A video region coding method and system fusing saliency detection

Abstract

The application relates to the technical field of video coding, in particular to a video region coding method and system fusing saliency detection, which comprises the following steps: generating a basic saliency map of a current video frame according to the spatial texture contrast feature, the temporal motion amplitude feature and the chroma feature of the current video frame; determining the motion reliability coefficient of each pixel in the current video frame based on the local statistical characteristics of the motion offset information between the current video frame and a historical video frame; fusing the basic saliency map and a historical frame cumulative saliency map according to the motion reliability coefficient to generate a target space-time saliency map of the current video frame; dividing the current video frame into coding regions of different levels according to the target space-time saliency map; and performing region coding on the current video frame according to the different coding parameters corresponding to the coding regions of different levels. Thus, the overall code rate is reduced and the coding efficiency is improved under the premise of ensuring the subjective picture quality.

Description

Technical Field

[0001] This application relates to the field of video coding technology, specifically to a video region coding method and system that integrates saliency detection. Background Technology

[0002] With the widespread adoption of multimedia services such as ultra-high-definition video, live streaming, remote interaction, and high-definition streaming media transmission, video data volume is growing exponentially. The contradiction between limited network transmission bandwidth and terminal storage and decoding computing resources is becoming increasingly acute. Video coding technology, as a core means of alleviating this contradiction, faces the critical challenge of balancing compression efficiency with visual perception quality. Since the human visual system naturally pays attention to different areas of a screen differently, how to fully utilize the human visual attention mechanism during the encoding process, and how to allocate bitrate resources towards areas of visual focus, has become an important research direction for improving the subjective quality and transmission efficiency of video coding.

[0003] Existing video coding methods mostly employ a globally uniform compression strategy, applying the same quantization coding intensity to all areas of the image, failing to fully consider the requirements of human visual characteristics for coding quality. While some technologies introduce saliency detection to distinguish visually relevant regions from background regions and attempt to perform differentiated coding based on region importance, existing saliency detection methods typically analyze only single-dimensional visual features. In complex dynamic scenes, they struggle to stably and accurately represent the visual importance of each pixel region, thus limiting the reliability of saliency-based region coding strategies in practical applications.

[0004] Due to the shortcomings of the above-mentioned encoding strategies in visual feature analysis and encoding resource allocation, it is difficult to achieve an effective balance between compression efficiency and subjective visual quality during the video encoding process. This can easily lead to insufficient encoding of key visual areas, resulting in loss of details, and excessive redundant encoding of background areas, resulting in wasted bitrate. Overall encoding performance and subjective viewing experience are difficult to balance. The encoding efficiency of existing video encoding methods in complex dynamic scenes still needs to be improved. Summary of the Invention

[0005] To address the technical challenge of improving coding efficiency in complex dynamic scenarios, this application aims to provide a video region coding method that integrates saliency detection. The specific technical solution adopted is as follows: Firstly, a video region coding method integrating saliency detection is provided, comprising: generating a basic saliency map of the current video frame based on its spatial texture contrast features, temporal motion amplitude features, and chromaticity features; determining the motion reliability coefficient of each pixel in the current video frame based on the local statistical characteristics of motion offset information between the current video frame and historical video frames; fusing the basic saliency map with the cumulative saliency map of historical frames based on the motion reliability coefficient to generate a target spatiotemporal saliency map of the current video frame, wherein the cumulative saliency map of historical frames is obtained by accumulating the basic saliency maps of multiple consecutive historical video frames through motion compensation and temporal decay; dividing the current video frame into coding regions of different levels based on the target spatiotemporal saliency map; and performing region coding on the current video frame according to the differentiated coding parameters corresponding to the different levels of coding regions.

[0006] One possible approach is to generate a basic saliency map of the current video frame based on its spatial texture contrast features, temporal motion amplitude features, and chromaticity features. This includes: determining the spatial texture contrast features centered on the target pixel based on the ratio of the grayscale variance of pixels in the target pixel's neighborhood to the grayscale variance of the global image; calculating the pixel-level motion offset between the current video frame and adjacent video frames, and determining the temporal motion amplitude features based on the magnitude of the pixel-level motion offset; converting the current video frame to a color space that separates luminance and chrominance, and determining the chromaticity features based on the deviation of each pixel's chromaticity component from the mean of the chromaticity components of the entire frame; and then normalizing the spatial texture contrast features, temporal motion amplitude features, and chromaticity features, and then summing them pixel by pixel to obtain the basic saliency map.

[0007] One possible approach is to determine the motion reliability coefficient of each pixel in the current video frame based on the local statistical characteristics of motion offset information between the current and historical video frames. This includes: constructing a local motion statistics window centered on the target pixel, based on the motion offset information between the current and historical video frames; calculating the standard deviation and mean of the magnitudes of the motion offset information for all pixels within the local motion statistics window; and determining the motion reliability coefficient based on the ratio of the standard deviation to the mean.

[0008] One possible approach to generating a historical frame cumulative saliency map includes: acquiring the historical base saliency maps corresponding to multiple consecutive historical video frames; performing a distortion mapping on the historical base saliency map corresponding to each historical video frame based on the pixel-by-pixel motion offset information from the current video frame to each historical video frame to obtain an aligned historical saliency map; and weighting and normalizing the aligned historical saliency maps according to a temporal decay factor to generate the historical frame cumulative saliency map, where the temporal decay factor characterizes the degree of attenuation of the saliency contribution from the time distance between the historical video frame and the current video frame.

[0009] As one possible approach, the process of determining the target spatiotemporal saliency map includes: determining the fusion weights of the base saliency map and the cumulative saliency map of historical frames based on the motion reliability coefficient; and fusing the base saliency map and the cumulative saliency map of historical frames pixel-by-pixel based on the fusion weights of the base saliency map and the cumulative saliency map of historical frames to obtain the target spatiotemporal saliency map.

[0010] One possible approach is to divide the current video frame into different levels of coding regions based on the target spatiotemporal saliency map, including: dividing the target spatiotemporal saliency map into multiple coding tree units; determining the average saliency value of pixels within each coding tree unit; determining a high saliency threshold benchmark based on the product of a first preset scaling factor and the average value; determining a low saliency threshold based on the product of a second preset scaling factor and the average value; for each pixel in the current video frame, determining the pixel's own high saliency threshold based on the high saliency threshold benchmark, a third preset scaling factor, and the motion reliability coefficient corresponding to the pixel; comparing the pixel's saliency value with the pixel's own high saliency threshold and the low saliency threshold of the coding tree unit to which the pixel belongs; classifying the region composed of pixels whose saliency value is greater than or equal to the pixel's own high saliency threshold as a high saliency core region; classifying the region composed of pixels whose saliency value is less than the low saliency threshold of the coding tree unit to which the pixel belongs as a low saliency background region; and classifying the region composed of pixels whose saliency value is greater than or equal to the low saliency threshold of the coding tree unit to which the pixel belongs and less than the pixel's own high saliency threshold as a medium saliency transition region.

[0011] One possible approach is to determine the fusion weights of the basic saliency map and the historical frame cumulative saliency map based on the motion reliability coefficient. This includes: obtaining a preset adjustment coefficient; and determining the historical frame cumulative saliency map fusion weights based on the adjustment coefficient and the motion reliability coefficient. The sum of the basic saliency map fusion weights and the historical frame cumulative saliency map fusion weights is a preset total weight value.

[0012] One possible approach is to perform region-based coding on the current video frame based on differentiated coding parameters corresponding to different levels of coding regions. This includes: obtaining global baseline quantization parameters set for the current video frame; dividing the target spatiotemporal saliency map into multiple coding tree units and determining the average value of the normalized saliency values ​​of pixels within each coding tree unit; determining the adaptive quantization parameters for the corresponding coding tree unit based on the global baseline quantization parameters, coding adjustment coefficients, and the average value of the normalized saliency values ​​within each coding tree unit; and performing coding on the highly saliency core region, the moderately saliency transition region, and the low saliency background region based on the adaptive quantization parameters. During inter-frame coding, motion offset information between the current video frame and historical video frames is reused to match the saliency region positions between consecutive video frames, ensuring that coding regions containing temporally consecutive saliency targets have consistent coding parameters. After completing the region-based differentiated coding, the global baseline quantization parameters are dynamically adjusted based on the pixel proportion of the highly saliency core region, and subsequent video frames are coded based on the adjusted global baseline quantization parameters to generate the target compressed bitstream.

[0013] One possible approach is to determine the current video frame, which includes: converting the acquired raw video data into a color space with a preset encoding format, and performing resolution and frame rate alignment. Then, performing video frame noise reduction and contrast enhancement on the converted video frame. Finally, dividing each video frame into blocks according to the size of the coding tree unit to obtain the preprocessed current video frame.

[0014] Secondly, a video region coding system integrating saliency detection is provided, including: The saliency map generation module is used to generate a basic saliency map of the current video frame based on its spatial texture contrast features, temporal motion amplitude features, and chromaticity features.

[0015] The reliability determination module is used to determine the motion reliability coefficient of each pixel in the current video frame based on the local statistical characteristics of motion offset information between the current video frame and historical video frames.

[0016] The spatiotemporal fusion module is used to fuse the basic saliency map with the historical frame cumulative saliency map according to the motion reliability coefficient to generate the target spatiotemporal saliency map of the current video frame. The historical frame cumulative saliency map is obtained by accumulating the basic saliency maps of multiple consecutive historical video frames through motion compensation and temporal decay.

[0017] The region segmentation module is used to divide the current video frame into different levels of coding regions based on the target spatiotemporal saliency map.

[0018] The encoding execution module is used to perform regional encoding on the current video frame according to the differentiated encoding parameters corresponding to different levels of encoding regions.

[0019] This application offers the following advantages: A basic saliency map is generated using spatial texture contrast features, temporal motion amplitude features, and chromaticity features. A motion reliability coefficient is introduced to fuse this basic saliency map with a historical frame cumulative saliency map obtained through motion compensation and temporal attenuation, resulting in a spatiotemporal saliency map. This spatiotemporal saliency map is then used to divide coding regions into different levels and assign differentiated coding parameters for regional coding. This improves the robustness of saliency detection in complex dynamic scenes (such as occlusion, sudden changes in illumination, and motion blur), effectively avoiding missed detections in key visual areas and over-detection in background areas. Simultaneously, through differentiated regional coding based on saliency levels, bitrate resources are tilted towards high-interest areas, significantly compressing redundant background data. This reduces the overall bitrate while maintaining subjective image quality, thereby improving coding efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a video region coding method that integrates saliency detection, provided as an embodiment of this application. Figure 2 This is a schematic diagram of the system architecture of a video region coding system that integrates saliency detection, provided as an embodiment of this application. Detailed Implementation

[0021] The following description, in conjunction with the accompanying drawings, details a specific scheme for a video region coding method that integrates saliency detection, as provided in this application.

[0022] Please see Figure 1 It illustrates a flowchart of a video region coding method that integrates saliency detection according to an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps: S100-S500.

[0023] S100. Generate the basic saliency map of the current video frame based on the spatial texture contrast features, temporal motion amplitude features, and chromaticity features of the current video frame.

[0024] The human eye's visual attention mechanism is naturally sensitive to abrupt changes in texture, object motion, and color contrast. However, in dynamic scenarios such as live streaming and surveillance, the visual characteristics of various scenes differ significantly: on the one hand, relying solely on spatial texture contrast to generate saliency maps can easily misclassify background areas with complex textures (such as swaying leaves or tile patterns) as salient targets, while missing foreground targets with flat textures (such as solid-color sportswear or smooth car surfaces) as background; on the other hand, when a salient target is stationary or in extremely slow motion, its motion characteristics are almost zero, and temporal motion information alone cannot distinguish it from the background. Furthermore, in scenes with flat textures and insignificant motion, color contrast often becomes crucial for the human eye to identify foreground targets.

[0025] To fully capture salient cues in different dimensions of video frames, three complementary features—spatial texture contrast, temporal motion amplitude, and chromaticity contrast—are used to generate a basic salient map that reflects the visual importance of each region through pixel-by-pixel fusion.

[0026] S200. Based on the local statistical characteristics of motion offset information between the current video frame and historical video frames, determine the motion reliability coefficient of each pixel in the current video frame.

[0027] In this embodiment, the basic saliency map cannot handle temporal anomalies in video sequences. For example, when a moving target is briefly occluded by a foreground object in a video sequence, the motion vector of the corresponding pixel of the moving target suddenly disappears in the occluded area, causing a sharp drop in the saliency response of the occluded area, and the basic saliency map shows a lack of saliency values ​​in the occluded area. To address the saliency loss problem caused by occlusion or sudden illumination changes, a feasible compensation strategy is to utilize the basic saliency map already generated in historical video frames and transfer the saliency information from historical frames to the current frame through motion vector distortion mapping to fill the saliency loss area. However, the above compensation strategy is effective only if the motion vectors between the current frame and historical frames are stable and reliable: when occlusion or sudden illumination changes occur, the optical flow field in the neighborhood of the pixel corresponding to the moving target becomes disordered, and the direction and magnitude of the motion vector fluctuate significantly. If the historical information is still transferred by relying on the motion vector, incorrect saliency information will be introduced into the current frame, causing incorrect compensation of the saliency map.

[0028] To address the aforementioned issues, this step analyzes the local statistical characteristics of motion offset information between the current video frame and historical video frames to construct a motion reliability coefficient for each pixel in the current frame. This motion reliability coefficient quantifies the stability of the optical flow field within the local region corresponding to the current pixel, thereby determining whether the current pixel is suitable for historical frame saliency transfer compensation.

[0029] In this embodiment, motion offset information refers to the pixel-by-pixel motion vector between the current video frame and the previous video frame. The motion vector includes horizontal and vertical offsets. Motion offset information can be calculated using dense optical flow and reflects the displacement of each pixel in the video frame between adjacent frames.

[0030] Local statistical characteristics refer to the statistical distribution of motion vector magnitudes of all pixels within a local window centered on the target pixel. Specifically, this includes the mean and standard deviation of the motion vector magnitudes. The mean reflects the overall motion intensity of the corresponding local region within the local window, while the standard deviation reflects the consistency of motion vectors within that local region. When the optical flow field is uniformly distributed within the local region and the direction and magnitude of the motion vectors of each pixel are similar, the standard deviation is small; when occlusion or abrupt changes in illumination occur within the local region, or when the direction and magnitude of the motion vectors become disordered, the standard deviation increases significantly.

[0031] The motion reliability coefficient is a scalar value calculated based on local statistical characteristics, ranging from 0 to 1. It quantifies the reliability of the optical flow field within the local region corresponding to a pixel: when the motion reliability coefficient approaches 1, it indicates that the local optical flow field is stable and reliable, suitable for historical frame saliency propagation compensation; when the motion reliability coefficient approaches 0, it indicates that the local optical flow field is disordered and unreliable, and historical frame saliency propagation should be weakened or avoided.

[0032] S300. Based on the motion reliability coefficient, the basic saliency map and the historical frame cumulative saliency map are fused to generate the target spatiotemporal saliency map of the current video frame.

[0033] Among them, the historical frame cumulative saliency map is obtained by accumulating the basic saliency maps of multiple consecutive historical video frames through motion compensation and temporal decay.

[0034] To compensate for the loss of saliency values ​​caused by occlusion or sudden changes in illumination, this application introduces a historical frame cumulative saliency map. The historical frame cumulative saliency map is obtained by accumulating the base saliency maps of multiple consecutive historical video frames preceding the current frame through motion compensation and temporal attenuation. Motion compensation refers to using pixel-by-pixel motion offset information from historical video frames to the current frame to distort and map the historical base saliency map, aligning the salient regions of moving targets in historical frames spatially with their corresponding positions in the current frame. Temporal attenuation assigns different accumulation weights to historical frames at different temporal distances from the current frame; the farther the historical frame, the lower the weight, to weaken the interference of potentially invalid background information in older frames on the current saliency judgment. Through motion compensation and temporal attenuation, the historical frame cumulative saliency map can provide the current frame with spatially aligned and temporally weighted saliency reference information from historical frames.

[0035] However, the transmission of historical frame saliency information depends on the accuracy of motion vectors. When motion vectors are unreliable, introducing historical frame information can lead to incorrect compensation. Therefore, based on the calculated motion reliability coefficient, the contribution ratio of the base saliency map and the historical frame cumulative saliency map during fusion is determined. The motion reliability coefficient ranges from 0 to 1 and is used to quantify the stability of the optical flow field within the local region corresponding to each pixel. When the motion reliability coefficient approaches 1, it indicates that the motion vector is reliable, and the fusion weight of the historical frame cumulative saliency map should be increased to fully utilize historical information to fill in the missing saliency value areas. When the motion reliability coefficient approaches 0, it indicates that the motion vector is unreliable, and the fusion weight of the base saliency map should be increased, prioritizing the features of the current frame to avoid the incorrect transmission caused by unreliable motion vectors. The final generated target spatiotemporal saliency map integrates the instantaneous saliency features of the current frame and the temporal saliency features of historical frames. It can effectively fill in the missing saliency value areas caused by occlusion and suppress illumination abrupt disturbances while preserving texture abrupt changes, motion changes, and color contrast information.

[0036] As one possible approach, the process for determining the target spatiotemporal saliency map includes: S301-S302.

[0037] S301. Determine the fusion weight of the basic saliency map and the fusion weight of the historical frame cumulative saliency map based on the motion reliability coefficient.

[0038] To dynamically balance the fusion ratio of current frame spatial salient features and historical temporal salient features, fusion weights are calculated based on the motion reliability coefficient. When the motion reliability coefficient is high, the fusion ratio of historical information is increased to fill occlusion gaps; when the motion reliability coefficient is low, the fusion ratio of historical information is reduced, with current frame features taking precedence, to avoid error propagation caused by sudden changes in illumination or occlusion.

[0039] In some embodiments, a preset control coefficient is obtained, and the historical frame cumulative saliency map fusion weight is determined based on the control coefficient and the motion reliability coefficient.

[0040] Specifically, historical frame cumulative saliency map fusion weights The calculation formula is: Basic saliency map fusion weights Weights fused with historical frame cumulative saliency maps The sum of these values ​​represents the preset total weight value. In this embodiment, the preset total weight value can be 1, in which case the base saliency map fusion weight is... The calculation formula is: in, This is a preset control coefficient used to constrain the proportion of historical information and prevent over-reliance on historical frames from causing target contour lag. In this embodiment... 0.8 can be taken. Represents pixels The motion reliability coefficient.

[0041] S302. Based on the fusion weights of the basic saliency map and the fusion weights of the historical frame cumulative saliency map, the basic saliency map and the historical frame cumulative saliency map are fused pixel by pixel to obtain the target spatiotemporal saliency map.

[0042] When the motion reliability coefficient is high, the fusion weight of the historical frame cumulative saliency map is larger, and the fusion result will be dominated by the historical frame cumulative saliency map, thus filling in the missing saliency value regions in the current frame caused by occlusion. When the motion reliability coefficient is low, the fusion weight of the basic saliency map is larger, and the fusion result will be dominated by the current frame basic saliency map, thus avoiding erroneous compensation caused by unreliable motion vectors. Through the above adaptive weighted fusion, the target spatiotemporal saliency map can integrate the texture abrupt changes, motion changes, color contrast features of the current frame, and reliable temporal features of historical frames to form a complete and robust saliency representation.

[0043] Specifically, the pixels in the spatiotemporal saliency map of the target Significance value : in, Represents pixels The saliency value of the corresponding target spatiotemporal saliency map; The coordinates in the basic saliency map are: The significant value of the pixel; Indicates the current pixel. Historical cumulative saliency values ​​after motion offset information distortion mapping. By adaptively balancing the fusion ratio of instantaneous spatial saliency features and historical temporal saliency features, the problem of missed detection of salient targets caused by the failure of a single feature is avoided.

[0044] For each pixel in the current video frame, the basic saliency map and the historical frame cumulative saliency map are fused pixel-by-pixel according to the fusion weights of the basic saliency map and the fusion weights of the historical frame cumulative saliency map to obtain the target spatiotemporal saliency map. The saliency value of each pixel in the target spatiotemporal saliency map comprehensively reflects the immediate saliency features of the current frame and the temporal saliency features of the historical frames, and can accurately characterize the visual importance of different pixels within the video frame.

[0045] S400. Based on the target spatiotemporal saliency map, divide the current video frame into different levels of coding regions.

[0046] In traditional video encoding and transmission, globally uniform encoding quantization parameters cannot adapt to the characteristics of human vision: if the same compression ratio is used for the entire frame, details will be lost and the image will be blurred for salient targets that the human eye focuses on; while retaining too much bitrate for background areas with no visual value will generate a large amount of redundant data and occupy transmission bandwidth. Therefore, based on the visual importance reflected by the spatiotemporal saliency map of the target, the video frame is adaptively divided into regions, and the pixels within the frame are divided into three levels: highly saliency core region, moderately saliency transition region, and low saliency background region.

[0047] S500: Perform regional coding on the current video frame according to the differentiated coding parameters corresponding to different levels of coding regions.

[0048] After region segmentation, the current video frame is encoded regionally based on differentiated coding parameters corresponding to different coding levels. Differential coding parameters refer to setting different quantization parameters for highly salient core regions, moderately salient transition regions, and low-salient background regions. Quantization parameters determine the compression intensity: smaller quantization parameters result in lower compression levels and more complete preservation of image details, but at a higher bitrate; larger quantization parameters result in higher compression intensity and a lower bitrate, but at a more significant loss of detail. This step adaptively adjusts the quantization parameters based on the saliency value of each pixel in the target spatiotemporal saliency map. This ensures that highly salient core regions use smaller quantization parameters to maintain image quality, low-salient background regions use baseline quantization parameters, and moderately salient transition regions use appropriate quantization parameters. This skewed allocation of bitrate resources is achieved by reducing the compression intensity in highly salient regions.

[0049] Furthermore, during inter-frame coding, the positions of salient regions between consecutive video frames are matched based on the calculated motion offset information between the current video frame and historical video frames. By reusing the motion offset information, temporally consecutive salient foreground targets maintain consistency in quantization parameters during cross-frame coding, avoiding image quality jitter or blurring caused by inter-frame jumps in quantization parameters. Simultaneously, after completing region-specific differential coding, this step dynamically adjusts the overall compression ratio based on the pixel proportion of highly salient core regions in the current frame. When the proportion of highly salient core regions is large, the overall compression intensity is reduced to ensure multi-target image quality; when the proportion of highly salient core regions is small, the overall compression ratio is increased to maximize bandwidth savings.

[0050] In one design, prior to performing the aforementioned saliency detection and coding steps, a preprocessing step for the raw video data is included: S600-S800. The purpose of the preprocessing step is to convert the acquired raw video data into a standardized video frame sequence with uniform format, noise suppression, and ease of subsequent processing, providing high-quality, time-aligned input data for saliency detection and region coding.

[0051] The S600 converts the acquired raw video data into a color space with a preset encoding format and performs resolution and frame rate alignment.

[0052] Raw, uncompressed video data is acquired via a camera, video capture card, or streaming media interface. The acquired RGB format raw video is uniformly converted to the YUV color space (a color encoding format that separates luminance and chrominance) to adapt to the encoding standard. Resolution cropping and scaling, frame rate alignment, removal of abnormal dropped frames and duplicate redundant frames are performed, and the frames are cached sequentially into the encoding buffer to form a continuous video frame data stream with normalized format.

[0053] S700 performs video frame noise reduction and contrast enhancement processing on the converted video frames.

[0054] Algorithms such as Gaussian filtering and mean filtering are used to remove salt-and-pepper noise and Gaussian noise that may be introduced into video frames during the acquisition process. For frames with insufficient lighting or blurriness, adaptive contrast enhancement processing is performed to weaken the negative impact of interference information on subsequent saliency calculation and improve the accuracy of saliency map calculation.

[0055] S800: Divide each video frame into blocks according to the size of the coding tree unit to obtain the preprocessed current video frame.

[0056] The video frame that has undergone noise reduction and enhancement is evenly divided into blocks according to the macroblock or coding tree unit size defined by the coding standard, forming basic coding tree units to facilitate region segmentation, feature calculation, and final coding operations. The processed video frame then becomes the current video frame.

[0057] In one design, the above S100 includes S101-S104.

[0058] S101. Taking the target pixel as the center, determine the spatial texture contrast feature based on the ratio of the gray-level variance of the pixels in the neighborhood of the target pixel to the gray-level variance of the global image.

[0059] Specifically, the current video frame is first converted into a grayscale image. For frames with coordinates of... For any target pixel, a 3×3 local neighborhood window is constructed centered on it. The local variance of the grayscale values ​​of all pixels within this neighborhood is calculated. Simultaneously, the global variance of the grayscale values ​​of all pixels in the entire frame is calculated. Furthermore, centered on the target pixel, the spatial texture contrast feature value of the target pixel is determined based on the ratio of the grayscale variance of the pixels within the target pixel's neighborhood to the global image grayscale variance. If the global image grayscale variance is zero, then let Among them, the spatial texture contrast feature value reflects the degree of change in texture complexity of a local region of a pixel relative to the entire image frame. Regions with abrupt texture changes (such as object edges and detailed foregrounds) Values ​​are typically high, while smooth regions... The value is low.

[0060] S102. Calculate the pixel-level motion offset between the current video frame and the adjacent video frames, and determine the temporal motion amplitude characteristics based on the magnitude of the pixel-level motion offset.

[0061] Specifically, the previous video frame is obtained from the current video frame. Dense optical flow is used to calculate the motion vector for each pixel from the current video frame to the previous video frame; this motion vector includes the horizontal offset. and vertical offset Pixel Time-domain motion amplitude eigenvalues The calculation formula is as follows: in, Represents pixels The time-domain motion amplitude characteristic value, and Each pixel Motion offset components in the horizontal and vertical directions. Temporal motion amplitude features quantify the intensity of pixel motion; moving foreground objects (such as walking pedestrians or swinging rackets) typically have higher [motion offset characteristics]. Value, while the static background area The value approaches zero.

[0062] S103. Convert the current video frame to a color space that separates luminance and chrominance, and determine the chrominance features based on the degree of deviation of the chrominance component of each pixel from the average chrominance component of the entire frame.

[0063] A uniformly colored object (such as a solid-color tracksuit or a smooth car surface) has a very low local standard deviation in the spatial domain. Traditional texture contrast features would classify it as background, leading to missed detections of salient areas. When a salient target is stationary, its temporal motion amplitude is almost zero, making it impossible for motion features to distinguish it from the background. For scenes with flat textures or insignificant motion, this paper leverages the human eye's sensitivity to color contrast to highlight potentially overlooked salient foreground objects from a color perspective. Specifically, the current video frame is converted from the Red Green Blue (RGB) color space to the CIELAB color space (Lab color space). The Lab space separates luminance (L channel) and chrominance (a and b channels), allowing for more accurate measurement of color differences. The mean values ​​of the entire frame in the a and b channels are calculated and denoted as follows: and For pixels Its chromaticity characteristic value The calculation formula is as follows: in, Represents pixels chromaticity characteristic values, and These represent the values ​​of the pixel in channel a and channel b, respectively. This feature value represents the Euclidean distance between the pixel's chromaticity and the dominant color of the image background. The larger the distance, the more likely the pixel is to be a prominent foreground element, thus effectively suppressing interference from large areas of background color.

[0064] S104. After normalizing the spatial texture contrast features, temporal motion amplitude features, and chromaticity features, add them pixel by pixel to obtain the basic saliency map.

[0065] The spatial texture contrast features, temporal motion amplitude features, and chromaticity features are normalized respectively to map their numerical ranges to the same preset interval. Normalization employs a linear mapping: for each feature, its global maximum and global minimum are calculated. If the difference between the global maximum and global minimum is not zero, the normalized value for each pixel is (original value - global minimum) / (global maximum - global minimum); if the difference is zero, the normalized value is 0. The values ​​at corresponding positions of the three normalized features are then summed pixel by pixel to obtain the basic saliency map.

[0066] In this way, by complementing the advantages of the three features, the basic saliency map can capture visual saliency information of different dimensions in the current frame more comprehensively.

[0067] In one design, the above S200 specifically includes: S201-S203.

[0068] S201. Using the target pixel as the center, construct a local motion statistics window based on the motion offset information between the current video frame and historical video frames.

[0069] Specifically, based on the calculated dense optical flow field from the current video frame to the previous video frame, for each pixel... A 5×5 local motion statistics window is constructed centered on the target pixel. This local motion statistics window covers the motion vector information of the target pixel and its surrounding neighborhood, and is used to evaluate the consistency of the optical flow field within a local area.

[0070] S202. Calculate the standard deviation and average value of the magnitude of the motion offset information of all pixels within the local motion statistics window.

[0071] Specifically, within this 5×5 local motion statistics window, the average value of the motion amplitude (i.e., motion vector magnitude) of all pixels is calculated. and standard deviation .average value This reflects the overall motion intensity of the local area, and its standard deviation. This reflects the degree of consistency of motion vectors within the region. When the optical flow field is uniform, the motion vectors of each pixel are similar. Smaller; when occlusion or sudden changes in lighting occur, the direction and magnitude of local motion vectors become disordered. Significantly increased.

[0072] S203. Determine the motion reliability coefficient based on the ratio of the standard deviation to the mean.

[0073] Motion reliability coefficient The formula used to quantify the stability of the optical flow field at the current pixel is as follows: in, Represents pixels The motion reliability coefficient, The standard deviation of the optical flow amplitude within the local motion statistics window. This represents the average amplitude of optical flow motion within the local motion statistics window. This is a very small constant, with a value of 0.001, used to prevent the denominator from being zero. This represents an exponential function with the natural constant e as its base. The motion reliability coefficient ranges from 0 to 1. When the target motion is smooth and there is no obstruction or interference, the optical flow distribution is uniform. Smaller than average Approaching 1, allowing for normal propagation of historical saliency; when target occlusion or sudden changes in illumination occur, local optical flow becomes disordered. Significantly increased, If the value approaches 0, the motion vector is deemed unreliable, thus weakening the transmission of historical frames.

[0074] As one possible approach, the process of generating a cumulative saliency map of historical frames includes: S204-S206.

[0075] S204. Obtain the historical basic saliency map corresponding to each of multiple consecutive historical video frames.

[0076] Get the continuous frames before the current video frame One historical video frame: , ,…, At the same time, obtain continuous The underlying saliency map generated from historical video frames: , ,…, . The preset number of historical frames can be set according to actual application needs, for example... The value is 7.

[0077] S205. Based on the pixel-by-pixel motion offset information from the current video frame to each historical video frame, perform a distortion mapping on the historical base saliency map corresponding to each historical video frame to obtain the aligned historical saliency map.

[0078] In order to accurately transmit historical salient information to the corresponding position in the current frame, spatial alignment is required through motion compensation.

[0079] Get from the current video frame to the 1st frame Pixel-by-pixel motion offset information of the frame The pixel-by-pixel motion offset information is obtained by concatenating the optical flow between adjacent frames, so that the historical frames... The median coordinate is The pixels correspond to the pixels in the current frame. Furthermore, pixel-by-pixel motion offset information is used to perform a distortion mapping on the historical baseline saliency map, so that the pixels in the historical baseline saliency map... Pixels corresponding to the current frame Location. After distortion mapping, salient regions of moving targets in historical frames are aligned to their corresponding coordinates in the current frame, providing a spatial consistency basis for subsequent accumulation and fusion.

[0080] S206. Based on the temporal decay factor, the aligned historical saliency map is weighted, accumulated, and normalized to generate a historical frame cumulative saliency map.

[0081] Introducing a time decay factor The aligned historical saliency map is then weighted and accumulated using the following formula: in, Indicates the current pixel Historical cumulative significant values ​​after motion offset information distortion mapping; Indicates multiplication operation; This is the time-series decay factor, which can be set to 0.75 to control the decay rate of historical information. This is the historical frame sequence number index, with values ​​ranging from 1 to... ; Indicates the first The historical saliency map of the frame in coordinates The significant value at that point, the coordinates of which correspond to the current frame pixel. The historical frame position is pointed to after optical flow offset. The temporal attenuation factor is used to characterize the degree of attenuation of the contribution of historical video frames to saliency due to the time distance between them and the current video frame. The further away a historical frame is from the current video frame, the lower the weight of its corresponding historical baseline saliency map. Through this accumulation mechanism, historical saliency can be passed to fill the information gaps in the current frame, while the attenuation rate of historical information can be controlled to avoid outdated information interfering with the current judgment.

[0082] In one design, the above S400 includes: S401-S408.

[0083] S401. Divide the target spatiotemporal saliency map into multiple coding tree units.

[0084] The generated spatiotemporal saliency map is divided into several non-overlapping coding tree units (CTUs) according to the size defined by the coding standard. Each coding tree unit contains multiple pixels. Subsequent region partitioning is performed using coding tree units as the basic granularity, which can balance coding efficiency and region partitioning accuracy.

[0085] S402. Determine the average value of pixel saliency within each coding tree unit.

[0086] For the Each coding tree unit calculates the average of the salient values ​​of all pixels within its coverage area. . It reflects the overall visual saliency level of the coding tree unit.

[0087] S403. Determine the high significance threshold benchmark based on the product of the first preset scaling factor and the average value; determine the low significance threshold based on the product of the second preset scaling factor and the average value; determine the high significance threshold adjustment factor based on the third preset scaling factor and the motion reliability coefficient corresponding to the pixel.

[0088] High significance threshold benchmark The calculation formula is: in, This is the first preset proportional coefficient.

[0089] Low significance threshold The calculation formula is: in, This is the second preset proportional coefficient.

[0090] Highly significant threshold regulation factor The calculation formula is: in, This is the third preset scaling factor. For example, The value can be 0.8. The value can be 0.4. The value can be 0.5.

[0091] S404. For each pixel in the current video frame, determine the pixel's own high significance threshold based on the high significance threshold benchmark and the high significance threshold adjustment factor.

[0092] The pixel's own high significance threshold The calculation formula is: S405. For each pixel in the current video frame, compare the saliency value of the pixel with the pixel's own high saliency threshold and the low saliency threshold of the coding tree unit where the pixel is located.

[0093] For each pixel in the current frame Obtain the saliency value of the pixel in the spatiotemporal saliency map of the target. and the calculated and Compare them.

[0094] S406. The region consisting of pixels whose significance value is greater than or equal to the high significance threshold of the pixel itself is divided into a high significance core region.

[0095] like If the saliency value is significantly higher than the average level of the coding tree unit, then the pixel is classified as a highly saliency core region. This region corresponds to pixels with saliency values ​​significantly higher than the average level of the unit in the coding tree, which are usually foreground objects that attract high attention from the human eye. The lowest compression ratio will be used to preserve complete details in this region.

[0096] S407. The region consisting of pixels whose salience values ​​are less than the low salience threshold of the coding tree unit where the pixel is located is divided into a low salience background region.

[0097] like If the pixel is not significantly different from the background pixel, it is classified as a low-salience background region. Low-salience background regions have low visual importance, and the human eye is not sensitive to the loss of image quality from them. Therefore, a high compression ratio can be used to save bitrate.

[0098] S408. The region consisting of pixels whose significance values ​​are greater than or equal to the low significance threshold of the coding tree unit where the pixel is located and less than the high significance threshold of the pixel itself is divided into a medium significance transition region.

[0099] Will satisfy The pixels are divided into medium-saliency transition regions. These regions are located in the transition zone between the foreground and background and are compressed at a moderate ratio to balance image quality and bitrate.

[0100] In one design, the aforementioned S500 includes S501-S507.

[0101] S501. Obtain the global baseline quantization parameters set for the current video frame.

[0102] Set a global baseline quantization parameter This serves as the baseline compression level for encoding. In this embodiment, A value of 32 can be used, which corresponds to a medium compression ratio in regular video encoding and can provide an acceptable balance between image quality and bitrate in most application scenarios.

[0103] S502. Normalize the pixel saliency values ​​in the target spatiotemporal saliency map to a preset numerical range to obtain normalized saliency values.

[0104] To eliminate interference caused by differences in the numerical range and distribution of saliency values ​​between different frames, the original saliency values ​​in the target spatiotemporal saliency map are normalized and mapped to the [0,1] interval to obtain normalized saliency values. The normalization calculation formula is as follows: in, The significance values ​​of the normalized target spatiotemporal saliency map are... and These are the minimum and maximum saliency values ​​in the spatiotemporal saliency map of the target in the current frame, respectively. If Then let After normalization, the saliency values ​​of all pixels in the target spatiotemporal saliency map are uniformly mapped to the [0,1] interval, which facilitates setting consistent quantization parameter adjustment rules across frames.

[0105] S503. Divide the target spatiotemporal saliency map into multiple coding tree units, and determine the average value of the normalized saliency values ​​of all pixels in each coding tree unit.

[0106] The target spatiotemporal saliency map is divided into multiple coding tree units according to the coding unit size defined by the coding standard. For each coding tree unit, the normalized saliency value of all pixels within its coverage area is calculated. average .

[0107] S504. Determine the adaptive quantization parameters of the coding tree unit based on the global baseline quantization parameters, the coding adjustment coefficients, and the mean normalized significance value of the coding tree unit.

[0108] Get the global baseline quantization parameters set for the current video frame. and preset coding adjustment coefficients (For example Adaptive quantization parameters for each coding tree unit. The calculation formula is as follows: To ensure Within the valid range (e.g., video coding standards typically define) The value ranges from 0 to 51. If the calculated value is... Then let ;like Then let The coding tree unit with the larger the mean of the normalized significance value, The smaller the value, the lower the compression level and the more complete the image quality is preserved.

[0109] S505. Based on the adaptive quantization parameters, encode the coding tree units corresponding to the highly significant core region, the moderately significant transition region, and the lowly significant background region.

[0110] Based on the adaptive quantization parameters of each coding tree unit calculated Encoding processing is performed separately for encoding tree units belonging to highly saliency core regions, moderately saliency transition regions, and low saliency background regions. Encoding tree units corresponding to highly saliency core regions are given smaller quantization parameters to preserve rich texture and edge details; encoding tree units corresponding to moderately saliency transition regions are given moderate quantization parameters; and encoding tree units corresponding to low saliency background regions are given baseline quantization parameters to discard texture information that is not sensitive to the human eye.

[0111] S506. During inter-frame coding, motion offset information between the current video frame and historical video frames is reused to match the positions of salient regions between consecutive video frames, and to ensure that the coding regions where temporally consecutive salient targets are located have consistent coding parameters.

[0112] During inter-frame predictive coding, the optical flow motion offset information calculated in steps S102 and S205 is reused to match and track salient regions between consecutive frames. This strategy ensures that moving foreground targets maintain consistent coding parameters in the temporal domain, avoiding quantization parameter jumps in salient targets between different frames due to inter-frame fluctuations in region division, thereby eliminating inter-frame image quality jitter or blurring. For low-salient background regions, a fast inter-frame matching algorithm can be used to reduce coding computation overhead.

[0113] S507. After completing the regional differential coding, dynamically adjust the global baseline quantization parameters according to the pixel ratio of the highly significant core region, and encode subsequent video frames based on the adjusted global baseline quantization parameters to generate the target compressed bitstream.

[0114] Before performing the final entropy encoding to generate the compressed bitstream, a region occupancy factor is set. This is used to represent the number of key targets in the current frame. The calculation formula is as follows: in, This represents the total number of pixels in the current frame that are divided into highly salient core regions. and These are the height and width (in pixels) of the current frame, respectively.

[0115] Set a state-holding counter Cnt, with an initial value of 0. When Entering a certain adjustment range ( or If Cnt=0, then the corresponding action is executed. Adjust Cnt and set it to M (e.g., M=5); if Cnt is greater than 0, simply decrement Cnt by 1 without performing a new adjustment. When Cnt reaches 0, if... If still within the current adjustment range, perform the adjustment again and reset Cnt. Falling back to the middle range ( If Cnt is zeroed out, then Cnt will be cleared. This mechanism ensures that adjustments are performed only once every M frames within the same adjustment interval, avoiding... Frequent jumps.

[0116] according to Dynamically adjust the global baseline quantization parameters of video frames In this embodiment of the application, the first preset percentage threshold is used as... The second preset percentage threshold is For example, satisfying For example, in the embodiments of this application, , .

[0117] The specific adjustment rules are as follows: like ,but This is used to reduce the baseline quantization parameter (i.e., reduce the compression intensity) when targets are densely packed, thus ensuring image quality for multiple targets. This is the function for finding the maximum value. The value of must satisfy In this embodiment .

[0118] like ,but It is used to improve the baseline quantization parameters (i.e., increase the compression strength) when the background is dominant. This is a function that takes the minimum value.

[0119] like ,but .

[0120] Finally, based on the global benchmark quantization parameters Encode subsequent video frames to generate the target compressed bitstream.

[0121] Please see Figure 2 The diagram illustrates a system architecture of a video region coding system integrating saliency detection according to an embodiment of the present invention. This system includes: a saliency map generation module 21, a reliability determination module 22, a spatiotemporal fusion module 23, a region division module 24, and a coding execution module 25. The units communicate bidirectionally via communication links to ensure real-time interaction of collected data and analysis results. The communication links can employ wired or wireless transmission methods to meet the communication needs of different application scenarios.

[0122] The saliency map generation module 21 is used to generate a basic saliency map of the current video frame based on the spatial texture contrast features, temporal motion amplitude features, and chromaticity features of the current video frame.

[0123] The reliability determination module 22 is used to determine the motion reliability coefficient of each pixel in the current video frame based on the local statistical characteristics of the motion offset information between the current video frame and the historical video frames.

[0124] The spatiotemporal fusion module 23 is used to fuse the basic saliency map with the historical frame cumulative saliency map according to the motion reliability coefficient to generate the target spatiotemporal saliency map of the current video frame. The historical frame cumulative saliency map is obtained by accumulating the basic saliency maps of multiple consecutive historical video frames through motion compensation and temporal decay.

[0125] The region segmentation module 24 is used to divide the current video frame into different levels of coding regions based on the target spatiotemporal saliency map.

[0126] The encoding execution module 25 is used to perform regional encoding on the current video frame according to the differentiated encoding parameters corresponding to different levels of encoding regions.

Claims

1. A video region coding method incorporating saliency detection, characterized in that, The method includes: Based on the spatial texture contrast features, temporal motion amplitude features, and chromaticity features of the current video frame, a basic saliency map of the current video frame is generated. Based on the local statistical characteristics of the motion offset information between the current video frame and historical video frames, the motion reliability coefficient of each pixel in the current video frame is determined. Based on the motion reliability coefficient, the base saliency map and the historical frame cumulative saliency map are fused to generate the target spatiotemporal saliency map of the current video frame, wherein the historical frame cumulative saliency map is obtained by accumulating the base saliency maps of multiple consecutive historical video frames through motion compensation and temporal decay. Based on the target spatiotemporal saliency map, the current video frame is divided into different levels of coding regions; Based on the differentiated coding parameters corresponding to the different levels of coding regions, the current video frame is subjected to regional coding. The step of generating the basic saliency map of the current video frame based on its spatial texture contrast features, temporal motion amplitude features, and chromaticity features includes: Centered on the target pixel, the spatial texture contrast feature is determined based on the ratio of the gray-level variance of the pixels in the neighborhood of the target pixel to the gray-level variance of the global image. Calculate the pixel-level motion offset between the current video frame and the adjacent video frames, and determine the temporal motion amplitude characteristics based on the magnitude of the pixel-level motion offset; The current video frame is converted to a color space that separates luminance and chrominance. The chrominance features are determined based on the degree of deviation of the chrominance component of each pixel from the average chrominance component of the entire frame. The spatial texture contrast feature, the temporal motion amplitude feature, and the chromaticity feature are normalized and then added pixel by pixel to obtain the basic saliency map. The determination of the motion reliability coefficient for each pixel in the current video frame based on the local statistical characteristics of the motion offset information between the current video frame and historical video frames includes: Centered on the target pixel, a local motion statistics window is constructed based on the motion offset information between the current video frame and historical video frames; Calculate the standard deviation and average value of the magnitude of the motion offset information of all pixels within the local motion statistics window; The motion reliability coefficient is determined based on the ratio of the standard deviation to the mean. The process of generating the historical frame cumulative saliency map includes: Obtain the historical saliency map corresponding to each of multiple consecutive historical video frames; Based on the pixel-by-pixel motion offset information from the current video frame to each historical video frame, the historical base saliency map corresponding to each historical video frame is distorted and mapped to obtain the aligned historical saliency map. The aligned historical saliency map is weighted, accumulated, and normalized according to the temporal decay factor to generate the historical frame cumulative saliency map. The temporal decay factor is used to characterize the degree of decay of the contribution of the time distance between the historical video frame and the current video frame to the significance. The process of determining the target spatiotemporal saliency map includes: Based on the motion reliability coefficient, determine the fusion weight of the basic saliency map and the fusion weight of the historical frame cumulative saliency map; Based on the fusion weights of the basic saliency map and the fusion weights of the historical frame cumulative saliency map, the basic saliency map and the historical frame cumulative saliency map are fused pixel by pixel to obtain the target spatiotemporal saliency map.

2. The video region coding method integrating saliency detection according to claim 1, characterized in that, The step of dividing the current video frame into different levels of coding regions based on the target spatiotemporal saliency map includes: The target spatiotemporal saliency map is divided into multiple coding tree units; Determine the average value of pixel saliency within each coding tree unit; The high significance threshold benchmark is determined by multiplying the first preset scaling factor by the average value. The low significance threshold is determined by multiplying the second preset scaling factor by the average value. For each pixel in the current video frame, the high significance threshold of the pixel itself is determined based on the high significance threshold benchmark, the third preset scaling factor, and the motion reliability coefficient corresponding to the pixel. The saliency value of a pixel is compared with the high saliency threshold of the pixel itself and the low saliency threshold of the coding tree unit to which the pixel belongs; The region consisting of pixels whose significance value is greater than or equal to the high significance threshold of the pixel itself is divided into a high significance core region; The region consisting of pixels whose salience values ​​are less than the low salience threshold of the coding tree unit to which the pixel belongs is divided into a low salience background region; The region consisting of pixels whose significance values ​​are greater than or equal to the low significance threshold of the coding tree unit to which the pixel belongs and less than the high significance threshold of the pixel itself is divided into a medium significance transition region. Highly significant threshold regulation factor The calculation formula is: ;in, This is the third preset proportional coefficient. For pixels The motion reliability coefficient; High significance threshold The calculation formula is: In the formula, This indicates a highly significant threshold benchmark.

3. The video region coding method integrating saliency detection according to claim 1, characterized in that, The step of determining the fusion weights of the basic saliency map and the cumulative saliency map of historical frames based on the motion reliability coefficient includes: Obtain a preset control coefficient, and determine the historical frame cumulative saliency map fusion weight based on the control coefficient and the motion reliability coefficient; the sum of the basic saliency map fusion weight and the historical frame cumulative saliency map fusion weight is a preset total weight value.

4. The video region coding method integrating saliency detection according to claim 2, characterized in that, The step of performing region-based coding on the current video frame according to the differentiated coding parameters corresponding to the different levels of coding regions includes: Get the global baseline quantization parameters set for the current video frame; The target spatiotemporal saliency map is divided into multiple coding tree units, and the average value of the normalized saliency value of pixels in each coding tree unit is determined. Based on the global baseline quantization parameters, coding adjustment coefficients, and the average of the normalized significance values ​​within each coding tree unit, the adaptive quantization parameters for the corresponding coding tree unit are determined; the adaptive quantization parameters for each coding tree unit... The calculation formula is as follows: In the formula, This represents the global baseline quantization parameter set for the current video frame. This represents the preset coding adjustment coefficient; Represents the normalized significance value of all pixels within the coverage area of ​​each coding tree unit. The average value; Encoding is performed on the highly significant core region, the moderately significant transition region, and the lowly significant background region according to the adaptive quantization parameters. During inter-frame coding, the motion offset information between the current video frame and the historical video frames is reused to match the positions of salient regions between consecutive video frames, and to ensure that the coding regions where salient targets are located in a temporally consecutive manner have consistent coding parameters. After completing the regional differential encoding, the global baseline quantization parameters are dynamically adjusted according to the pixel ratio of the highly significant core region, and the subsequent video frames are encoded based on the adjusted global baseline quantization parameters to generate the target compressed bitstream.

5. The video region coding method integrating saliency detection according to claim 1, characterized in that, The process of determining the current video frame includes: The acquired raw video data is converted into a color space with a preset encoding format, and then aligned for resolution and frame rate. Perform video frame noise reduction and contrast enhancement processing on the converted video frames; Each video frame is divided into blocks according to the size of the coding tree unit to obtain the preprocessed current video frame.

6. A video region coding system integrating saliency detection, characterized in that, The steps for implementing a video region coding method for fusing saliency detection as described in claim 1 include: The saliency map generation module is used to generate a basic saliency map of the current video frame based on the spatial texture contrast features, temporal motion amplitude features, and chromaticity features of the current video frame. The reliability determination module is used to determine the motion reliability coefficient of each pixel in the current video frame based on the local statistical characteristics of the motion offset information between the current video frame and historical video frames. The spatiotemporal fusion module is used to fuse the basic saliency map and the historical frame cumulative saliency map according to the motion reliability coefficient to generate the target spatiotemporal saliency map of the current video frame, wherein the historical frame cumulative saliency map is obtained by accumulating the basic saliency maps of multiple consecutive historical video frames through motion compensation and temporal decay. The region segmentation module is used to divide the current video frame into different levels of coding regions based on the target spatiotemporal saliency map; The encoding execution module is used to perform regional encoding on the current video frame according to the differentiated encoding parameters corresponding to the different levels of encoding regions.

Images