are as follows:
[0045] when hour, ;when hour, ;when hour, ;
[0046] The segmentation thresholds 22 and 37 were determined through joint evaluation experiments on texture direction decision accuracy and computational cost in low, medium, and high bitrate encoding scenarios. The optimal segmentation point, balancing both factors, was selected after multiple iterative experiments. (This is related to adaptive grayscale levels.) Building upon this foundation, a dynamic adjustment mechanism based on histogram flatness is introduced. The uniformity of the probability of each gray level in the current sub-block's histogram is used as the adjustment criterion. If the histogram tends towards uniformity, the number of levels is appropriately reduced; if the histogram is concentrated, the number of levels is appropriately increased. (Interval width) The calculation formula is expressed as follows:
[0047] ;
[0048] in, and These are the maximum and minimum grayscale values of the current sub-block, respectively. Gray level range The definition is stated as follows:
[0049] ;
[0050] No. Gray level probability The calculation formula is expressed as follows:
[0051] ;
[0052] in, For the first Gray level range The number of pixels inside, This represents the total number of pixels in the current sub-block. The sub-block entropy value for each sub-block. The calculation formula is expressed as follows:
[0053] ;
[0054] Among them, when hour, Defined as 0. The sum of horizontal entropy values. Sum of entropy values in the vertical direction The calculation formula is expressed as follows:
[0055] ;
[0056] ;
[0057] in, For the first The entropy value of a horizontally arranged sub-block. For the first The entropy values of vertically arranged sub-blocks. The absolute value of the entropy difference. The calculation formula is expressed as follows:
[0058] ;
[0059] Second adaptive threshold The calculation formula is expressed as follows:
[0060] ;
[0061] The coefficient 0.025 was determined through iterative experiments by statistically analyzing the misclassification rate of direction decisions across various texture types of video sequences and selecting the coefficient value that minimizes the misclassification rate. and If the texture is more concentrated in the horizontal direction, the vertical division mode will be selected first, and the traversal of the horizontal division mode will be skipped; if and If the texture is more concentrated in the vertical direction, the horizontal division mode will be selected first, and the traversal of the vertical division mode will be skipped; if If the above decision result is not met, then all candidate directions for division are retained. This decision result serves as the reference input for the probability of prioritizing the division direction in step S04.
[0062] In step S03, the Gaussian kernel scale of Gaussian pre-smoothing is... The Gaussian kernel size is automatically selected based on the estimated noise variance of the current frame; the larger the estimated noise variance, the larger the Gaussian kernel size. The larger the Gaussian kernel size, the greater the size of the Gaussian kernel. The value ranges from 0.5 to 3.0. Gaussian kernel scale. The specific value is determined by comparing the gradient signal-to-noise ratio on test frames with different estimated noise variance levels, and selecting the Gaussian kernel scale that maximizes the gradient signal-to-noise ratio. After multiple experimental iterations, the estimated noise variance and Gaussian kernel scale were established. The mapping table is determined. Laplacian gradient calculation is performed on the Gaussian-smoothed sub-block pixels, using the Laplacian operator. The formal expression is as follows:
[0063] ;
[0064] In pixels A 3×3 pixel neighborhood matrix centered on The formal expression is as follows:
[0065] ;
[0066] Mean pixel neighborhood difference for each sub-block The calculation formula is expressed as follows:
[0067] ;
[0068] The fading severity parameter is obtained by real-time estimation of the Nakagami-m parameter of the current frame signal. Channel confidence weighting function The calculation formula is expressed as follows:
[0069] ;
[0070] in, As a normalized reference parameter, experimental statistics were conducted on the accuracy of classification decision under different fading conditions, and the parameter that resulted in the highest accuracy was selected. The value is determined by multiplying the Laplacian response value of each pixel within each sub-block by the channel confidence weight. Then, the mean is calculated again to obtain the weighted pixel neighborhood difference mean for each sub-block, which is denoted as follows: , , , The four sub-blocks are numbered sequentially as Block 1, Block 2, Block 3, and Block 4. Weighted central gradient difference. The calculation formula is expressed as follows:
[0071] ;
[0072] Weighted marginal gradient difference The calculation formula is expressed as follows:
[0073] ;
[0074] Weighted gradient difference ratio The calculation formula is expressed as follows:
[0075] ;
[0076] in, It is a very small positive number, used to prevent the denominator from being zero. If Then a binary tree partitioning method is preferred; if Therefore, a ternary tree partitioning method is preferred. The weighted gradient difference ratio... As a reference input for the binary tree partitioning probability and the ternary tree partitioning probability in step S04, the preferred partitioning direction is used as a reference input for the preferred partitioning direction probability in step S04.
[0077] The effect of the Nakagami-m channel confidence weighting technique is as follows: Satellite channels exhibit Nakagami-m fading characteristics under the combined effects of orbital altitude, atmospheric refraction, and the Doppler effect. When the fading is severe, the pixel gradient response is contaminated by noise. The Laplacian operator, as a second-order differential operator, will amplify the noise a second time, leading to differences in the weighted center gradient. Difference with weighted marginal gradient All are dominated by noise and lose their structural characterization significance. A parameter based on the severity of fading is introduced. Channel confidence weighting function as independent variable Subsequently, the gradient calculation results are coupled with channel quality in real time. The deeper the fading, the lower the gradient contribution weight of each pixel, causing the gradient mean to automatically shrink to zero rather than fluctuate randomly during periods of severe noise pollution, thereby improving the weighted gradient difference ratio. Maintaining statistical stability under deep fading conditions ensures the reliability of binary tree and ternary tree partitioning decisions, avoiding erroneous partitioning decisions caused by gradient random fluctuations that could lead to a decrease in compression efficiency.
[0078] The specific structure of the channel-aware dual-path residual block partitioning model is as follows: The main path adopts a depthwise separable convolutional residual chain, containing four depthwise separable convolutional residual blocks. Each depthwise separable convolutional residual block is composed of a channel-wise convolutional layer and a pointwise convolutional layer connected in series. The kernel size of the channel-wise convolutional layer is 3×3, and the kernel size of the pointwise convolutional layer is 1×1. Batch normalization layers and linear rectified activation functions are used within the block. The residual connections directly add the input to the block output. The number of output channels of the four depthwise separable convolutional residual blocks are 16, 32, 32, and 64, respectively. The auxiliary path is a three-layer fully connected network. The input is channel quality parameters, which consist of signal-to-noise ratio, bit error rate, Doppler estimate, and quantization parameters. The system consists of a 4-dimensional scalar vector. The three fully connected layers have 32, 64, and 64 neurons respectively, with a linear rectified activation function used between layers. The output is a 64-dimensional channel state embedding vector. The output feature map of each depthwise separable convolutional residual block in the main path is cross-fused with the corresponding 64-dimensional channel state embedding vector in the auxiliary path through an adaptive instance normalization variant layer. The fusion method involves using the 64-dimensional channel state embedding vector to generate scaling and bias coefficient vectors through two independent fully connected layers. An affine transformation is then performed on each channel of the main path feature map to achieve dynamic channel state-to-image feature representation. Modulation; the neuron connection weights between the main path and the auxiliary path share GPU memory allocation at the adaptive instance normalized variant layer. Intermediate activation values from each layer of the main path are stored in contiguous memory blocks according to the number of output channels. The fully connected weight matrix of the auxiliary path is stored in row-major order. During fusion, zero-copy access is performed through a unified memory pool to reduce data movement overhead between GPU memory and system memory. Interlayer data flow uses two independent CUDA streams to schedule the forward computation of the main path and the auxiliary path respectively. A synchronization barrier is set between the two CUDA streams at the adaptive instance normalized variant layer. Before the synchronization barrier, the main path's... Convolutional kernel weights are grouped and bound to corresponding CUDA thread blocks according to their output channels. Fully connected weights in auxiliary paths are batch-distributed to independent CUDA thread bundles. The affine transformations of the fusion layer after the synchronization barrier are executed serially on a single CUDA stream, eliminating write-after-read data contention during two-way parallel processing. The fused features are compressed into 128-dimensional vectors using global average pooling and fed into a three-way parallel softmax classification head. The three softmax classification heads output the termination partition probability, the priority partition direction probability, the binary tree partition probability, and the ternary tree partition probability, respectively. A confidence-gated loop mechanism is introduced. If the maximum probability value output by any softmax classifier is lower than the confidence threshold, the current 128-dimensional feature vector is fed into the shared parameter refinement block for one to two feature refinement iterations. The shared parameter refinement block structure consists of two fully connected layers plus residual connections. The upper limit of the number of iterations is determined by experimental calibration of the encoding delay constraint. If the confidence level still does not meet the confidence threshold after the iteration is completed, the current class with the highest probability is output. The confidence threshold ranges from 0.6 to 0.9, with an initial experimental step size of 0.05. The step size is then reduced to 0.01 near the optimal interval for a fine search to determine the threshold.
[0079] The steps for establishing the training dataset for the channel-aware dual-path residual block partitioning model specifically include: acquiring high-definition video sequences covering various content types; dividing each frame into 16×16 and 32×32 coded block sizes; recording the pixel data corresponding to each coded block; obtaining the optimal partitioning type for each coded block using a full search method under the reference encoder as a label, with the label including three items: whether partitioning was terminated, optimal partitioning direction, and binary tree or ternary tree partitioning type; and collecting parameters of different fading severity levels under satellite channel simulation environment. Channel quality parameters under different signal-to-noise ratios, bit error rates, and Doppler shifts are recorded and paired with corresponding coding block samples. The dataset is divided into training, validation, and test sets in a 7:1:2 ratio, and the pixel values are normalized and the channel quality parameters are standardized.
[0080] The training steps of the channel-aware dual-path residual block partitioning model specifically include: calculating the classification loss of the three softmax classification heads using the cross-entropy loss function, and weighting the three classification losses to obtain the total loss, with the weight coefficients determined experimentally; using an adaptive moment estimation optimizer to update parameters, with the initial learning rate determined through segmented experiments, and the learning rate decaying based on the changes in the validation set loss during training; allocating the parameters of the four depthwise separable convolutional residual blocks of the main path to different GPU memory partitions according to the block number, and centrally allocating the parameters of the auxiliary path to independent memory partitions, and accumulating the gradients according to the memory partitions during backpropagation and updating them uniformly to avoid gradient write conflicts; after training, determining the confidence threshold through an exhaustive search on the validation set, and selecting the confidence threshold value that minimizes the product of the overall accuracy of the three softmax classification heads and the average inference delay.
[0081] The channel-aware dual-path residual block partitioning model also includes a learning rate adjustment function, which is based on the classification loss of the current validation set. Current training round Compared with the current channel average signal-to-noise ratio Calculate the comprehensive adjustment index The formula is expressed as follows:
[0082] ;
[0083] in, The classification loss is used for the initial validation set. For reference signal-to-noise ratio, The attenuation coefficient is one of the three parameters, all of which were determined experimentally; when When, the current learning rate remains unchanged; when When the current learning rate is multiplied by a factor of 0.5, a moderate decay is applied; when When the current learning rate is multiplied by a coefficient of 0.1, it is significantly decayed; when When the learning rate is reset to the initial learning rate multiplied by a coefficient of 0.01 and the learning rate warm-up phase is restarted, the model performance is dealt with in response to the sharp decline caused by sudden changes in channel conditions.
[0084] The technical advantages of the channel-aware dual-path residual block partitioning model are as follows: Video coding partitioning decisions are essentially a classification problem of local structural features of an image. However, the time-varying fading characteristics of satellite channels introduce random noise into pixel data during transmission and acquisition, causing the accuracy of partitioning decisions based solely on image features to decrease when channel conditions change. By injecting channel quality parameters into the main path feature extraction process as side-channel conditions through the auxiliary path, and using an adaptive instance normalization variant layer, the feature map of the main path is dynamically scaled and biased according to the channel state. This is equivalent to real-time correction of feature offsets introduced by channel noise in the feature space, ensuring that the model's output termination partitioning probability, priority partitioning direction probability, binary tree partitioning probability, and ternary tree partitioning probability are always aligned with the current channel conditions, thus maintaining the stability of partitioning decisions when channel quality fluctuates. The confidence-gated loop mechanism allows the model to automatically refine when feature representation is insufficient, avoiding direct output of low-confidence decisions that could compromise compression quality.
[0085] In step S04, the speculative execution mechanism is specifically implemented as follows: during the pixel standard deviation calculation in step S01, the sub-block data prefetching and sub-block entropy initialization operations of step S02 are simultaneously speculatively initiated. If step S01 determines to terminate the partitioning, the calculation results of the speculative initiation are discarded. The ping-pong buffer mechanism allocates independent buffers to two adjacent coding blocks, so that when the current coding block executes step S03, steps S01 and S02 of the next coding block are executed in parallel in another independent buffer. The three steps are staggered and overlapped in time, eliminating pipeline bubbles caused by strict serial data dependency chains. The technical effect of the speculative execution mechanism is that the three-stage decision-making has strict serial data dependency; step S02 requires the result of step S01, and step S03 requires the result of step S02. On resource-constrained platforms, pipeline bubbles are formed, leading to a serious decrease in throughput. The speculative execution mechanism utilizes the statistical prior that step S01 does not terminate the partitioning in most cases, and starts the preparation work of step S02 in advance during the calculation of step S01, overlapping the data waiting time with the calculation time. The ping-pong buffer mechanism further arranges the three stages of adjacent coding blocks in an interleaved manner on the time axis, so that each functional unit of the processor is continuously occupied, thereby improving the actual coding throughput without increasing hardware resources and meeting the timing requirements of satellite real-time video coding.
[0086] In step S05, the block boundary pixel copying strategy stores a 1-pixel-wide guard band pixel in an independent memory region for each coding block. This ensures that the Laplacian gradient calculation accesses local independent memory throughout, eliminating cross-block pixel read / write contention during multi-threaded parallel processing. The Z-order curve data layout arranges the pixel data of spatially adjacent coding blocks continuously in physical memory according to the Z-order curve order, allowing cache line loading to simultaneously cover the pixel data of multiple spatially adjacent coding blocks, improving cache hit rate. The technical effects of the block boundary pixel copying strategy and the Z-order curve data layout are as follows: When processing adjacent coding blocks in parallel with multiple threads, the Laplacian gradient calculation needs to access pixels outside the block boundary. Concurrent read / write operations on shared boundary pixels by multiple threads lead to cache line contention, causing false sharing and increasing the cache miss frequency, resulting in parallel efficiency significantly lower than the theoretical value. The block boundary pixel copying strategy transforms cross-block dependency into local independent memory access, fundamentally eliminating false sharing. The Z-order curve data layout aggregates the pixel data of spatially adjacent coding blocks in physical memory, allowing a single cache line load to cover the pixels required by multiple coding blocks, reducing cache misses and bringing multi-core parallel efficiency closer to the theoretical peak.
[0087] The Gaussian pre-smoothing technique works as follows: The Laplacian operator, as a second-order differential operator, is highly sensitive to local abrupt changes in pixel values. Image noise is amplified quadratically after the Laplacian operation, causing the weighted pixel neighborhood difference mean to be dominated by noise and unable to reflect the true image structure. By introducing scale-adaptive Gaussian pre-smoothing, high-frequency components caused by noise are effectively suppressed before entering the Laplacian operator, while the edges of the true structure, due to their strong low- to mid-frequency components, can still generate significant gradient responses, thus reducing the gradient difference at the weighted center. Difference with weighted marginal gradient The calculation results accurately reflect the structural distribution within the coded block, ensuring the accuracy of the binary tree partitioning and ternary tree partitioning decisions in step S03.
[0088] The adaptive instance normalization variant layer is a normalization layer that performs affine transformation on the feature map based on an external conditional vector. Its working principle is as follows: first, the input feature map is instance normalized by channel; then, a scaling coefficient vector and a bias coefficient vector are generated by the 64-dimensional channel state embedding vector through two independent fully connected layers, respectively, and a channel-wise linear transformation is performed on the normalized feature map, so that the statistical properties of the feature map are controlled by the 64-dimensional channel state embedding vector. The depthwise separable convolutional residual block is a lightweight convolutional module that decomposes standard convolution into two steps: a channel-wise convolutional layer and a pointwise convolutional layer. The channel-wise convolutional layer performs spatial convolution independently on each input channel, and the pointwise convolutional layer performs channel mixing through 1×1 convolution. The total number of parameters and computational cost of the two steps is lower than that of standard convolution with the same receptive field. The residual connection directly adds the input to the block output to alleviate gradient vanishing and accelerate convergence. The Z-order curve is a space-filling curve that maps two-dimensional spatial coordinates to one-dimensional linear addresses. Its mapping rule involves interleaving the binary bits of the coordinate components to form a linear address, ensuring that adjacent points in the two-dimensional space remain close in the one-dimensional address, thereby improving cache line utilization during multi-block parallel access. The ping-pong buffering mechanism is a double-buffering technique that allocates independent buffers to two consecutively processed coding blocks, allowing the computation of one coding block to overlap with the data preparation of another, eliminating serial waiting. The confidence-gated loop is an iterative refinement mechanism based on output probability confidence. It determines whether to initiate the feature refinement iteration of the shared parameter refinement block by judging whether the maximum probability value output by the softmax classification head is lower than the confidence threshold, enabling the model to automatically enhance feature representation capabilities before outputting the decision result when feature representation is insufficient. The rate-distortion optimization calculation is an optimization process in video coding that predicts, transforms, quantizes, and entropy-encodes a given partitioning pattern, and calculates the weighted sum of the number of encoded bits and the distortion, ultimately selecting the partitioning pattern with the smallest weighted sum as the output result. The pseudo-sharing refers to the phenomenon in a multi-threaded environment where multiple threads frequently read and write to different memory addresses located in the same cache line, causing the cache line to be repeatedly invalidated and reloaded, leading to an increased cache miss frequency. The pipeline bubble refers to the phenomenon in a pipelined processor where data dependencies prevent subsequent stages from starting before the preceding stages complete, resulting in idle waiting cycles in the pipeline. The Nakagami-m parameter is a statistical parameter describing the severity of wireless channel fading, with values ranging from 0.5 to 0.5. Smaller values indicate more severe channel fading, and values approaching infinity indicate a channel approaching a fading-free additive white Gaussian noise channel. The CUDA stream is an execution sequence on an NVIDIA GPU used to organize and schedule concurrent computing tasks. Operations within the same CUDA stream are strictly executed serially, while operations between different CUDA streams are executed concurrently provided their dependencies are satisfied.
[0089] Optionally, the present invention also provides a lightweight video block partitioning and compression system for unstable satellite communication channels, implemented by a computer, wherein the computer is provided with a readable storage medium storing program instructions, which are used to execute the above-described method when the computer is run.
[0090] The specific implementation of step S01 is as follows: For an encoding block of size 16×16 or 32×32, traverse all pixel values within the block, and first calculate the average pixel value. This means summing all pixel values within a block and then dividing by the total number of pixels. ; and then Using the mean as a baseline, calculate the square of the difference between each pixel value and the mean. Then, take the square root of the average of the squared differences of all pixels to obtain the pixel standard deviation. First adaptive threshold From quantization parameters The value is determined by the product of the coefficient 0.6, and the reference value is... .Will and In comparison, if If the pixel grayscale changes within the block are gradual and the texture complexity is low, the current coding block will be treated as a whole coding unit, and subsequent partitioning will be terminated; if If the condition is met, proceed to step S02. This step utilizes the standard deviation principle in statistics to measure the dispersion of pixel distribution within a block, using this as a proxy index for texture complexity to achieve an early termination decision with extremely low computational cost.
[0091] The specific implementation of step S02 is as follows: the encoded block is evenly divided into four sub-blocks, numbered sequentially from block 1 to block 4. Adaptive grayscale level number. Based on quantization parameters Segmentation is determined, when Time to take ,when Time to take ,when Time to take Based on the segmentation, adjust the grayscale histogram of the current sub-block according to its flatness. Dynamic fine-tuning is performed: the number of levels is appropriately reduced when the histogram tends to be uniform, and the number of levels is appropriately increased when the histogram becomes concentrated. For each sub-block, the maximum gray value within the block is used as the starting point. and minimum value Difference divided by Get the interval width Divide the grayscale range into The probability is obtained by calculating the proportion of pixels in each level interval to the total number of pixels. Using the information entropy formula in information theory Calculate the entropy value of each sub-block. Add the entropy values of the four horizontally arranged sub-blocks to obtain the total horizontal entropy value. The sum of the entropy values of the four vertically arranged sub-blocks is obtained by adding the entropy values in the vertical direction. ,calculate Second adaptive threshold ,like and Then prioritize vertical partitioning and skip horizontal partitioning traversal. and Then prioritize horizontal partitioning and skip vertical partitioning traversal. Then all directional candidates are retained.
[0092] The specific implementation of step S03 is as follows: For each sub-block in the priority partitioning direction, the Gaussian kernel scale is retrieved from the pre-built mapping table based on the noise variance estimated in the current frame. , The value ranges from 0.5 to 3.0; the larger the noise variance, the lower the value. The larger the value, the more Gaussian convolution pre-smoothing is performed on sub-block pixels to suppress high-frequency noise components. After smoothing, a Laplacian operator is applied to each non-boundary pixel. With a 3×3 pixel neighborhood matrix Perform convolution, calculate the absolute value of the Laplacian response, and average the absolute values of all non-boundary pixels within the block to obtain the mean of the pixel neighborhood difference. Simultaneously, the Nakagami-m parameters are estimated in real time for the current frame signal to obtain the fading severity parameter. , by normalized reference parameters and Jointly calculate channel confidence weights ,parameter Determined through experimental calibration. The Laplacian response value of each pixel is multiplied by... Then, the mean was calculated again to obtain the weighted pixel neighborhood difference mean of the four sub-blocks. to Weighted central gradient difference Weighted marginal gradient difference Weighted gradient difference ratio ,in To prevent extremely small positive numbers with a denominator of zero. If Then prioritize binary tree partitioning, if Then, prioritize ternary tree partitioning.
[0093] The specific implementation of step S04 is: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] The sum of horizontal entropy values Sum of vertical entropy values Weighted gradient difference ratio The 4-dimensional channel quality parameters, consisting of signal-to-noise ratio, bit error rate, Doppler estimation, and quantization parameters, are input into the main and auxiliary paths of the channel-aware dual-path residual block partitioning model, respectively. The main path extracts image structural features from pixel data layer by layer through four depthwise separable convolutional residual blocks, outputting 16, 32, 32, and 64 channels respectively. The auxiliary path encodes the 4-dimensional channel quality parameters into a 64-dimensional channel state embedding vector through a three-layer fully connected network. At the output of each depthwise separable convolutional residual block, an adaptive instance normalization variant layer generates scaling and bias coefficient vectors from the 64-dimensional channel state embedding vector, performing an affine transformation on the main path feature map channel by channel. The fused features are compressed into a 128-dimensional vector by global average pooling and fed into a three-way parallel softmax classification head, outputting the termination partitioning probability, priority partitioning direction probability, binary tree partitioning probability, and ternary tree partitioning probability, respectively. If the maximum probability value output by any classification head is lower than the confidence threshold, the 128-dimensional feature vector is fed into the shared parameter refinement block for one to two feature refinement iterations before being re-output. The confidence threshold reference range is 0.6 to 0.9. The two computation paths are independently scheduled by CUDA streams, with a synchronization barrier set at the adaptive instance normalization variant layer. The speculative execution mechanism simultaneously initiates data prefetching for step S02 during the computation of step S01. The ping-pong buffering mechanism staggers the three stages of adjacent coding blocks on the time axis, eliminating pipeline bubbles.
[0094] The specific implementation of step S05 is as follows: For each coding block, a 1-pixel-wide guard band pixel is pre-copied in an independent memory region, allowing the Laplacian gradient calculation to access local independent memory throughout, eliminating concurrent read / write contention for shared boundary pixels during multi-threaded parallel processing. Pixel data in physical memory is rearranged according to the Z-order curve, storing pixel data of spatially adjacent coding blocks contiguously at physical addresses, enabling a single cache line load to cover the pixels required for multiple coding blocks, improving cache hit rate. For the partitioning patterns not skipped in steps S02 and S03, prediction, transformation, quantization, and entropy coding are performed sequentially. The number of encoded bits and the weighted sum of reconstruction distortion for each partitioning pattern, i.e., the rate-distortion cost, are calculated. The partitioning pattern with the lowest rate-distortion cost is selected as the output result of the current coding block.
[0095] It should be noted that the key technologies of this invention include: the Nakagami-m channel confidence weighting mechanism directly couples the real-time fading state of the satellite channel to the Laplacian gradient calculation, causing the gradient statistics to automatically shrink to zero during deep fading, structurally eliminating the root cause of noise-driven gradient random fluctuations and ensuring the statistical stability of the weighted gradient difference ratio; the channel-aware dual-path residual block partitioning model uses an adaptive instance normalization variant layer to embed the channel state into the vector and modulate the main path feature map in real time through affine transformation, ensuring that the partitioning decision probability is always aligned with the current channel quality; the confidence-gated loop mechanism further automatically refines the feature representation when it is insufficient, ensuring the output reliability in low-confidence scenarios; the speculative execution mechanism and ping-pong buffer mechanism transform the three-stage serial dependency chain into a time-interleaved pipeline structure; and the block boundary pixel copying strategy and Z-order curve data layout eliminate multi-threaded false sharing at the physical memory level. The three technologies mentioned above work together. The first two ensure the accuracy of partitioning decisions under unstable channel conditions from an algorithmic perspective, while the latter ensures the throughput of lightweight real-time coding from an engineering perspective. Together, they achieve a dual guarantee of compression efficiency and real-time performance under satellite channel conditions.
[0096] It's important to note that in satellite video coding systems employing multi-threaded parallel processing of adjacent coding blocks, Laplacian gradient calculation requires accessing neighboring pixels outside the current coding block boundary. These pixels outside the boundary are also internal pixels of adjacent coding blocks, continuously read and written by another thread. Multiple threads repeatedly reading and writing to different memory addresses within the same cache line leads to repeated invalidation and reloading of the cache line, creating a false sharing phenomenon. This results in actual parallel efficiency being far lower than the theoretical peak, a problem particularly pronounced under the limited processor resources of satellite payloads. The reason for this technical problem is that the processor cache synchronizes reads and writes at the cache line level. When two threads access different memory addresses but are within the same cache line, a write operation by either thread will trigger the invalidation of the other's cache line, forcing the other to reload data from memory. This process is unrelated to whether the data is actually used by the other thread; it's an inherent mechanism of the hardware consistency protocol. Common solutions to this problem include inserting padding bytes into the data structure to distribute data accessed by different threads across different cache lines, or using lock-free data structures to reduce the frequency of cache line contention. However, in video coding block partitioning scenarios, the pixel data of a coding block is determined by the content of the image frame, and its spatial layout is fixed, making it impossible to arbitrarily insert padding bytes for dispersion. Furthermore, lock-free data structures cannot eliminate structural conflicts caused by spatially adjacent pixels being allocated to the same cache line, leading to persistent false sharing during multi-core parallel coding. This invention effectively solves this technical problem. The block boundary pixel copying strategy pre-stores a 1-pixel-wide guard band pixel in an independent memory region for each coding block, ensuring that all pixels required for Laplacian gradient calculation come from the local memory region exclusively occupied by this thread. This completely cuts off cross-block pixel sharing dependencies, eliminating the conditions for false sharing at the root. The Z-order curve data layout further arranges the pixel data of spatially adjacent coding blocks consecutively in physical memory addresses, allowing a single cache line load to cover the pixel data required by multiple coding blocks. While eliminating write contention with the guard band pixel copying strategy, it also reduces the number of read misses, bringing cache utilization close to the theoretical upper limit during multi-core parallel coding and substantially improving multi-threaded coding throughput.
[0097] Specifically, the principle of this invention is:
[0098] The fundamental reason why this invention can solve the above-mentioned core technical problems is that it transforms channel state information from passively subjected to external interference into an endogenous variable that actively participates in the partitioning decision, thereby enabling dynamic coupling between block partitioning logic and channel quality.
[0099] At the gradient calculation level, the second-order differential property of the Laplacian operator makes it highly sensitive to local abrupt changes in pixels. After the Laplacian operation, the amplitude of noise is amplified quadratically, and under deep fading conditions, the gradient statistics will be completely dominated by noise. This invention introduces scale-adaptive Gaussian pre-smoothing, which automatically selects the Gaussian kernel scale based on the noise variance estimated in the current frame. This effectively suppresses high-frequency components caused by noise before entering the Laplacian operator, while the edges of the real structure, due to their strong low- to mid-frequency components, can still generate significant gradient responses, restoring the gradient statistics' ability to represent the image structure. Furthermore, the Laplacian response value of each pixel is multiplied by a channel confidence weight function with the Nakagami-m fading severity parameter as the independent variable. The deeper the fading, the lower the weight. The gradient mean automatically shrinks to zero at deep fading times, causing the numerator and denominator of the weighted gradient difference ratio to shrink synchronously under noise dominance without generating random ratio fluctuations, thus ensuring the statistical stability of the binary and ternary tree partitioning decisions.
[0100] At the partitioning decision level, the main path of the channel-aware dual-path residual block partitioning model extracts local structural features of the image through a depthwise separable convolutional residual chain, while the auxiliary path encodes channel quality parameters into channel state embedding vectors through a three-layer fully connected network. An adaptive instance normalization variant layer generates scaling and bias coefficient vectors from the channel state embedding vectors, performing an affine transformation on each channel of the main path feature map. This is equivalent to real-time correction of feature offsets introduced by channel noise in the feature space. This mechanism ensures that the model's output termination partitioning probability, priority partitioning direction probability, binary tree partitioning probability, and ternary tree partitioning probability are always aligned with the current channel conditions, maintaining the stability of partitioning decisions even when channel quality fluctuates. A confidence-gated loop mechanism automatically initiates feature refinement iterations when the classification head outputs insufficient confidence, preventing low-confidence decisions from being directly output and compressing quality.
[0101] At the engineering implementation level, the speculative execution mechanism and ping-pong buffering mechanism transform the three-stage serial dependency chain into a time-interleaved pipeline structure, eliminating pipeline bubbles. The block boundary pixel copying strategy and Z-order curve data layout fundamentally eliminate the false sharing problem in multi-threaded parallelism, bringing the multi-core parallel efficiency closer to the theoretical peak. These mechanisms together ensure that this invention achieves lightweight real-time video coding on resource-constrained satellite payload platforms, while maintaining the accuracy of partitioning decisions under unstable channel conditions.
[0102] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.
[0103] The specific implementation method of step S01 is as follows.
[0104] For size The coded block first calculates the pixel mean. The formula is expressed as follows:
[0105] ;
[0106] In the formula, For pixels The pixel grayscale value at that location, and These are the width and height of the coded block, respectively, in pixels. and These are the column index and row index of the pixel, respectively. The pixel standard deviation is then calculated based on these indices. The formula is expressed as follows:
[0107] ;
[0108] This reflects the dispersion of pixel grayscale within a coding block. A smaller value indicates a flatter texture within the block, making it more suitable as a whole coding unit without further subdivision. First adaptive threshold. From quantization parameters The linearity is determined by the following formula:
[0109] ;
[0110] In the formula, the coefficient 0.6 is determined by applying different values to multiple sets of standard test video sequences. Encoding experiments were conducted to compare the combined cost of coding efficiency loss and computational savings under different coefficient values. The final value was determined after multiple iterative experiments. If so, the division of the coding block is terminated, and it is treated as a whole coding unit; if Then proceed to step S02.
[0111] The specific implementation method of step S02 is as follows.
[0112] Adaptive grayscale level Depend on With block size The joint determination and segmentation rules are as follows: when hour, ;when hour, ;when hour, The segmentation thresholds of 22 and 37 were determined through joint evaluation experiments on texture direction decision accuracy and computational cost in low, medium, and high bitrate encoding scenarios. Based on this, a dynamic adjustment mechanism based on histogram flatness is introduced: Let the current sub-block's histogram be... The probability of grayscale level is ( The reference value for uniform distribution is If each If it tends to be uniform, reduce it appropriately. If the distribution is concentrated, then increase appropriately. Interval width The calculation formula is expressed as follows:
[0113] ;
[0114] In the formula, and These are the maximum and minimum grayscale values of the current sub-block, respectively. and , They have the same dimensions, both being units of grayscale values. Gray level range The definition is stated as follows:
[0115] ;
[0116] In the formula, This is a grayscale level index. Gray level probability The calculation formula is expressed as follows:
[0117] ;
[0118] In the formula, For interval The number of pixels inside, This represents the total number of pixels in the current sub-block. This is a dimensionless probability value. The sub-block entropy value for each sub-block. The calculation formula is expressed as follows:
[0119] ;
[0120] when When this occurs, the contribution is defined as 0. The unit is bits, and the range of values is... When the pixel grayscale distribution within a sub-block is completely uniform Take the maximum value When all pixels are concentrated at the same gray level In the formula The sub-blocks are indexed, with values ranging from 1 to 4. The sum of horizontal entropy values. Sum of entropy values in the vertical direction The calculation formula is expressed as follows:
[0121] , ;
[0122] In the formula, For the first The entropy value of a horizontally arranged sub-block. For the first The entropy value of a vertically arranged sub-block. and The units are all bits. The absolute value of the entropy difference. The calculation formula is expressed as follows:
[0123] ;
[0124] The unit is bits. Second adaptive threshold. The calculation formula is expressed as follows:
[0125] ;
[0126] In the formula, For dimensionless threshold, Here, it participates in linear operations as a dimensionless parameter. The coefficient 0.025 is determined through iterative experiments by statistically analyzing the misclassification rate of direction decisions on video sequences of various texture types, selecting the coefficient value that minimizes the misclassification rate. and If so, then the vertical partitioning mode should be selected first; if and If so, then the horizontal partitioning mode should be selected first; if If so, all candidate partitioning directions are retained.
[0127] The specific implementation method of step S03 is as follows.
[0128] Gaussian kernel scaling with pre-smooth Gaussian kernel The noise variance is automatically selected based on the current frame estimate, with a value ranging from 0.5 to 3.0. A larger estimated noise variance results in a higher noise variance. The larger the value, the more precise the mapping is determined by establishing a lookup table after comparing the gradient signal-to-noise ratio on test frames with different noise variance levels. Laplacian gradient calculation is performed on the pre-smoothed sub-block pixels, using the Laplacian operator... The matrix form is expressed as follows:
[0129] ;
[0130] In pixels Centered Pixel Neighborhood Matrix The formal expression is as follows:
[0131] ;
[0132] Mean pixel neighborhood difference for each sub-block The calculation formula is expressed as follows:
[0133] ;
[0134] In the formula, Represents the pixel neighborhood matrix With the Laplace operator The scalar Laplace response obtained by taking the absolute value of the element-wise weighted summation. and All measurements are of the same dimension, using grayscale values as the unit. Real-time estimation of the nanocargami-m parameters is performed on the current frame signal to obtain the fading severity parameter. Channel confidence weighting function The calculation formula is expressed as follows:
[0135] ;
[0136] In the formula, The fading parameter is a nano-cagami-m fading parameter, with a value range of [value range missing]. , The smaller the value, the more severe the channel fading. As the value approaches infinity, the channel approaches a fading-free additive white Gaussian noise channel. As a normalized reference parameter, experimental statistics were conducted on the accuracy of classification decision under different fading conditions, and the parameter that maximized the accuracy was selected. Value determined The range of values is , where is a dimensionless weight value. For the ... Sub-blocks ( Let its width be... Height is pixel The absolute value of the Laplace response at that point is Then the first The weighted pixel neighborhood difference mean of each sub-block The calculation formula is expressed as follows:
[0137] ;
[0138] In the formula, and The first The width and height of each sub-block, in pixels. The pixel neighborhood matrix after Gaussian pre-smoothing and the Laplacian operator The result is obtained by performing a weighted summation element by element and then taking the absolute value. and All measurements are in the same unit, grayscale value. Weighted center gradient difference. Difference with weighted marginal gradient The calculation formula is expressed as follows:
[0139] , ;
[0140] and All units are grayscale values. Weighted gradient difference ratio. The calculation formula is expressed as follows:
[0141] ;
[0142] In the formula, It is a very small positive number, and the default value is 0. This is used to prevent the denominator from being zero. This is a dimensionless ratio. If Then a binary tree partitioning method is preferred; if In this case, a ternary tree partitioning method is preferred.
[0143] The specific implementation method of step S04 is as follows.
[0144] The main path of the channel-aware dual-path residual block partitioning model uses four depthwise separable convolutional residual blocks, each consisting of a channel-wise convolutional layer (convolutional kernel). ) and pointwise convolutional layers (convolution kernels) The four blocks are cascaded, each containing a batch normalization layer and a linear rectified activation function. Residual connections directly add the input to the block output. The number of output channels for the four blocks are 16, 32, 32, and 64, respectively. The auxiliary path is a three-layer fully connected network, with inputs consisting of signal-to-noise ratio, bit error rate, Doppler estimation, and... The 4-dimensional channel quality parameter vector The number of neurons in the three layers is 32, 64, and 64 respectively, and the output is a 64-dimensional channel state embedding vector. Adaptive instance normalization variant layer on main path feature map First, perform instance normalization based on the channel, let the first... The feature map of each channel contains a total of The feature value of the i-th spatial location, the th The feature value of each position is Then the average value of the instances in that channel variance of instances The calculation formula is expressed as follows:
[0145] , ;
[0146] Normalization results The calculation formula is expressed as follows:
[0147] ;
[0148] In the formula, This is the numerical stability constant, which defaults to 0. Used to prevent division by zero. These are dimensionless normalized eigenvalues. (From...) Scaling factor vectors are generated through two independent fully connected layers. With bias coefficient vector ,in The total number of channels in the current main path feature map. The Each channel undergoes an affine transformation, as expressed by the following formula:
[0149] ;
[0150] In the formula, For the fusion of the first The first channel Dimensionless eigenvalues at each position, This represents the dimensionless scaling factor for the corresponding channel. These are the dimensionless bias coefficients for the corresponding channels. The fused features are then compressed into a 128-dimensional vector using global average pooling. Send in three parallel paths Classification header, let the logical value vector output by a certain classification header after mapping through a fully connected layer be... , The number of categories in this classification head, the th Class output probability The calculation formula is expressed as follows:
[0151] ;
[0152] In the formula, For the first Logical value of a class For the summation index, output satisfy , which is a dimensionless probability value, and the three classification heads output the termination probability respectively. Prioritize the division of directional probabilities Binary tree partitioning probability Probability of partitioning a ternary tree ,in The corresponding categories are termination of partitioning and continuation of partitioning. Corresponding to three categories: horizontal priority, vertical priority, and no obvious direction. and This corresponds to two categories: binary trees and ternary trees. In the confidence-gated loop mechanism, if the maximum probability value output by any classifier head is lower than the confidence threshold... Then the 128-dimensional feature vector The data is fed into a shared parameter refinement block for 1 to 2 iterations of refinement. The shared parameter refinement block consists of two fully connected layers plus a residual connection structure. If the confidence level is still not met after the iterations, the class with the highest probability is output. The value ranges from 0.6 to 0.9, with an initial search step size of 0.05, which is then narrowed down to 0.01 near the optimal interval for fine-tuning. Total training loss. The weighted sum of the cross-entropy losses from the three classification heads is expressed by the following formula:
[0153] ;
[0154] In the formula, , , These are the cross-entropy losses for the termination partitioning head, the priority partitioning direction partitioning head, and the binary tree and ternary tree partitioning head, respectively. , , To correspond to the dimensionless weight coefficients, all were determined through experimental adjustment. The three losses and the total loss are all dimensionless scalars. Two independent unified computing device architecture streams schedule the main path and auxiliary path forward computation respectively. A synchronization barrier is set at the adaptive instance normalization variant layer, and the affine transformation of the fusion layer is executed serially on a single unified computing device architecture stream. The learning rate adjustment function is based on the classification loss of the current validation set. Current training round Compared with the current channel average signal-to-noise ratio Calculate the comprehensive adjustment index The formula is expressed as follows:
[0155] ;
[0156] In the formula, The ratio of the current validation set loss to the initial validation set loss is a dimensionless quantity. middle The unit is the reciprocal of the round. The unit is rounds, and the product of the two is a dimensionless quantity. middle and Since they have the same unit, their ratio is dimensionless. Overall, it is a dimensionless comprehensive regulation index. The classification loss is used for the initial validation set. For reference signal-to-noise ratio, The attenuation coefficient is defined by these three parameters, all of which were determined experimentally. The learning rate remains constant; when Multiply the learning rate by 0.5; when Multiply the learning rate by 0.1; when The learning rate is then reset to the initial learning rate multiplied by 0.01, and the learning rate warm-up phase is restarted. The training dataset is divided into training, validation, and test sets in a 7:1:2 ratio. An adaptive moment estimation optimizer is used for parameter updates, with a confidence threshold... The value that minimizes the product of the overall accuracy of the three-way classification heads and the average inference delay is determined by exhaustive search on the validation set. During the calculation of step S01, the speculative execution mechanism simultaneously initiates the sub-block data prefetching and entropy initialization of step S02. If step S01 determines that the partitioning is terminated, the speculative result is discarded. The ping-pong buffer mechanism allocates independent buffers to two adjacent coding blocks, so that when the current block executes step S03, steps S01 and S02 of the next block are executed in parallel in another buffer. The three stages are staggered and overlapped in time to eliminate pipeline bubbles.
[0157] The specific implementation method of step S05 is as follows.
[0158] The block boundary pixel copying strategy pre-copys the surrounding 1-pixel-wide guard band pixels in an independent memory region for each coding block, ensuring that the Laplacian gradient calculation accesses local independent memory throughout, eliminating cross-block pixel read / write contention during multi-threaded parallelism. The Z-order curve data layout arranges the pixel data of spatially adjacent coding blocks continuously in physical memory according to the Z-order curve order. The mapping rule for the Z-order curve is to map the two-dimensional coordinates... The binary bits are interleaved to form a one-dimensional linear address, ensuring that pixels in spatially adjacent coding blocks remain close together in physical memory, thus improving cache hit rate. Rate-distortion optimization is performed on the partitioning patterns that are not skipped, and the partitioning pattern with the lowest rate-distortion cost is selected for output.
[0159] The specific implementation method of step S06 is as follows.
[0160] After the current coding block is processed, the next coding block is processed in the coding order. Steps S01 to S05 are repeated until all coding blocks of the current frame are processed and then the subsequent frames are processed until the video sequence coding is completed.
[0161] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: Technicians set up a satellite video coding test environment, selecting high-definition video sequences covering three categories: urban buildings, cloud movement, and surface texture, as test materials. The video resolution was 1920×1080, and the frame rate was 25 frames per second. The method of this invention was used to quantize the parameters of each video sequence. Coding experiments were conducted under conditions of 22, 27, 32, and 37, respectively. The channel simulation used the Nakagami-m fading model, with fading severity parameters... Varying within the range of 0.5 to 4.0, the channel signal-to-noise ratio is 10. Up to 25 Variations within the range.
[0162] In the test environment, the encoder runs on a resource-constrained embedded processor platform with 4 processor cores and limited on-chip cache capacity. Channel quality parameters are provided in real time by the channel estimation module, including the current frame signal-to-noise ratio, bit error rate estimate, Doppler shift estimate, and quantization parameters, which constitute the auxiliary path of the 4D vector real-time input channel-aware dual-path residual block partitioning model.
[0163] For the urban architectural sequence in , The signal-to-noise ratio is 15. In a typical deep fading scenario, step S01 calculates the pixel standard deviation of each 16×16 coding block. First adaptive threshold Encoding blocks with a pixel standard deviation below 19.2 are directly terminated, accounting for approximately 38% of all encoding blocks. Step S02 calculates the sub-block entropy value for the unterminated encoding blocks, and applies a second adaptive threshold. Approximately 54% of the coded blocks skipped some partitioning directions based on the absolute value of the entropy difference. In step S03, the noise variance of the current frame estimation is relatively large, and the Gaussian kernel scale... Automatically selected as 2.0, Nakagami-m parameter estimation results. Channel confidence weight The gradient statistics are reduced accordingly to keep them stable under noise interference.
[0164] The partitioning decision results of the channel-aware dual-path residual block partitioning model on the test sequence are shown in Table 1.
[0165] Table 1. Statistical table of partitioning decision accuracy for different coding block types under different channel conditions.
[0166]
[0167] As shown in Table 1, under deep fading conditions ( Signal-to-noise ratio 10 The accuracy rate of each decision remained above 86%, indicating that the synergistic effect of the channel confidence weighting mechanism and the dual-path model effectively suppressed the interference of channel noise on the partitioning decision.
[0168] In step S04, the speculative execution mechanism simultaneously prefetches sub-block data of the next coding block during the calculation in step S01. The ping-pong buffering mechanism causes the three processing stages of adjacent coding blocks to be staggered on the time axis, such as... Figure 2 As shown, the time overlap of each processing stage keeps each functional unit of the processor continuously occupied, significantly reducing the pipeline bubble ratio.
[0169] In step S05, the block boundary pixel copying strategy allocates a 1-pixel-wide guard band memory independently for each coding block. The Z-order curve data layout arranges the pixel data of spatially adjacent coding blocks consecutively in physical address, such as... Figure 3 As shown, when loading cache lines, pixel data from multiple encoding blocks can be covered simultaneously, significantly reducing the number of cache misses during multi-threaded parallel processing.
[0170] In the comprehensive test of three types of test sequences, rate-distortion optimization computation sequentially performs prediction, transformation, quantization, and entropy coding on the partition patterns that are not skipped, and selects the partition pattern with the lowest rate-distortion cost as the output. Each sequence is tested under different conditions. Rate-distortion performance curves under channel conditions are as follows Figure 4 As shown, under deep fading conditions, the rate-distortion curve of the method of the present invention maintains a better position than the baseline method without introducing a channel-aware mechanism, indicating that the improvement in the accuracy of partitioning decision directly leads to the improvement in compression efficiency.
[0171] The advancements of this invention over traditional methods are reflected in the following principles. Traditional block partitioning methods rely solely on gradient and entropy statistics in the image domain for partitioning decisions. While this approach yields good results under terrestrial channel conditions, under deep fading conditions in satellite channels, the second-order differential property of the Laplacian operator amplifies noise quadratically, causing gradient statistics to be completely dominated by noise and lose their structural representational significance. The partitioning decision then degenerates into a random decision. This invention, by coupling Nakagami-m fading parameters to gradient weights in real time, makes channel quality an endogenous modulation variable in gradient calculation, thus fundamentally blocking the path of random noise fluctuations to partitioning decisions. The dual-path residual model further eliminates feature distribution drift introduced by channel noise through affine correction of the feature space, ensuring statistical consistency in partitioning decisions even with large fluctuations in channel quality. At the engineering level, the block boundary pixel replication strategy transforms cross-block pixel dependencies into local independent memory access, eliminating the parallel efficiency bottleneck of pseudo-sharing from a physical memory structure perspective. The Z-order curve layout preserves spatial locality in the physical address dimension. Both factors combined bring the multi-core parallel efficiency close to the theoretical peak, satisfying the temporal constraints of satellite real-time video coding.
[0172] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.
[0173] Table 2. Variable Explanation Table (Part 1)
[0174]
[0175] Table 3. Variable Explanation Table (Part Two)
[0176]
[0177] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A lightweight video block partitioning and compression method for unstable satellite communication channels, characterized in that, Includes the following steps: Calculate the pixel standard deviation for the coding block, compare the pixel standard deviation with the first adaptive threshold, and if the pixel standard deviation is less than the first adaptive threshold, terminate the division of the coding block; otherwise, proceed to the next step. The coded block is divided into four sub-blocks. The sub-block entropy value of each sub-block is calculated based on the number of adaptive gray levels. The sum of horizontal entropy values and the sum of vertical entropy values are calculated. The absolute value of the entropy difference is obtained from the absolute value of the difference between the two. The absolute value of the entropy difference is compared with the second adaptive threshold to determine the priority division direction. For each sub-block in the priority partitioning direction, scale-adaptive Gaussian pre-smoothing is introduced before the Laplacian gradient calculation. The Laplacian response value of each pixel in each sub-block is multiplied by the channel confidence weight, and the weighted pixel neighborhood difference mean of each sub-block is calculated. Then, the weighted center gradient difference and the weighted edge gradient difference are calculated. The weighted gradient difference ratio is obtained from the ratio of the two. The binary tree partitioning or ternary tree partitioning is determined according to the weighted gradient difference ratio. The pixel standard deviation, the sum of horizontal entropy values, the sum of vertical entropy values, the weighted gradient difference ratio, and the channel quality parameters are input into the channel-aware dual-path residual block partitioning model. The model outputs the termination partitioning probability, the priority partitioning direction probability, the binary tree partitioning probability, and the ternary tree partitioning probability. The model also combines the speculative execution mechanism and the ping-pong buffering mechanism to eliminate pipeline bubbles. A block boundary pixel copying strategy is adopted to pre-copy the protection band pixels of each coding block to an independent memory area. The pixel data in physical memory is rearranged in conjunction with the Z-order curve data layout. Rate distortion optimization calculation is performed on the partitioning modes that are not skipped, and the partitioning mode with the minimum rate distortion cost is selected for output. After completing the processing of the current coding block, process the next coding block in the coding order, and repeat the above steps.
2. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 1, characterized in that, The pixel standard deviation is calculated by taking the square root of the square of the difference between each pixel value and the mean, based on the mean of all pixel values within the coding block.
3. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 2, characterized in that, The first adaptive threshold is determined by the product of the quantization parameter and the first coefficient, with a reference value of 0.6 for the first coefficient.
4. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 3, characterized in that, The number of adaptive gray levels is determined by the quantization parameters and the block size, and a dynamic adjustment mechanism based on histogram flatness is introduced to dynamically correct the number of gray levels according to the uniformity of the probability of each gray level in the current sub-block histogram.
5. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 4, characterized in that, The second adaptive threshold is determined by the product of the quantization parameter and the second coefficient, with a reference value of 0.025 for the second coefficient. If the absolute value of the entropy difference is greater than the second adaptive threshold and the sum of the horizontal entropy values is less than the sum of the vertical entropy values, then the vertical division mode is selected first. If the absolute value of the entropy difference is greater than the second adaptive threshold and the sum of the horizontal entropy values is greater than the sum of the vertical entropy values, then the horizontal division mode is selected first. If the absolute value of the entropy difference is not greater than the second adaptive threshold, then all candidates for division directions are retained.
6. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 5, characterized in that, The Gaussian kernel scale of the adaptive Gaussian pre-smoothing is automatically selected based on the estimated noise variance of the current frame. The larger the estimated noise variance, the larger the Gaussian kernel scale. The value range of the Gaussian kernel scale is 0.5 to 3.
0.
7. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 6, characterized in that, The channel confidence weight is determined by the fading severity parameter obtained by real-time estimation of the Nakagami-m parameter of the current frame signal. The smaller the fading severity parameter, the lower the channel confidence weight.
8. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 7, characterized in that, The main path of the channel-aware dual-path residual block partitioning model adopts a depthwise separable convolutional residual chain, which contains 4 depthwise separable convolutional residual blocks; the auxiliary path is a three-layer fully connected network, with the input being a 4-dimensional channel quality parameter composed of signal-to-noise ratio, bit error rate, Doppler estimation value and quantization parameter, and the output being a 64-dimensional channel state embedding vector.
9. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 8, characterized in that, The output feature map of each depth-separable convolutional residual block of the main path and the 64-dimensional channel state embedding vector corresponding to the auxiliary path are cross-fused through an adaptive instance normalization variant layer. After fusion, the features are compressed by global average pooling and then fed into a three-way parallel softmax classification head to output the classification probabilities of each item.
10. The lightweight video block partitioning and compression method for unstable satellite communication channels according to claim 9, characterized in that, The channel-aware dual-path residual block partitioning model introduces a confidence-gated loop mechanism. If the maximum probability value output by any softmax classification head is lower than the confidence threshold, the current feature vector is sent to the shared parameter refinement block for feature refinement iteration. The confidence threshold ranges from 0.6 to 0.9.
Citation Information
Patent Citations
VVC fast coding method based on decision tree
CN116320436A
VVC coding unit accelerated division method based on feature difference and storage medium
CN117135350A
