Video Coding Method Based on Multi-Domain Perceptual Feature Fusion

Abstract

The present invention discloses a video encoding method based on multi-domain perceptual feature fusion, which belongs to the field of image communication technology. The method comprises the following steps: extracting perceptual features in the spatial domain, time domain, and frequency domain from an input video frame sequence; normalizing the multi-domain perceptual features, and generating a perceptual importance factor for each coding unit through a dynamic fusion weight mechanism; constructing a perceptual distortion evaluation based on the perceptual importance factor, replacing the traditional rate-distortion objective function, and guiding the coding unit bit rate allocation; based on the perceptual importance factor, performing nonlinear mapping using a hyperbolic tangent function, calculating the coding unit-level quantization parameter adjustment amount, and combining the basic quantization parameter to obtain the final quantization parameter, thereby achieving adaptive rate-distortion optimization. This method can accurately identify visually sensitive areas and significantly reduce the bit rate while ensuring subjective visual quality.

Description

Technical Field

[0001] The present invention belongs to the field of image communication technology, and specifically relates to a video encoding method based on multi-domain perception feature fusion. Background Art

[0002] With the innovation of video technology and the diversification of its application scenarios, the amount of video data is increasing exponentially, posing severe challenges to data storage and communication. While standard video coding methods have achieved theoretical coding efficiency by removing statistical redundancy, redundancy still exists at the level of visual perception. The essence of perceptual coding is to rationally allocate bitrate based on the characteristics of human vision. However, existing perceptual video coding methods often rely on a single visual feature or simply superimpose multiple features, failing to construct distortion assessment models that are consistent with the characteristics of the human eye. This results in a mismatch between bitrate allocation and the distribution of visual importance.

[0003] The human eye has significant differences in sensitivity to visual signals of different frequencies and is more sensitive to mid-frequency details. However, existing methods do not fully utilize this characteristic, resulting in low coding efficiency of key visual information. Although some studies have attempted to integrate multi-domain features, they have limitations such as a single feature extraction dimension, static weight design, and rate-distortion optimization that does not incorporate perceptual distortion. For example, the video coding standard method uses pixel-level distortion indicators and does not consider the differences in human eye perception of different regions, resulting in insufficient bitrate allocation in key areas and redundant coding of non-key areas. Therefore, there is an urgent need for a coding method that dynamically integrates multi-dimensional perceptual features in the spatial, temporal, and frequency domains, adaptively adjusts weights, and optimizes rate-distortion balance based on perceptual distortion to improve coding efficiency and achieve a precise match between bitrate allocation and visual importance. Summary of the Invention

[0004] This invention provides a novel solution to the shortcomings of existing video coding standards, including multi-domain feature fusion, static weight design, and linearization of rate-distortion optimization. By dynamically integrating perceptual features in spatial, temporal, and frequency domains, it achieves more accurate representation of perceptual importance factors and bitrate allocation, effectively improving coding efficiency.

[0005] The present invention provides a video encoding method based on multi-domain perceptual feature fusion, the method comprising the following steps:

[0006] S1: Extract multi-domain perceptual features from the input video frame sequence. The multi-domain perceptual features include spatial domain perceptual features, temporal domain perceptual features, and frequency domain perceptual features. The specific extraction method is as follows:

[0007] Spatial domain perceptual feature extraction: For each coding unit, the edge strength, contrast, and spatial masking features of its covered pixels are calculated and fused to generate a spatial domain comprehensive perceptual feature;

[0008] Temporal domain perceptual feature extraction: For each coding unit, based on its covered pixels, the brightness change features reflecting the significance of inter-frame motion are extracted as the temporal domain perceptual features of the coding unit.

[0009] Frequency domain perceptual feature extraction: For each coding unit, Fourier transform is performed on its covered pixels, and the intermediate frequency components that are sensitive to the human eye are extracted through bandpass filtering. The mean value of the area corresponding to this component is calculated as the frequency domain perceptual feature of the coding unit.

[0010] S2: Normalize the multi-domain perceptual features extracted in S1 and design a fusion weight mechanism based on dynamic adaptive adjustment of feature values. The perceptual importance factor of each coding unit is generated through weighted aggregation. Specifically, it includes:

[0011] Feature normalization: Use the Min-Max method to map all features to the [0,1] interval to eliminate the impact of dimensional differences on feature representation;

[0012] Dynamic fusion weights: Calculate weights that are adaptively adjusted based on eigenvalues, enhance the influence of key perceptual features, and dynamically adjust weights using the Sigmoid function.

[0013] Perceptual importance factor calculation: The normalized multi-domain perceptual features are weighted and summed with the dynamic weights to obtain the perceptual importance factor that represents the visual importance of the coding unit.

[0014] S3: Constructing a perceptual distortion evaluation value D based on the perceptual importance factor p , and the rate-distortion objective function in the video coding standard method Replace with , used to guide the code rate allocation of the coding unit, including the bit resource allocation of the coding unit and the iterative optimization of the quantization parameter; where D is the distortion term in the video coding standard method, is the Lagrange multiplier, and R is the bit rate used for encoding.

[0015] S4: Based on the perceptual importance factor calculated in S2, the quantization parameter adjustment amount of each coding unit is calculated through a nonlinear mapping function; the basic quantization parameter value preset in the video coding standard method is added to the quantization parameter adjustment amount to obtain the final quantization parameter value of the coding unit, which is used for video coding to achieve adaptive rate-distortion optimization; the nonlinear mapping uses a hyperbolic tangent function to achieve nonlinear transformation of the quantization parameter adjustment amount.

[0016] Preferably, in the spatial domain perceptual feature extraction:

[0017] For each pixel point (x, y) in the coding unit CU, the horizontal gradient G is calculated by the Sobel operator x(x,y) and vertical gradient G y (x,y):

[0018] ,

[0019] ,

[0020] Where I(x,y) represents the 3×3 neighborhood pixel matrix centered at (x,y), * represents the convolution operation; the pixel-level gradient amplitude M(x,y) is calculated as:

[0021] ,

[0022] CU-level edge strength feature E CU is the normalized value of the gradient amplitude of all valid pixels in the CU:

[0023] ,

[0024] Where N is the number of valid pixels in the CU.

[0025] For each pixel (x, y) in the CU, calculate the mean brightness value within its 3×3 neighborhood W(x, y) and standard deviation :

[0026] ,

[0027] ,

[0028] CU-level contrast characteristic C CU The standard deviation normalized value of all pixels in the CU:

[0029] .

[0030] For each non-boundary pixel (x, y) in the CU, the predicted value Ȋ(x, y) is calculated based on its four neighboring pixels:

[0031] ,

[0032] Prediction error The calculation formula is:

[0033] .

[0034] CU-level spatial masking feature M CU The normalized value of the prediction error of all valid pixels in the CU:

[0035] ,

[0036] Among them CUvalid Represents the set of non-boundary pixels in the CU.

[0037] Preferably, in the temporal domain perceptual feature extraction, for the coding units of two adjacent frames in the video sequence, the current frame CU t and the previous frame CU t-1 , calculate the time domain difference map D(x,y) of the corresponding position pixel:

[0038] ,

[0039] in and Represents the pixel brightness value at position (x, y) in the current frame and the previous frame respectively;

[0040] CU-level temporal domain differential feature T CU The time domain differential normalized value of all pixels in the CU:

[0041] ,

[0042] Where N is the total number of pixels in the CU.

[0043] Preferably, in the frequency domain perceptual feature extraction, a two-dimensional discrete Fourier transform (DFT) is first performed on the input current frame image I(x, y) to convert the spatial domain information into a frequency domain representation, and obtain its spectrum F(u, v):

[0044] ,

[0045] Among them, I(x, y) is the original spatial domain image with a size of H×W; F(x, y) is the frequency domain complex matrix obtained after transformation, Table Fourier transform operation.

[0046] Construct an inner radius R inner , the outer radius is R outer The annular bandpass filter M(u,v) is used to retain the intermediate frequency components of the image, and its mathematical definition is:

[0047] ,

[0048] in Represents a spectrum point distance to the center of the spectrum; is the center coordinate of the spectrum graph; R inner and R outer are the inner and outer radii of the bandpass filter, which are 10 and 60 respectively.

[0049] Multiply the designed filter M(u, v) by the frequency domain data F(u, v) point by point, retaining only the intermediate frequency components that fall within the annular band, and obtain the filtered frequency domain result :

[0050] .

[0051] The filtered frequency domain data Perform two-dimensional inverse Fourier transform IDFT to convert the intermediate frequency component from the frequency domain back to the spatial domain to obtain the intermediate frequency feature map :

[0052] ,

[0053] in This step represents the inverse Fourier transform operation, which can observe and process the intermediate frequency information in the spatial domain.

[0054] Calculate the average intermediate frequency superposition value of all pixels in the CU:

[0055] .

[0056] Preferably, in the feature normalization, the normalization formula is:

[0057] ,

[0058] in, , min(X) and max(X) are the minimum and maximum feature values ​​of all CUs in the current frame, respectively.

[0059] Preferably, in the spatial domain perceptual feature fusion, the spatial domain comprehensive perceptual feature is a fusion of the edge strength, contrast and spatial masking features of each coding unit with a weight of 0.4:0.3:0.3.

[0060] Preferably, in the dynamic weight setting:

[0061] The spatial domain weight calculation formula is: ,

[0062] The time domain weight calculation formula is: ,

[0063] The frequency domain weight calculation formula is: ,

[0064] Among them, the default , , .

[0065] Preferably, the perceptual importance factor of each coding unit is Indicator is being constructed:

[0066] .

[0067] Preferably, in the perceptual distortion assessment:

[0068] The weighting coefficient , is suitable for most video scenes, and the default value is 0.5; the perceptual distortion evaluation value D p Ratio of the distortion term D to the video coding standard method , which reflects the modulation effect of visual importance on distortion evaluation, for example, when When the distortion weight can be increased to the maximum value of the video coding standard method times.

[0069] Preferably, the quantization parameter adjustment amount Calculating:

[0070] ,

[0071] The default parameters are A=8, B=2, C=-1, and D=2. Clip is a clipping function that limits the calculation results to the interval [-5, 5].

[0072] Final quantization parameters The values ​​are:

[0073] ,

[0074] in is the base quantization parameter of the encoder, and .

[0075] Compared with the prior art, the present invention has at least the following beneficial effects:

[0076] 1. This invention dynamically integrates multi-domain perceptual features from the spatial, temporal, and frequency domains to construct an importance assessment model that aligns with the human visual characteristics. Unlike existing methods that rely on a single feature or static feature superposition, this model enables accurate identification of visually sensitive areas across different scenes. This model further constructs a perceptual distortion assessment mechanism based on perceptual importance factors. Through adaptive adjustment of quantization parameters, it shifts the rate-distortion optimization objective from pixel error to visual perceptual error. This significantly improves coding efficiency compared to standard video coding methods while maintaining the same subjective quality.

[0077] 2. During the feature extraction phase, this invention innovatively incorporates frequency-domain intermediate-frequency features, which, together with spatial and temporal perceptual features, form a multi-domain perceptual feature fusion system. This fusion approach fully captures the complexity of video content. Compared to methods that rely solely on a single or small number of features, this multi-domain perceptual feature fusion mechanism more accurately assesses the visual importance of video content. This allows for more bit resources to be allocated to critical areas during encoding, while reducing redundant encoding in non-critical areas, achieving superior rate-distortion performance.

[0078] 3. The perceptual distortion assessment model constructed by this invention integrates perceptual importance factors with the distortion terms of standard methods to form an assessment mechanism that conforms to the visual characteristics of the human eye. This significantly improves the weight of distortion assessment in key areas, optimizing bitrate allocation and reducing overall bitrate. Substituting this into the rate-distortion objective function, the coding optimization goal is upgraded to minimizing visual perceptual error. Compared with the VVC coding method, the proposed method achieves significant bitrate savings at the same subjective quality, with a more pronounced advantage in dynamic scenes.

[0079] 4. This paper proposes a quantization parameter adjustment method based on perceptual importance factors. This method uses nonlinear mapping to accurately calculate the quantization parameter value for each coding unit. Compared with fixed quantization parameter values ​​or simple linear adjustment methods used in video coding standards, this method allows for more precise control of bit allocation during the encoding process, thereby further reducing the encoding bit rate while maintaining video quality.

[0080] 5. This invention was validated on standard video coding test sequences Class B through Class E. The experimental results show that compared to the VVC standard coding method, this invention can achieve an average bitrate reduction of 19.2% while maintaining video quality. This demonstrates that this invention strikes a good balance between coding efficiency and video quality, and has high practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0082] Figure 1 This is the core flow chart of the video encoding method based on multi-domain perception feature fusion of the present invention.

[0083] Figure 2 This is a diagram of the architecture of the multi-domain perception feature extraction and perception importance factor calculation module of the present invention.

[0084] Figure 3 A flow chart is provided for constructing a perceptual rate-distortion optimization model based on perceptual importance factors according to the present invention.

[0085] Figure 4This is a schematic diagram of the principle of the quantization parameter adjustment algorithm based on the perceptual importance factor of the present invention. DETAILED DESCRIPTION

[0086] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several variations and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0087] like Figure 1 The figure shows the core flow chart of the video encoding method based on multi-domain perception feature fusion of the present invention. This embodiment provides a video encoding method based on multi-domain perception feature fusion, including the following steps:

[0088] S1: extracting multi-domain perceptual features from an input video frame sequence, where the multi-domain perceptual features include spatial domain perceptual features, temporal domain perceptual features, and frequency domain perceptual features.

[0089] S2: Normalize the multi-domain perceptual features extracted by S1 and design a fusion weight mechanism based on dynamic adaptive adjustment of feature values ​​to generate the perceptual importance factor of each coding unit through weighted aggregation .

[0090] S3: Constructing a perceptual distortion evaluation value based on the perceptual importance factor , and the rate-distortion objective function of the video coding standard method Replace with , which is used to guide the code rate allocation of the coding unit, including the bit resource allocation of the coding unit and the iterative optimization of the quantization parameters. Where D is the distortion term in the video coding standard method, is the Lagrange multiplier, R is the code rate used for encoding;

[0091] S4: Based on the perceptual importance factor calculated in S2, a quantization parameter adjustment amount for each coding unit is calculated through a nonlinear mapping function; a base quantization parameter value is added to the quantization parameter adjustment amount to obtain a final quantization parameter value for the coding unit, which is used for video encoding to achieve adaptive rate-distortion optimization.

[0092] Example 1: Multi-domain perception feature extraction

[0093] The present invention provides a preferred embodiment for executing S1, processing videos of any length and type, such as surveillance, movies, and animation, through a multi-domain perceptual feature extraction mechanism. This mechanism breaks through the limitations of single feature extraction in standard video coding methods and achieves the coordinated extraction of perceptual features in the spatial, temporal, and frequency domains, providing data support for subsequent bitrate allocation. The specific steps are as follows:

[0094] S11. Spatial Domain Aware Feature Extraction

[0095] For each coding unit CU of the input video frame, for pixel points , use the Sobel operator to calculate the horizontal gradient G x (x,y) and vertical gradient G y (x,y). The horizontal template of the Sobel operator is , the vertical template is , through The 3×3 neighborhood pixel matrix centered at Convolution obtains the gradient value, and then calculates the pixel-level gradient amplitude , after normalization, we get the CU-level edge strength feature E CU .

[0096] At the same time, calculate the pixel The average brightness in a 3×3 neighborhood and standard deviation , after normalization, we get the CU-level contrast feature C CU For non-boundary pixels within the CU, the predicted value is calculated based on the four neighboring pixels. , and get the prediction error , after normalization, we get the CU-level spatial masking feature M CU Finally, the normalized edge intensity, contrast and spatial masking features are fused with a weight of 0.4:0.3:0.3 to obtain the spatial domain comprehensive perception feature S CU .

[0097] S12. Time-Domain Aware Feature Extraction

[0098] For the coding unit of two adjacent frames in a video sequence, the current frame and the previous frame , calculate the time domain difference map of the corresponding position pixel , after normalization, we get the CU-level time domain differential feature T CU Note that the first frame has no previous frame, so the temporal perception feature is set to 0.

[0099] S13. Frequency Domain Perceptual Feature Extraction

[0100] For the input image Perform a two-dimensional discrete Fourier transform to obtain the frequency domain matrix , construct an annular bandpass filter with an inner radius of 10 and an outer radius of 60 ,Will and Multiply and extract the intermediate frequency components , the intermediate frequency feature map is obtained by two-dimensional inverse Fourier transform Finally, for each coding unit CU, calculate the coverage area The pixel mean As the frequency domain perceptual features of the coding unit CU.

[0101] Example 2: Calculation of Perceived Importance Factor

[0102] The present invention provides a preferred embodiment to perform S2, which converts the normalized multi-domain perception features into a quantitative index of perception importance factor through dynamic weight fusion and nonlinear mapping. This embodiment supports adaptive adjustment of the influence weights of perceptual features in various domains under different scenarios. For example, in motion scenes, the weights of temporal perceptual features are automatically increased to ensure that coding resources are tilted toward visually sensitive areas.

[0103] Figure 2 This is a module architecture diagram of this embodiment, showing the complete link of "feature normalization → dynamic weight calculation → perception importance factor generation". The specific steps are as follows:

[0104] S21. Feature Normalization

[0105] The Min-Max method is used to map spatial domain perception features, including edge strength, contrast, spatial masking features, time domain perception features and frequency domain perception features to The normalized formula is , where X includes the spatial, temporal, and frequency domain perceptual features of the coding unit, and X' is the normalized feature value.

[0106] S22. Dynamic Fusion Weight Design

[0107] The weights that are adaptively adjusted with the eigenvalues ​​are designed through the Sigmoid function:

[0108] Spatial domain weight: ,

[0109] Time domain weight: ,

[0110] Frequency domain weight: ,

[0111] Among them, the default , , .

[0112] S23. Calculation of Perceived Importance Factor

[0113] The normalized multi-domain perception features are integrated and weighted summed according to dynamic weights to obtain the perception importance factor. , the calculation formula is , characterizes the visual importance of CU.

[0114] Example 3: Perceptual Distortion Evaluation and Rate-Distortion Objective Function Optimization

[0115] This invention provides a preferred embodiment for S3, constructing a perceptual distortion assessment mechanism based on perceptual importance factors and reshaping the rate-distortion objective function. By correcting the pixel-level distortion of standard video coding methods to a distortion metric that aligns with human perception, and dynamically optimizing coding decisions based on bitrate constraints, coding resources are precisely matched to the needs of visually sensitive areas, addressing the irrational bitrate allocation problem caused by video coding standards ignoring perceptual characteristics.

[0116] Figure 3 This is a flowchart for constructing a perceptual rate-distortion optimization model based on perceptual importance factors in this embodiment. It presents the core logic of "perceptual distortion correction → rate-distortion objective function reconstruction." The specific steps are as follows:

[0117] S31. Video Coding Standard Method Distortion Term Calculation:

[0118] For a certain CU, the mean square error (MSE) between the reconstructed pixels and the original pixels is calculated as:

[0119] ,

[0120] Among them, N is the number of CU pixels, and are the original pixel values ​​and the reconstructed pixel values ​​respectively.

[0121] If the absolute error MAE is calculated, the formula is:

[0122] .

[0123] S32. Generating a perceptual distortion evaluation value:

[0124] Assuming that the calculated value of the CU's perceived importance factor is 0.7, take , then the perceived distortion evaluation value is: This result shows that regions with high visual importance such as The distortion is amplified by 35%, prompting the encoder to allocate more bit rate.

[0125] S33. Rate-distortion objective function replacement:

[0126] The rate-distortion objective function of the video coding standard method Replace with For example, if , , bit rate R=100, then the standard objective function value is 8+16×100=1608, and the optimized objective function value is 10.4+16×100=1610.4, indicating that the encoder will give priority to reducing the distortion of this CU.

[0127] Example 4: Adaptive Rate-Distortion Coding Optimization

[0128] The present invention provides a preferred embodiment for performing S4, which achieves adaptive adjustment of the quantization parameter value through nonlinear mapping of the hyperbolic tangent function. This mechanism overcomes the linear quantization limitations of standard video coding methods and dynamically fine-tunes the quantization parameter of the CU based on the value of the perceptual importance factor. In texture-rich scenes such as Park Scene, the quantization parameter is automatically reduced to preserve detail, while in flat areas, the quantization parameter is increased to reduce redundant coding.

[0129] Figure 4 The algorithm principle of this embodiment is demonstrated: based on the basic quantization parameter, the quantization parameter adjustment amount is calculated through the formula to finally generate the optimal quantization parameter value. The specific steps are as follows:

[0130] S41. Quantization Parameter Adjustment Calculation

[0131] The perceptual importance factor is nonlinearly mapped to the quantization parameter adjustment through the hyperbolic tangent function to avoid the limitations of linear mapping. The calculation formula is:

[0132] ,

[0133] The default parameters are A=8, B=2, C=-1, D=2, and clip is a clipping function used to limit the calculation results to the interval [-5,5]. is the perceptual importance factor at the coding unit level.

[0134] S42. Final quantization parameter determination

[0135] The final quantization parameter value is obtained by adding the basic quantization parameter and the quantization parameter adjustment amount to achieve adaptive rate-distortion optimization of video coding. The calculation formula is: , and must meet .

[0136] To validate the effectiveness of the proposed video coding method based on multi-domain perceptual feature fusion, a comparative experiment was conducted using a standard test video sequence (Class BE), using the VVC coding method as a baseline. The experiment covered scenes of varying resolutions, including Class B (1920×1080), Class C (834×420), Class D (416×240), and Class E (1280×720), encompassing typical video types, including dynamic scenes such as BasketballDrive, texture-rich scenes such as Cactus, and static scenes such as BQSquare.

[0137] This experiment uses common industry quantitative indicators for comparative analysis: BD-MOS is used to characterize the improvement in subjective visual quality of the method of the present invention compared to the VVC encoding method, with larger values ​​indicating a more significant quality advantage; BD-Rate (%) quantifies the bitrate savings, with negative values ​​indicating a bitrate reduction compared to the VVC encoding method.

[0138] In the experimental process, we first extract and fuse multi-domain perceptual features from the spatial, temporal, and frequency domains for each coding unit (CU). We then dynamically adjust the quantization parameters using perceptual importance factors. We also embed a perceptual distortion assessment model into the rate-distortion optimization objective function to achieve adaptive and precise bitrate allocation. The specific experimental results are shown in Table 1.

[0139] Table 1 Experimental results of the method of the present invention

[0140]

[0141] As shown in Table 1, the proposed method achieves an average bitrate reduction of 19.2% compared to the VVC video coding method while maintaining subjective visual quality. BQTarace achieves a bitrate reduction of 32.0% for dynamic scenes, and Cactus achieves a 19.8% reduction for texture-rich scenes, demonstrating the method's universality and advancement across diverse scenarios. This achievement is primarily due to the following innovative designs:

[0142] First, a dynamic fusion mechanism for multi-domain perceptual features is proposed, breaking through the limitations of standard video coding methods that rely on a single feature or simply superimpose it. By collaboratively extracting perceptual features from the spatial, temporal, and frequency domains, it accurately captures the visual importance of video content and provides a multi-dimensional basis for bitrate allocation.

[0143] Secondly, an adaptive weight and perceptual distortion evaluation model is constructed. Based on the Sigmoid function, a weight mechanism that is dynamically adjusted with the eigenvalue is designed, and combined with the perceptual distortion formula , upgrading pixel-level distortion evaluation to perceptual distortion evaluation that conforms to the visual characteristics of the human eye, and achieving accurate bitrate allocation in key areas;

[0144] Finally, a nonlinear quantization parameter adaptive adjustment algorithm is designed. The perceptual importance factor is mapped to the quantization parameter adjustment amount through the hyperbolic tangent function, realizing the adaptive quantization strategy of "low quantization parameters in visually important areas to retain details, and high quantization parameters in flat areas to reduce redundancy".

[0145] The synergistic effect of the above-mentioned innovative modules breaks through the limitations of insufficient feature fusion, static weights, and rate-distortion optimization that are separated from perception in standard video coding methods, realizes the deep coupling of visual perception and rate-distortion optimization, and provides a new technical path for improving video coding efficiency.

Claims

1. A video coding method based on multi-domain perceptual feature fusion, characterized in that: The following steps are involved: S1: Extract multi-domain perceptual features from the input video frame sequence. The multi-domain perceptual features include spatial domain perceptual features, temporal domain perceptual features, and frequency domain perceptual features. The specific extraction method is as follows: Spatial domain perception feature extraction: For each coding unit, the edge strength, contrast, and spatial masking characteristics of its covered pixels are calculated and weighted and fused into a spatial domain comprehensive perception feature; Temporal perceptual feature extraction: For each coding unit, based on its covered pixels, the brightness change features reflecting the significance of inter-frame motion are extracted as the temporal perceptual features of the coding unit; Frequency domain perceptual feature extraction: For each coding unit, Fourier transform is performed on its covered pixels, and the intermediate frequency components that are sensitive to the human eye are extracted through bandpass filtering. The mean value of the corresponding area of ​​the intermediate frequency components is calculated as the frequency domain perceptual feature of the coding unit. S2: Normalize the multi-domain perceptual features extracted in S1 and design a fusion weight mechanism based on dynamic adaptive adjustment of feature values. The perceptual importance factor of each coding unit is generated through weighted aggregation. Specifically, it includes: Feature normalization: Use the Min-Max method to map all features to the [0,1] interval to eliminate the impact of dimensional differences on feature representation; Dynamic fusion weights: Calculate weights that are adaptively adjusted based on spatial, temporal, and frequency domain perception features, strengthen the influence of key perception features, and dynamically adjust weights using a Sigmoid function. The dynamic fusion weight is adaptively adjusted by the Sigmoid function. The calculation formula includes: The spatial domain weight calculation formula is: , The time domain weight calculation formula is: , The frequency domain weight calculation formula is: , The default parameters are , , ; They are respectively the normalized spatial domain perception features, time domain perception features and frequency domain perception features; Perceptual importance factor calculation: The normalized multi-domain perceptual features are weighted summed with the dynamic fusion weights to obtain the perceptual importance factor of the coding unit. , the specific calculation formula is: ， in are spatial domain weight, time domain weight and frequency domain weight respectively. They are respectively the normalized spatial domain perception features, time domain perception features and frequency domain perception features; S3: Constructing a perceptual distortion evaluation value based on the perceptual importance factor , and the rate-distortion objective function in the video coding standard method Replace with , used to guide the code rate allocation of the coding unit, including the bit resource allocation of the coding unit and the iterative optimization of the quantization parameter; is the distortion term in the video coding standard method, is the Lagrange multiplier, The bitrate used for encoding; S4: Based on the perceptual importance factor calculated in S2, the quantization parameter adjustment amount of each coding unit is calculated through a nonlinear mapping function; the basic quantization parameter value preset in the video coding standard method is added to the quantization parameter adjustment amount to obtain the final quantization parameter value of the coding unit, which is used for video coding to achieve adaptive rate-distortion optimization; the nonlinear mapping uses a hyperbolic tangent function to achieve nonlinear transformation of the quantization parameter adjustment amount.

2. The video coding method based on multi-domain perceptual feature fusion according to claim 1, wherein: The spatial domain perception feature extraction includes: calculating the pixel gradient amplitude through the Sobel operator; calculating the brightness standard deviation based on a 3×3 neighborhood; using the four-neighborhood prediction error as the spatial masking value, accumulating the pixel-level features and taking the average value to obtain the edge, contrast and spatial masking features of each coding unit, and fusing them into the spatial domain comprehensive perception features with a weight of 0.4:0.3:0.

3.

3. The video coding method based on multi-domain perceptual feature fusion according to claim 1, wherein: The temporal domain perception feature extraction is as follows: calculating the absolute brightness difference between the current frame and the corresponding position of the previous frame, taking the average of the difference of pixels covered by each coding unit as the temporal domain perception feature, and setting the temporal domain perception feature of the first frame to 0.

4. The video coding method based on multi-domain perceptual feature fusion according to claim 1, wherein: The frequency domain perceptual feature extraction includes: performing Fourier transform on the current frame to obtain a spectrum diagram, using an annular bandpass filter with an inner radius of 10 and an outer radius of 60 to filter out low-frequency and high-frequency components, and retaining intermediate-frequency components; then obtaining an intermediate-frequency feature map through inverse Fourier transform, and calculating the pixel mean of the intermediate-frequency feature map corresponding to the area covered by each coding unit as the frequency domain perceptual feature of the coding unit.

5. The video coding method based on multi-domain perceptual feature fusion according to claim 1, wherein: The feature normalization adopts the Min-Max method, and the formula is: ， in is the spatial domain perception feature, time domain perception feature, or frequency domain perception feature of the coding unit, is the normalized eigenvalue.

6. The video coding method based on multi-domain perceptual feature fusion according to claim 1, wherein: The video coding standard method is AVC, HEVC, VVC video coding standard.

7. The video coding method based on multi-domain perceptual feature fusion according to claim 1, wherein: The perceptual distortion evaluation value is obtained by Calculate; where is the distortion term in the video coding standard method, is the perceptual importance factor of the coding unit, The default weighting factor is 0.

5.

8. The video coding method based on multi-domain perceptual feature fusion according to claim 1, wherein: The quantization parameter adjustment amount Calculation by hyperbolic tangent function mapping: ， The default parameters are , That is, the clipping function is used to limit the calculation results to the interval [-5,5]. is the perceptual importance factor of the coding unit; final The values ​​are: ， in is the base quantization parameter of the encoder, and .

Images