Film transmission quality control method and system based on AI intelligent analysis

Abstract

The invention provides a film transmission quality control method and system based on AI intelligent analysis, and the method comprises the steps: obtaining a real-time frame sequence data set in a target film transmission link, carrying out the transmission quality feature extraction of the real-time frame sequence data set, and carrying out the transmission quality feature extraction of the real-time frame sequence data set; the method comprises the following steps: generating frame quality evaluation features and frame sequence dynamic features, calling a pre-trained transmission quality analysis model, fusing the features to obtain a transmission quality attenuation prediction result, matching a preset transmission parameter adjustment strategy set according to the transmission quality attenuation prediction result, and generating a dynamic transmission optimization instruction set. And finally, the dynamic transmission optimization instruction set is fed back to an encoder node and a routing node corresponding to the target film transmission link, transmission parameter adaptive adjustment operation is triggered, effective control of film transmission quality is realized, and the transmission effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for controlling the transmission quality of a video based on AI intelligent analysis. Background Art

[0002] In the field of video transmission technology, with the rapid development of applications such as streaming media services, video conferencing, and virtual reality, users' demand for high-definition, low-latency, and high-stability video transmission is increasing day by day. However, the existing video transmission quality control technologies face multiple challenges.

[0003] Specifically, the existing technologies lack a collaborative optimization mechanism for the dynamic interaction of multiple nodes in the transmission link. The parameter adjustments of the encoder node and the routing node are often carried out independently, and it is impossible to achieve cross-node adaptive collaborative control according to the real-time transmission quality changes, thus resulting in problems such as low transmission efficiency and large quality fluctuations. In addition, the transmission quality prediction in the existing technologies is mostly based on simple statistical models or rule engines, and it is difficult to effectively capture the dynamic correlation and non-linear characteristics between frame sequences, resulting in insufficient accuracy of quality attenuation prediction and inability to provide a reliable basis for transmission parameter adjustment. Summary of the Invention

[0004] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for controlling the transmission quality of a video based on AI intelligent analysis, and the method includes:

[0005] Obtain a set of real-time frame sequence data in a target video transmission link, where the set of real-time frame sequence data includes encoding parameters, transmission path node identifiers, and decoding delay metrics of multiple continuously transmitted frames;

[0006] Perform transmission quality feature extraction processing on the set of real-time frame sequence data to generate frame quality evaluation features of each transmission frame and frame sequence dynamic features between adjacent transmission frames;

[0007] Call a pre-trained transmission quality analysis model to perform feature fusion processing on the frame quality evaluation features and the frame sequence dynamic features to generate a transmission quality attenuation prediction result of the target video transmission link;

[0008] Match a preset set of transmission parameter adjustment strategies based on the transmission quality attenuation prediction result to generate a dynamic transmission optimization instruction set;

[0009] Feed back the dynamic transmission optimization instruction set to the encoder node and the routing node corresponding to the target video transmission link to trigger an adaptive adjustment operation of transmission parameters.

[0010] In another aspect, an embodiment of the present invention further provides a film transmission quality control system based on AI intelligent analysis, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiment of the present application obtains multi-modal frame sequence data (including coding parameters, path node identifiers, and decoding delay metrics) in the target film transmission link in real time, and combines a transmission quality feature extraction algorithm to generate frame-level quality evaluation features and inter-frame dynamic features. Further, a pre-trained transmission quality analysis model is used to deeply fuse multi-dimensional features to accurately predict the quality attenuation trend of the transmission link. On this basis, an adaptive optimization instruction set is generated by dynamically matching a preset policy set and is fed back to the encoder and the routing node in real time, forming a cross-node parameter collaborative adjustment mechanism, effectively solving the problems of lagging transmission parameter adjustment and lack of global optimization in traditional methods. Thereby, not only the accuracy and real-time performance of transmission link quality prediction are improved, but also the dynamic optimization of film transmission quality in a complex network environment is realized through end-to-end intelligent closed-loop control, significantly enhancing the robustness and adaptive ability of the transmission system. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a schematic execution flow diagram of a film transmission quality control method based on AI intelligent analysis provided by an embodiment of the present invention.

[0013] Figure 2 is a schematic diagram of exemplary hardware and software components of a film transmission quality control system based on AI intelligent analysis provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 is a schematic flow diagram of a film transmission quality control method based on AI intelligent analysis provided by an embodiment of the present invention. The film transmission quality control method based on AI intelligent analysis will be introduced in detail below.

[0015] Step S110, obtain a real-time frame sequence data set in the target film transmission link, where the real-time frame sequence data set includes coding parameters, transmission path node identifiers, and decoding delay metrics of multiple continuously transmitted frames.

[0016] In this embodiment, a link that is transmitting a 3D animated film is used as an application scenario for illustration. Specifically, for the encoding parameters therein, taking a certain transmission frame as an example, the encoding format is AVC (Advanced Video Coding), and its encoding bit rate is 5 Mbps. This encoding bit rate determines the size of the transmitted data volume. The image resolution is set to 1920x1080 pixels, which is a common resolution for high-definition films. The color mode is YUV4:2:0, and this color mode can effectively compress data while ensuring a certain image quality.

[0017] In terms of the transmission path node identification, assume that the entire transmission link passes through 8 network devices from the film source to the receiving end. Then they can be respectively identified as nodes numbered 1-8, thus forming the transmission path node identification.

[0018] For a specific frame, the decoding delay index may be a time parameter generated by the combined influence of the receiving end device performance and the network condition. For example, the measured decoding delay is 80 milliseconds, which can be understood as the decoding delay index. On this basis, multiple consecutive transmission frames (such as 200 consecutive frames) can be continuously monitored, so as to collect the encoding parameters, transmission path node identification, and decoding delay index of each frame, and form the real-time frame sequence data set.

[0019] Step S120, perform transmission quality feature extraction processing on the real-time frame sequence data set to generate the frame quality evaluation feature of each transmission frame and the frame sequence dynamic feature between adjacent transmission frames.

[0020] In this embodiment, for each transmission frame in the above-mentioned 3D animated film transmission, the encoding parameters are first subjected to block processing. For example, since the encoding parameters of a certain frame contain multiple parameter parts with different functions and properties, they can be divided into 5 encoding parameter sub-block sets. Taking one of the encoding parameter sub-block sets related to image texture encoding as an example, dynamic feature extraction operations can be performed on it. For example, if there are obvious changes in some parameters of the texture encoding between different sub-blocks within this encoding parameter sub-block set, such as the quantization step of the texture encoding changes from 10 to 12 between adjacent sub-blocks, thus forming the inter-block quality fluctuation feature. And the intra-block encoding distortion feature may be manifested as that within this sub-block, the encoding of some texture details may have certain losses, resulting in an unclear display effect of the image in a specific texture area. Thus, the above-mentioned inter-block quality fluctuation feature and intra-block encoding distortion feature of the same transmission frame can be aggregated to obtain the frame quality evaluation feature of this transmission frame.

[0021] Next, consider the situation between adjacent transmission frames: First, extract the difference features of the decoding delay metrics between adjacent transmission frames. For example, if the decoding delay of a certain frame is 80 milliseconds and the decoding delay of the next adjacent frame becomes 90 milliseconds, then the difference feature of the decoding delay metrics between them is 10 milliseconds. Then, perform path delay correlation analysis in combination with the transmission path node identifiers. If it is found that when the transmission passes through the node marked as No. 3, the decoding delay difference of adjacent frames increases significantly, it indicates that this node may affect the stability of the transmission. Thus, the inter-frame decoding delay correlation metric can be obtained. Suppose this metric is 0.7 after calculation and statistical analysis, indicating a strong correlation between the inter-frame decoding delays. At the same time, regarding the correlation degree of path node load fluctuations, if during the transmission process, when passing through a certain node (such as node No. 5), a large fluctuation in the transmission quality of adjacent frames is detected, and the load of this node also rapidly rises from 40% utilization to 70% during this period, then the correlation degree of path node load fluctuations may be 0.8, indicating a high correlation degree. On this basis, the above-mentioned inter-frame decoding delay correlation metric and the correlation degree of path node load fluctuations together constitute the dynamic features of the frame sequence.

[0022] Step S130, call the pre-trained transmission quality analysis model to perform feature fusion processing on the frame quality evaluation features and the dynamic features of the frame sequence, and generate a prediction result of the transmission quality attenuation of the target video transmission link.

[0023] In this embodiment, the generated frame quality evaluation features can be input into the first feature encoding layer of the transmission quality analysis model. For example, after the inter-block quality fluctuation feature and the intra-block coding distortion feature in the frame quality evaluation features are processed by the first feature encoding layer, a new feature vector can be generated, which may contain more comprehensive and abstract information about the frame quality in different dimensions (such as coding accuracy, image detail retention, etc.), that is, the first-order high-quality feature.

[0024] Then, input the inter-frame decoding delay correlation metric and the correlation degree of path node load fluctuations into the second feature encoding layer of the transmission quality analysis model. Suppose the inter-frame decoding delay correlation metric is 0.7 and the correlation degree of path node load fluctuations is 0.8. After being processed by the second feature encoding layer, a feature vector reflecting multi-faceted information such as the inter-frame dynamic relationship and the impact of node load on the transmission quality can be generated, that is, the second-order high-order dynamic feature.

[0025] Next, cross-modal attention weighting processing can be performed on the first high-order quality feature and the second high-order dynamic feature. For example, assume that in this transmission quality analysis model, through a large amount of training data, the weight of the first high-order quality feature in the overall quality assessment is determined to be 0.55, and the weight of the second high-order dynamic feature in reflecting the transmission dynamic situation is 0.45. After weighting processing, a fused quality-dynamic feature vector is generated.

[0026] Then, similarity matching can be performed according to the fused quality-dynamic feature vector in a preset attenuation mode library. Assume that there are various predefined quality attenuation modes in the attenuation mode library. For example, mode A indicates that due to some unreasonable parameter settings during the encoding process, the transmission quality gradually and slowly decreases; mode B indicates that due to congestion at a key node in the transmission path, the transmission quality rapidly decreases with fluctuations; mode C indicates that there may be a performance bottleneck in the receiver decoder, resulting in intermittent decreases in transmission quality, etc. If the fused quality-dynamic feature vector has the highest similarity with mode B, then it is determined that the quality attenuation mode category corresponding to the target video transmission link is mode B.

[0027] Finally, a preset attenuation prediction parameter table can be queried based on this quality attenuation mode category. In this attenuation prediction parameter table, for the transmission quality problem caused by path node congestion in mode B, its corresponding transmission quality attenuation prediction result may include multiple key attenuation factors. For example, the encoding complexity exceeding the standard flag may be "yes" because node congestion causes data backlog, making the encoding processing complexity exceed the normal range; the path node congestion level may be "high", indicating a relatively serious congestion situation; the decoding buffer overflow probability may increase to 40% due to data backlog and transmission delay at the congested node. With this design, the above key attenuation factors together constitute the transmission quality attenuation prediction result.

[0028] Step S140: Match a preset set of transmission parameter adjustment strategies based on the transmission quality attenuation prediction result to generate a dynamic transmission optimization instruction set.

[0029] In this embodiment, the set of key attenuation factors in the previously obtained transmission quality attenuation prediction result can be parsed. In this example, the set of key attenuation factors includes the encoding complexity exceeding the standard flag being "yes", the path node congestion level being "high", and the decoding buffer overflow probability being 40%.

[0030] Query the preset encoding parameter adjustment rule library according to the encoding complexity exceeding the standard flag. Since the encoding complexity exceeds the standard, after querying the encoding parameter adjustment rule library, it is found that the encoding complexity can be reduced by adjusting the quantization parameter and the number of reference frames of the encoding. For example, the original quantization parameter was 30, which is now adjusted to 25, and at the same time, the number of reference frames is reduced from 3 to 2, thereby generating the first optimized instruction subset.

[0031] Call the dynamic routing decision model according to the path node congestion level to generate an alternative path switching instruction subset. Because the path node congestion level is "high", a better transmission path needs to be found. Obtain the real-time topology status data of the current transmission path nodes. Assume that the available bandwidth of the node is 8 Mbps, the number of hops of adjacent nodes is 4, and the historical transmission packet loss rate is 8%. Thus, the above data can be input into the feature extraction layer of the dynamic routing decision model to generate a node status feature vector.

[0032] Call the multi-path decision layer of the dynamic routing decision model to calculate the path score for the node status feature vector. Perform a convolution operation on the node status feature vector to extract the spatial association features between path nodes. For example, through the convolution operation, it is found that there is an uneven bandwidth allocation in space for the connection between some adjacent nodes. On this basis, the spatial association feature can be temporally aligned with the historical path switching record to generate a path stability temporal feature. Call the pre-trained stability prediction sub-model to perform regression processing on the path stability temporal feature and output the transmission stability score. Assume that the transmission stability score is 0.5. Then, based on the available bandwidth of the node and the number of hops of adjacent nodes, construct a delay estimation equation to calculate the delay guarantee score. For example, the calculated delay guarantee score is 0.6. Standardize these two scores so that they fall into the same numerical range between 0 and 1. Then, according to the preset transmission stability weight coefficient of 0.45 and the delay guarantee weight coefficient of 0.55, perform weighted aggregation on the standardized transmission stability score and the delay guarantee score to generate the comprehensive optimization score of each alternative path.

[0033] Assume that there are four alternative paths, and the calculated comprehensive optimization scores are 0.65, 0.7, 0.55, and 0.6 respectively. Then, based on the comprehensive optimization score, sort the alternative paths in descending order to generate an alternative path priority sequence, obtain the identifier of the alternative path with the highest priority, and match and verify the identifier of the alternative path with the preset path switching conditions. If the verification passes, for example, the preset condition is that the comprehensive optimization score is greater than 0.6 to be able to switch, then output it as the optimal alternative path identifier; otherwise, trigger the operation of regenerating the alternative path priority sequence. Finally, generate an alternative path switching instruction subset that includes the optimal alternative path identifier and the switching time window (such as switching within the next 3 minutes).

[0034] Adjust the buffer threshold parameter of the target decoder according to the decoding buffer overflow probability. Since the decoding buffer overflow probability is 40%, in order to reduce this decoding buffer overflow probability, the buffer threshold parameter of the target decoder can be adjusted from the original 80MB to 100MB to generate a subset of buffer control instructions.

[0035] Finally, aggregate the first optimized instruction subset, the alternative path switching instruction subset, and the buffer control instruction subset together to generate the dynamic transmission optimization instruction set.

[0036] Step S150, feedback the dynamic transmission optimization instruction set to the encoder node and the routing node corresponding to the target video transmission link to trigger the transmission parameter adaptive adjustment operation.

[0037] In this embodiment, the encoding parameter adjustment instructions in the dynamic transmission optimization instruction set can be parsed to generate the updated value of the quantization parameter and the adjusted value of the reference frame number of the encoder node. As mentioned before, the quantization parameter is updated from 30 to 25, and the reference frame number is reduced from 3 to 2. The control parameter table of the encoder is reconstructed according to the above values.

[0038] Then, the path switching instructions in the dynamic transmission optimization instruction set can be parsed to update the next-hop address mapping table of the routing node. Suppose the original next-hop address is node X, and now it is updated to node Y according to the path switching instruction. Send a reconfiguration command containing the control parameter table to the encoder node, and send a path update command containing the next-hop address mapping table to the routing node.

[0039] Next, status query requests can be sent to the encoder node and the routing node respectively to obtain the currently effective quantization parameter value and the next-hop address. If the currently effective quantization parameter value of the encoder node is 30, while the target value in the control parameter table is 25, compare the two for differences and find that they are inconsistent, generating an encoder synchronization status flag of unsynchronized. For the routing node, if the currently effective next-hop address is node X, while the target address in the path update command is node Y, after performing a consistency check and finding that they are inconsistent, generate a routing synchronization status flag of unsynchronized.

[0040] When the encoder synchronization status flag or the routing synchronization status flag is unsynchronized, increment the retry counter and resend the corresponding command. Suppose the initial value of the retry counter is 0, and it increments by 1 every time an unsynchronization is found. If the retry counter exceeds the preset threshold, for example, the preset threshold is 3, then an alarm log is generated and a manual intervention process is triggered. There may be professional technicians to check where the problem is and make manual adjustments.

[0041] Based on the above steps, in the embodiments of this application, by obtaining multi-modal frame sequence data (including encoding parameters, path node identifiers, and decoding delay metrics) in the target video transmission link in real time, and combining with the transmission quality feature extraction algorithm, frame-level quality evaluation features and inter-frame dynamic features are generated. Further, a pre-trained transmission quality analysis model is used to deeply fuse the multi-dimensional features to accurately predict the quality attenuation trend of the transmission link. On this basis, an adaptive optimization instruction set is generated by dynamically matching a preset policy set and is fed back to the encoder and routing nodes in real time, forming a cross-node parameter collaborative adjustment mechanism, effectively solving the problems of lagging transmission parameter adjustment and lack of global optimization in traditional methods. Thereby, not only the accuracy and real-time performance of transmission link quality prediction are improved, but also through end-to-end intelligent closed-loop control, the dynamic optimization of video transmission quality in a complex network environment is realized, significantly enhancing the robustness and adaptive ability of the transmission system.

[0042] In a possible implementation manner, step S120 includes:

[0043] Step S121, perform block processing on the encoding parameters of each transmission frame to obtain a plurality of encoding parameter sub-block sets.

[0044] For the transmission frames in the 3D animation video link being transmitted, taking a certain transmission frame as an example, its encoding parameters cover information in multiple aspects. For example, in the parts related to encoding format, image resolution, color mode, and frame rate control. In this embodiment, the above-mentioned encoding parameter parts with different functions and properties can be block-processed to obtain a plurality of encoding parameter sub-block sets. For example, the parameters related to the encoding format (such as some specific parameter settings in AVC encoding) are used as one sub-block set, and the image resolution-related parameters (such as the specific value of 1920x1080 pixels and the related pixel encoding method) are used as another sub-block set, etc.

[0045] Step S122, perform dynamic feature extraction operations on the encoding parameter sub-block sets to generate the inter-block quality fluctuation features and intra-block encoding distortion features of each transmission frame.

[0046] For example, in the set of sub-blocks of coding parameters related to image resolution, if there are inconsistent adjustment methods for resolution-related parameters in different parts within this set of coding parameter sub-blocks, for example, the horizontal resolution coding method of the image changes in some sub-blocks, from one coding method to another, thus generating an inter-block quality fluctuation feature. In terms of the intra-block coding distortion feature, within this set of coding parameter sub-blocks, due to certain limitations of the coding algorithm in processing resolution-related data, the coding of some pixels may not be precise enough, resulting in slight blurring or distortion of the image in specific areas, thereby constituting the intra-block coding distortion feature, which is used to represent the transmission frame.

[0047] Step S123, aggregate the inter-block quality fluctuation feature and the intra-block coding distortion feature of the same transmission frame to generate the frame quality evaluation feature of the transmission frame.

[0048] For example, the inter-block quality fluctuation feature indicates that there are certain fluctuations in the resolution-related parameters of the image among sub-blocks, while the intra-block coding distortion feature shows that there are inaccurate pixel codings in some areas. By comprehensively considering these two features, a feature that can comprehensively evaluate the quality of the transmission frame is formed, that is, the frame quality evaluation feature. This frame quality evaluation feature can reflect the quality status of the transmission frame from multiple perspectives, including aspects such as the stability and clarity of the image.

[0049] Step S124, extract the difference feature of the decoding delay index between adjacent transmission frames, and perform path delay correlation analysis in combination with the transmission path node identifier to generate the dynamic feature of the frame sequence. Among them, the dynamic feature of the frame sequence includes the inter-frame decoding delay correlation index and the path node load fluctuation correlation degree.

[0050] For example, in a possible implementation manner, step S124 includes:

[0051] Step S1241, perform sliding window statistics on the decoding delay indexes of N consecutive transmission frames to generate a delay mean feature and a delay variance feature.

[0052] For example, assume that N is taken as 10. Then, a sliding window statistic can be performed on the decoding delay metrics of these 10 consecutive transmission frames. For instance, the decoding delays of these 10 transmission frames are 80 milliseconds, 85 milliseconds, 90 milliseconds, 82 milliseconds, 88 milliseconds, 92 milliseconds, 84 milliseconds, 86 milliseconds, 90 milliseconds, and 83 milliseconds respectively. Calculate the mean delay feature. Add these 10 values together, and the sum is 860 milliseconds. Then divide by 10 to obtain the mean delay feature of 86 milliseconds. When calculating the variance delay feature, first calculate the square of the difference between each value and the mean. For example, the difference between the first value of 80 milliseconds and the mean of 86 milliseconds is -6 milliseconds, and its square is 36. Calculate the squares of the 10 differences in this way, add them together, the sum is 236, and then divide by 10 to obtain the variance delay feature of 23.6.

[0053] Step S1242: Match the historical path delay record set according to the transmission path node identifier, and extract the load fluctuation feature of the current path node within a preset time period.

[0054] For example, assume that the preset time period is the past 5 minutes. According to the transmission path node identifier (such as the 1-8th nodes mentioned above), search for relevant records passing through a certain node (such as the 3rd node) in the historical path delay record set. If within these 5 minutes, the load of the 3rd node fluctuates from 30% to 50%, this is the load fluctuation feature of the 3rd node within the preset time period.

[0055] Step S1243: Perform time alignment processing on the mean delay feature, the variance delay feature, and the load fluctuation feature to generate a delay-load joint analysis matrix.

[0056] Assume that the mean delay feature of 86 milliseconds, the variance delay feature of 23.6, and the load fluctuation feature of the 3rd node (from 30% to 50%) are time-aligned, and a matrix is constructed. The rows of the matrix can represent different features (such as the first row is the mean delay feature, the second row is the variance delay feature, and the third row is the load fluctuation feature), and the columns can represent time or serial numbers, etc.

[0057] Step S1244: Invoke the pre-trained delay correlation model to perform feature mapping on the delay-load joint analysis matrix, and output the inter-frame decoding delay correlation metric.

[0058] In this embodiment, the pre-trained delay correlation model is trained based on a large amount of historical data. It can analyze the relationship between the inter-frame decoding delay and factors such as load fluctuation according to the input delay-load joint analysis matrix, and thus output the inter-frame decoding delay correlation metric. Assume that after the analysis and calculation of the model, the output inter-frame decoding delay correlation metric is 0.7, indicating a strong correlation between the inter-frame decoding delay and factors such as load fluctuation.

[0059] Step S1245, construct a path topology graph based on the transmission path node identifiers, and generate the path node load fluctuation correlation degree according to the real-time load data of each node in the path topology graph.

[0060] For example, a path topology graph can be constructed according to the previously mentioned nodes 1-8. In this path topology graph, the connection relationship between nodes represents the transmission path, and each node has its corresponding real-time load data. For example, the real-time load of node 1 is 20%, the real-time load of node 2 is 30%, the real-time load of node 3 is 50%, etc. By analyzing the real-time load data of the above nodes and their positional relationships in the topology graph, the path node load fluctuation correlation degree can be generated. Suppose after a series of analyses and calculations (such as considering factors such as the change amplitude of node load and the impact on adjacent nodes), the path node load fluctuation correlation degree is 0.8, indicating a strong correlation relationship between path node load fluctuations. Thus, the above inter-frame decoding delay correlation index and the path node load fluctuation correlation degree together constitute the frame sequence dynamic characteristics, which can reflect the dynamic relationship between adjacent transmission frames in terms of decoding delay and path node load.

[0061] For example, in a possible implementation manner, step S1244 includes:

[0062] Step S1244-1, input the delay-load joint analysis matrix into the first convolutional layer of the delay correlation model for spatio-temporal feature extraction to generate a primary spatio-temporal feature matrix.

[0063] In this embodiment, the delay-load joint analysis matrix contains information such as the previously calculated delay mean feature, delay variance feature, and load fluctuation feature. The first convolutional layer contains multiple convolutional kernels, and the above convolutional kernels slide on the matrix for convolutional operations. For example, assume that the size of the convolutional kernel is 3x3 and the stride is 1, and it slides sequentially on the rows and columns of the matrix. For each 3x3 sub-region in the matrix, the convolutional kernel performs multiplication operations on the corresponding elements and then sums them. This process is carried out on the entire matrix to extract the primary spatio-temporal features, which reflect the local relationships of delay-related features at different time points and different load conditions, and finally generate a primary spatio-temporal feature matrix.

[0064] Step S1244-2, perform channel attention weight assignment on the primary spatio-temporal feature matrix to generate a channel attention weighted primary spatio-temporal feature matrix.

[0065] Among them, step S1244-2 includes:

[0066] Step S1244-21 performs a global average pooling operation on the primary spatio-temporal feature matrix in the channel dimension to generate a channel statistical feature vector.

[0067] Step S1244-22 inputs the channel statistical feature vector into the attention weight generation layer of the time-delay correlation model for non-linear transformation to generate a channel attention weight vector.

[0068] Step S1244-23 performs a per-channel multiplication operation on the channel attention weight vector and the primary spatio-temporal feature matrix to generate a primary spatio-temporal feature matrix weighted by channel attention.

[0069] Assume that the primary spatio-temporal feature matrix has 3 channels. Calculate the average value of all elements on each channel. For example, the sum of all elements in the first channel is 100 and the number of elements is 20, then the average value of this channel is 5. Calculate the average values of the three channels in this way to form a channel statistical feature vector. Then, input this channel statistical feature vector into the attention weight generation layer of the time-delay correlation model for non-linear transformation to generate a channel attention weight vector. It should be noted that this non-linear transformation may be implemented through structures such as a multi-layer perceptron. After calculation, the weight value corresponding to each channel is obtained. Assume that the generated weight vector is [0.3, 0.5, 0.2]. Then perform a per-channel multiplication operation on the channel attention weight vector and the primary spatio-temporal feature matrix. Multiply each element in the first channel of the primary spatio-temporal feature matrix by 0.3, the second channel by 0.5, and the third channel by 0.2. In this way, a primary spatio-temporal feature matrix weighted by channel attention is generated. Through this method, the model can pay more attention to the channel features that have an important impact on the time-delay correlation.

[0070] Step S1244-3 inputs the primary spatio-temporal feature matrix weighted by channel attention into the second convolutional layer of the time-delay correlation model for feature abstraction to generate a high-level spatio-temporal feature matrix.

[0071] The second convolutional layer also contains convolutional kernels. Its convolution operation is similar to that of the first convolutional layer, but the parameters and purposes of the convolutional kernels may be different. This convolutional layer can further extract more abstract and higher-level features. By performing a convolution operation on the primary spatio-temporal feature matrix weighted by channel attention, local features are combined to form more representative high-level spatio-temporal features, which can better reflect the relationship between time-delay and load at different spatio-temporal scales, and finally generate a high-level spatio-temporal feature matrix.

[0072] Step S1244-4 performs a sliding window pooling operation on the high-level spatio-temporal feature matrix in the time dimension to generate a time aggregation feature vector.

[0073] Among them, step S1244-4 includes:

[0074] Step S1244-41 divides multiple overlapping sliding windows in the time dimension of the high-level spatio-temporal feature matrix.

[0075] Step S1244-42 performs a max pooling operation on the eigenvalues within each sliding window to generate a pooled feature sub-vector for each sliding window.

[0076] Step S1244-43 concatenates the pooled feature sub-vectors of all sliding windows in chronological order to generate the time aggregation feature vector.

[0077] Assume that the length of the time dimension of the high-level spatio-temporal feature matrix is 10. In this embodiment, 3 sliding windows can be divided, each window has a size of 4, and adjacent windows overlap by 2 elements. A max pooling operation is performed on the eigenvalues within each sliding window to generate a pooled feature sub-vector for each sliding window. For example, if the eigenvalues within the first sliding window are [2, 5, 3, 4], then the result after max pooling is 5. In this way, the pooled feature sub-vectors of each sliding window are obtained. The pooled feature sub-vectors of all sliding windows are concatenated in chronological order. Assume that the pooled feature sub-vectors of the three sliding windows are 5, 7, and 6 respectively. The generated time aggregation feature vector after concatenation is [5, 7, 6].

[0078] Step S1244-5 inputs the time aggregation feature vector into the fully connected classification layer of the time delay correlation model to predict the correlation probability, and generates the time delay correlation probability distribution for each time window.

[0079] Among them, step S1244-5 includes:

[0080] Step S1244-51 performs a matrix multiplication operation on the time aggregation feature vector and the weight matrix of the fully connected classification layer to generate an initial classification score vector.

[0081] Step S1244-52 performs a sigmoid function transformation on the initial classification score vector to generate a normalized time delay correlation probability distribution.

[0082] Assume that the time aggregation feature vector is [5, 7, 6], and the weight matrix of the fully connected classification layer is a 3x2 matrix with elements [[0.1, 0.2], [0.3, 0.4], [0.5, 0.6]] respectively. According to the matrix multiplication rule, the initial classification score vector is calculated as [5×0.1 + 7×0.3 + 6×0.5, 5×0.2 + 7×0.4 + 6×0.6] = [6.6, 8.2]. Then, the initial classification score vector is subjected to a softmax function transformation to generate a normalized time delay correlation probability distribution. The softmax function transformation process is as follows: First, calculate the exponential value of each score, that is, e to the power of 6.6 and e to the power of 8.2. Assume that e to the power of 6.6 is 736.99 and e to the power of 8.2 is 3640.95. Then, divide these two values by their sum (736.99 + 3640.95 = 4377.94) respectively to obtain the normalized time delay correlation probability distribution as [736.99÷4377.94, 3640.95÷4377.94] = [0.168, 0.832], which represents the time delay correlation probabilities of different time windows.

[0083] Step S1244-6: Extract the time window identifier corresponding to the maximum probability value in the time delay correlation probability distribution, and map the time window identifier to a preset time delay correlation level table, and output the inter-frame decoding time delay correlation index.

[0084] For example, in the above example, the maximum probability value is 0.832, and the corresponding time window identifier is assumed to be 2. This time window identifier can be mapped to the preset time delay correlation level table. If the time delay correlation level corresponding to the identifier 2 in the level table is specified as "high", then the output inter-frame decoding time delay correlation index is "high". This inter-frame decoding time delay correlation index can reflect the degree of correlation between the inter-frame decoding time delay and factors such as load fluctuations in the time dimension.

[0085] For example, in a possible implementation manner, step S1245 includes:

[0086] Step S1245-1: Analyze the node connection relationships included in the transmission path node identifier to generate an initial path connection topology.

[0087] For example, in the transmission link of the above 3D animated film, the transmission path node identifiers from 1 to 8 identify each node. By analyzing the connection relationships implied in the above transmission path node identifiers, for example, in this embodiment, it is found that node 1 is connected to node 2, and node 2 is connected to nodes 3 and 4, etc. According to the above connection information, an initial path connection topology is constructed. This initial path topology structure preliminarily depicts how each node is connected, forming a basic framework that can reflect the possible flow paths of data during the film transmission process.

[0088] Step S1245-2: Add bidirectional transmission channel identifiers between each pair of nodes according to the node hierarchy order in the initial path connection topology to generate a complete path topology graph.

[0089] In the initial path connection topology, there is a certain hierarchical relationship among the nodes. For example, Node 1 may be at a relatively upper layer, which is a node near the initial sending end of the video data, and the subsequent node levels gradually extend to the receiving end. According to this hierarchical order, add bidirectional transmission channel identifiers for each pair of connected nodes. This means that data can not only be unidirectionally transmitted from one node to another, but in some cases, there may be reverse transmission or two-way interaction. For example, between Node 2 and Node 3, in addition to the normal transmission direction from 2 to 3, there may also be a transmission possibility from 3 to 2 due to certain network protocols or transmission mechanisms. By adding the above bidirectional transmission channel identifiers, a complete path topology graph is generated, which more comprehensively shows the connection and transmission relationship between the nodes.

[0090] Step S1245-3: Extract the historical load collection interface addresses of each node from the path topology graph, send real-time load query requests to the historical load collection interface addresses, and obtain the real-time load data of each node within the current time window.

[0091] In this embodiment, each node has a corresponding historical load collection interface address in network transmission, and the above historical load collection interface address is used to obtain node load-related data. After finding the above historical load collection interface addresses from the path topology graph, send real-time load query requests to each historical load collection interface address. For example, for Node 3, after its historical load collection interface address is determined, send a query request to obtain the load data within the current time window. Assume that the current time window is set to the most recent 10 minutes. Within these 10 minutes, the load data of Node 3 may include information such as the number of data packets it processes and the proportion of bandwidth occupied. Obtain the real-time load data of Nodes 1 to 8 within these 10 minutes in the same way.

[0092] Step S1245-4: Perform sliding window mean calculation on the real-time load data to generate a load mean sequence for each node, and perform time alignment processing on the load mean sequences of adjacent nodes to obtain the time-aligned load mean sequences.

[0093] Taking Node 3 as an example, assume that it obtains load data every 1 minute within 10 minutes, resulting in 10 load data values. Further, in this embodiment, the sliding window size is set to 3 minutes. Starting from the first data, calculate the average value of the load data within the first sliding window (the 1st, 2nd, and 3rd minutes), then slide the window and calculate the average value of the load data within the second sliding window (the 2nd, 3rd, and 4th minutes), and so on, to generate the load average value sequence of Node 3. Calculate the load average value sequences of Nodes 1 to 8 in the same way. For adjacent nodes, such as Node 2 and Node 3, since there may be slight differences in the time points when they obtain load data, time alignment processing is required. Adjust the load average value sequences of Node 2 and Node 3 in chronological order to ensure comparison and analysis at the same time point, thereby obtaining the load average value sequence after time alignment.

[0094] Step S1245-5: Based on the load average value sequence after time alignment, calculate the load fluctuation covariance matrix for each pair of adjacent nodes, and extract the maximum eigenvalue in the covariance matrix as the strength of the load fluctuation correlation between nodes.

[0095] For example, for an adjacent pair of nodes, such as Node 2 and Node 3, use their load average value sequences after time alignment to calculate the load fluctuation covariance matrix. The calculation process is as follows: Assume that the load average value sequence of Node 2 is [20, 22, 25, 23, 21], and the load average value sequence of Node 3 is [18, 20, 23, 21, 19]. First, calculate the average value of each sequence. The average value of Node 2 is (20 + 22 + 25 + 23 + 21) ÷ 5 = 22, and the average value of Node 3 is (18 + 20 + 23 + 21 + 19) ÷ 5 = 20. Then, calculate the elements of the covariance matrix. For the element (1, 1) in the covariance matrix, calculate (20 - 22) × (20 - 22) + (22 - 22) × (22 - 22) + (25 - 22) × (25 - 22) + (23 - 22) × (23 - 22) + (21 - 22) × (21 - 22) = 10. Calculate the other elements of the covariance matrix in the same way to obtain the covariance matrix. Then, calculate the eigenvalues of this covariance matrix. Assume that the obtained eigenvalues are [3, 7], and the maximum eigenvalue is 7. This 7 is used as the strength of the load fluctuation correlation between Node 2 and Node 3, indicating the degree of correlation between the load fluctuations of these two nodes. Calculate the strength of the load fluctuation correlation for all adjacent node pairs in the path topology diagram in the same way.

[0096] Step S1245-6: Traverse all directly connected node pairs in the path topology diagram, aggregate the strength of the load fluctuation correlation of each node pair, and generate the load fluctuation association degree of the path nodes.

[0097] For example, in a path topology graph, find all directly connected node pairs, such as the node 1 and node 2, node 2 and node 3, node 3 and node 4 mentioned above, etc. Aggregate the load fluctuation correlation strengths of the above-mentioned node pairs. For example, the load fluctuation correlation strength between node 1 and node 2 is 5, that between node 2 and node 3 is 7, and that between node 3 and node 4 is 4, etc. Sum up the above-mentioned values or perform other reasonable aggregation operations (the specific operations can be set according to actual requirements and algorithms). The result obtained is the load fluctuation correlation degree of path nodes, and this load fluctuation correlation degree of path nodes can comprehensively reflect the correlation relationship between node load fluctuations in the entire path topology graph.

[0098] In a possible implementation manner, step S130 includes:

[0099] Step S131, input the frame quality evaluation feature into the first feature encoding layer of the transmission quality analysis model to generate a first high-order quality feature.

[0100] For example, for the frame quality evaluation feature in the transmission of 3D animated films, it contains the comprehensive information obtained by aggregating the inter-block quality fluctuation feature and the intra-block coding distortion feature before. When it is input into the first feature encoding layer of the transmission quality analysis model, this first feature encoding layer extracts higher-level and more abstract features through a series of neural network operations. For example, the first feature encoding layer may contain multiple neurons, which are used to perform linear and non-linear transformations on the input frame quality evaluation feature. Suppose a certain dimension in the frame quality evaluation feature represents a numerical value related to the coding accuracy of the image. After the weighted sum of the neurons and the processing of the activation function (such as the ReLU function), a new numerical value is obtained, and this numerical value reflects the impact of coding accuracy on the overall frame quality at a higher level. By processing each dimension of the frame quality evaluation feature in this way, a first high-order quality feature is finally generated, and this first high-order quality feature can describe the quality status of the frame from a more comprehensive and in-depth perspective.

[0101] Step S132, input the inter-frame decoding delay correlation index and the load fluctuation correlation degree of path nodes into the second feature encoding layer of the transmission quality analysis model to generate a second high-order dynamic feature.

[0102] For example, the inter-frame decoding delay correlation metric (e.g., the previously obtained value is 0.7, indicating a strong correlation between the inter-frame decoding delay and factors such as load fluctuations) and the path node load fluctuation correlation degree (e.g., 0.8, indicating a strong correlation between path node load fluctuations) are fed as inputs into the second feature encoding layer. This feature encoding layer also has multiple neurons, which process these two input features. Taking the inter-frame decoding delay correlation metric as an example, the neuron performs a linear combination operation based on the value of this metric and its own weight, and then performs a non-linear transformation through an activation function. A similar operation is also performed on the path node load fluctuation correlation degree. After the above processing, the information of these two features is fused and more advanced features are extracted, that is, the second higher-order dynamic feature is generated. This second higher-order dynamic feature synthesizes the dynamic information of the inter-frame decoding delay and the path node load fluctuation, and can better reflect the impact of the dynamic relationship in the transmission link on the transmission quality.

[0103] Step S133: Perform cross-modal attention weighting processing on the first higher-order quality feature and the second higher-order dynamic feature to generate a fused quality-dynamic feature vector.

[0104] In a possible implementation manner, step S133 includes:

[0105] Step S1331: Use the first higher-order quality feature as the query vector, and use the second higher-order dynamic feature as the key vector and value vector.

[0106] Step S1332: Calculate the dot product similarity between the query vector and the key vector to generate an initial attention weight distribution.

[0107] Suppose the first higher-order quality feature has 5-dimensional values [1, 2, 3, 4, 5] respectively, and the second higher-order dynamic feature has 5-dimensional values [2, 3, 4, 5, 6] respectively. When calculating the dot product similarity, for each dimension, multiply the value of the query vector by the value of the key vector and then sum. For example, the calculation for the first dimension is 1×2 = 2, the second dimension is 2×3 = 6, and so on, obtaining the dot product result [2, 6, 12, 20, 30], and this dot product result is the initial attention weight distribution.

[0108] Step S1333: Perform non-linear activation processing on the initial attention weight distribution to generate a normalized attention weight.

[0109] The non-linear activation process here may adopt functions such as the Softmax function. For the initial attention weight distribution [2, 6, 12, 20, 30], first calculate the exponential value of each value. For example, the exponential value of 2 is e to the power of 2 (assuming e to the power of 2 is 7.39), the exponential value of 6 is e to the power of 6 (assuming e to the power of 6 is 403.43), and so on. Then divide the above exponential values by their sum (7.39 + 403.43 + … + other calculated exponential values) to obtain the normalized result, assumed to be [0.01, 0.05, 0.12, 0.25, 0.57], which is the normalized attention weight.

[0110] Step S1334, perform weighted summation on the value vector according to the normalized attention weight to generate a dynamic feature enhancement vector.

[0111] For example, perform weighted summation on the normalized attention weight [0.01, 0.05, 0.12, 0.25, 0.57] and the value vector (i.e., the second-order high-level dynamic feature [2, 3, 4, 5, 6]). The calculation process is: 0.01×2 + 0.05×3 + 0.12×4 + 0.25×5 + 0.57×6 = 0.02 + 0.15 + 0.48 + 1.25 + 3.42 = 5.32 (this is just an example of the calculation for one dimension, and actually such calculations are performed for each dimension), and a dynamic feature enhancement vector is obtained through such calculations.

[0112] Step S1335, perform residual connection on the dynamic feature enhancement vector and the first-order high-level quality feature to generate the fused quality-dynamic feature vector.

[0113] Assume the dynamic feature enhancement vector is [5.32, a certain value, a certain value, a certain value, a certain value], and the first-order high-level quality feature is [1, 2, 3, 4, 5]. When performing residual connection, for each dimension, add the value of the dynamic feature enhancement vector to the value of the first-order high-level quality feature. For example, for the first dimension, it is 5.32 + 1 = 6.32. In this way, the fused quality-dynamic feature vector [6.32, the added value of the second dimension, the added value of the third dimension, the added value of the fourth dimension, the added value of the fifth dimension] is obtained. This fused quality-dynamic feature vector combines the information of both frame quality and frame sequence dynamics.

[0114] Step S134, perform similarity matching on the fused quality-dynamic feature vector in a preset attenuation mode library to determine the quality attenuation mode category corresponding to the target video transmission link.

[0115] In this embodiment, the preset attenuation mode library contains a variety of predefined quality attenuation modes. For example, mode A represents a slow and continuous decline in transmission quality due to the gradual deterioration of encoding parameters; mode B represents a sharp decline in transmission quality caused by a sudden high load on a certain key transmission node, etc. Calculate the similarity between the fusion quality-dynamic feature vector and each attenuation mode. The calculation of this similarity may be based on a certain distance metric method, such as the Euclidean distance. Assume that the feature vector corresponding to mode A is [2, 3, 4, 5, 6], and the feature vector corresponding to mode B is [5, 6, 7, 8, 9]. Calculate the Euclidean distance between the fusion quality-dynamic feature vector (for example, [6.32, a certain value, a certain value, a certain value, a certain value]) and the feature vector of mode A. The calculation process is as follows: First, calculate the square of the difference in each dimension. For example, for the first dimension, (6.32 - 2)^2 = 18.66. Then, add up the above squares and take the square root to obtain the distance value from mode A. Calculate the distance value from mode B in the same way. Finally, compare the above distance values and find that the distance value from mode B is the smallest. Then, determine that the quality attenuation mode category corresponding to the target video transmission link is mode B.

[0116] Step S135, query the preset attenuation prediction parameter table based on the quality attenuation mode category, and output the transmission quality attenuation prediction result.

[0117] In this embodiment, in the attenuation prediction parameter table, there are corresponding prediction parameters for mode B. For example, it may include that within the next 5 minutes, the encoding complexity will increase by 20%, the congestion level of nodes in the transmission path will further increase by 30%, and the probability of decoding buffer overflow will reach 50%, etc. The above parameters together constitute the transmission quality attenuation prediction result, which can provide a basis for subsequent transmission parameter adjustment strategies to optimize the transmission link of 3D animated videos.

[0118] In a possible implementation manner, step S140 includes:

[0119] Step S141, analyze the set of key attenuation factors in the transmission quality attenuation prediction result. The set of key attenuation factors includes an encoding complexity overrun flag, a path node congestion level, and a decoding buffer overflow probability.

[0120] In this embodiment, among the previously determined prediction results of transmission quality attenuation, key attenuation factors such as an indication of excessive coding complexity, the congestion level of path nodes, and the probability of decoder buffer overflow are included. Suppose the prediction result of transmission quality attenuation shows that during the next transmission, due to certain factors (such as an increase in the amount of transmitted data, a decrease in the efficiency of the coding algorithm, etc.), the coding complexity has reached the indication of exceeding the standard, the congestion level of path nodes is in the "high" state, meaning that some nodes in the transmission path are severely congested, and at the same time, the probability of decoder buffer overflow has reached 30%, indicating that there is a high possibility of data overflow in the decoder buffer.

[0121] Step S142, query the preset coding parameter adjustment rule library according to the indication of excessive coding complexity, and generate a first subset of optimization instructions.

[0122] In this embodiment, the preset coding parameter adjustment rule library contains adjustment strategies for different coding complexity situations. Since the coding complexity exceeds the standard, when querying this coding parameter adjustment rule library, it may be found that for the coding of the current 3D animation film (for example, using AVC coding), the coding complexity can be reduced by reducing the quantization parameter. For example, the original quantization parameter was 30, and according to the suggestion of the rule library, it is reduced to 25. At the same time, it may also be necessary to adjust the number of reference frames for coding. The original number of reference frames was 3, and it is adjusted to 2. The above adjustment instructions, including the adjustment value of the quantization parameter and the adjustment value of the number of reference frames, together constitute the first subset of optimization instructions. This first subset of optimization instructions aims to reduce the coding complexity by adjusting the coding parameters, thereby alleviating the impact of the coding process on the transmission quality.

[0123] Step S143, call the dynamic routing decision model according to the congestion level of the path nodes to generate a subset of alternative path switching instructions.

[0124] In a possible implementation manner, step S143 includes:

[0125] Step S1431, obtain the real-time topology status data of the current transmission path nodes, and the real-time topology status data includes the available bandwidth of the nodes, the number of hops of adjacent nodes, and the historical transmission packet loss rate.

[0126] Step S1432, input the real-time topology status data into the feature extraction layer of the dynamic routing decision model to generate a node status feature vector.

[0127] In this embodiment, for example, since the congestion level of the path node is "high", it is necessary to find a better transmission path. Assume that the available bandwidth of a certain node (such as node 3) in the current transmission path is 5 Mbps, the number of hops to the adjacent node is 3, and the historical transmission packet loss rate is 10%. The feature extraction layer of the dynamic routing decision model converts the above data into a node state feature vector through a series of mathematical transformations and feature extraction operations. For example, it may perform weighted combination or other transformations on the available bandwidth, the number of hops to the adjacent node, and the historical transmission packet loss rate, and finally generate a vector that can comprehensively represent the node state.

[0128] Step S1433: Invoke the multi-path decision layer of the dynamic routing decision model to calculate the path scores for the node state feature vector, and generate the transmission stability scores and delay guarantee scores for each alternative path.

[0129] In this embodiment, the multi-path decision layer is used to evaluate the performance of different alternative paths. For the calculation of the transmission stability score, multiple factors in the node state feature vector may be considered. For example, a lower available bandwidth may reduce the transmission stability score, while a lower historical transmission packet loss rate will increase the transmission stability score. Assume that through a series of calculations (which may involve model analysis based on historical data, consideration of the network topology structure, etc.), the transmission stability score for a certain alternative path is 0.6. For the delay guarantee score, relevant calculation logic will be constructed based on the node available bandwidth and the number of hops to the adjacent node. For example, if the available bandwidth is low and the number of hops to the adjacent node is large, then the delay guarantee score may be low. Assume that the delay guarantee score for this alternative path is calculated to be 0.5.

[0130] Step S1434: Perform normalization processing on the transmission stability score and the delay guarantee score to generate a normalized transmission stability score and a normalized delay guarantee score.

[0131] In this embodiment, the normalization processing is to unify these two scores with different ranges and meanings into the same numerical interval for subsequent weighted aggregation. Assume that the value range of the transmission stability score is 0 - 1, while the original value range of the delay guarantee score is 0 - 2, and the delay guarantee score is normalized. For example, using the method of linear transformation, if the original delay guarantee score is 0.5 (within the range of 0 - 2), the normalized delay guarantee score is 0.25 (within the range of 0 - 1), and the transmission stability score is normalized in the same way (if its original value is already within the range of 0 - 1, it remains unchanged).

[0132] Step S1435: According to the preset transmission stability weight coefficient and delay guarantee weight coefficient, perform weighted aggregation on the normalized transmission stability score and the normalized delay guarantee score to generate the comprehensive optimization scores of each alternative path.

[0133] Suppose the preset transmission stability weight coefficient is 0.6 and the delay guarantee weight coefficient is 0.4. For a certain alternative path mentioned above, its normalized transmission stability score is 0.6 and its normalized delay guarantee score is 0.25. When calculating the comprehensive optimization score, perform it in the way of weighted summation, that is, 0.6×0.6 + 0.4×0.25 = 0.36 + 0.1 = 0.46. Calculate the comprehensive optimization scores of other alternative paths in the same way.

[0134] Step S1436: Based on the comprehensive optimization scores, perform a descending order sorting on each alternative path to generate an alternative path priority sequence, and extract the alternative path identifier with the highest ranking in the alternative path priority sequence as the reference alternative path identifier. Match and verify the reference alternative path identifier with the preset path switching condition. If the verification passes, output it as the optimal alternative path identifier; otherwise, trigger the operation of regenerating the alternative path priority sequence.

[0135] Suppose there are three alternative paths, and the calculated comprehensive optimization scores are 0.46, 0.4, and 0.35 respectively. After sorting in descending order, an alternative path priority sequence is obtained. Suppose the alternative path identifier with the highest ranking is the identifier of Path A. The preset path switching condition may be that the comprehensive optimization score is greater than 0.45 to allow path switching. Match and verify the identifier of Path A with this path switching condition. It is found that 0.46 is greater than 0.45, and the verification passes. Then, output the identifier of Path A as the optimal alternative path identifier. If the verification fails, for example, the comprehensive optimization score is less than 0.45, it is necessary to recalculate the comprehensive optimization scores of each alternative path, regenerate the alternative path priority sequence and perform the verification again.

[0136] Step S1437: Generate a subset of alternative path switching instructions including the optimal alternative path identifier and the switching time window.

[0137] Suppose the optimal alternative path identifier is the identifier of Path A, and the switching time window is set to the next 3 minutes. Then the subset of alternative path switching instructions includes the identifier of Path A and the information of switching within the next 3 minutes. This subset of alternative path switching instructions will be used to guide the routing node to perform path switching operations to avoid congested nodes and improve the transmission quality of 3D animation films.

[0138] Step S144: Adjust the buffer threshold parameter of the target decoder according to the decoding buffer overflow probability to generate a subset of buffer control instructions.

[0139] Since the probability of decoding buffer overflow has reached 30%, in order to reduce this probability of decoding buffer overflow, it is necessary to adjust the buffer threshold parameter of the target decoder. Assume that the original buffer threshold parameter of the target decoder is 80MB. According to relevant strategies (which may be based on factors such as the performance of the decoder and the rate of transmitted data), the buffer threshold parameter is adjusted to 100MB. This adjusted parameter constitutes a subset of buffer control instructions, whose purpose is to reduce the risk of decoding buffer overflow by increasing the buffer threshold, thereby ensuring the normal decoding and playback of 3D animated films.

[0140] Step S145, aggregate the first optimization instruction subset, the alternative path switching instruction subset, and the buffer control instruction subset to generate the dynamic transmission optimization instruction set.

[0141] In this embodiment, the first optimization instruction subset (including instructions such as adjusting the quantization parameter to 25 and the number of reference frames to 2) generated previously, the alternative path switching instruction subset (including instructions such as the optimal alternative path identifier being path A and a switching time window of 3 minutes), and the buffer control instruction subset (including the instruction to adjust the buffer threshold parameter to 100MB) can be aggregated. These instructions together form a complete dynamic transmission optimization instruction set, which will be sent to relevant nodes (such as encoder nodes and routing nodes) in the 3D animated film transmission link to trigger the adaptive adjustment operation of transmission parameters, thereby improving the transmission quality of the entire transmission link and ensuring the smooth transmission of 3D animated films.

[0142] In a possible implementation manner, step S1433 includes:

[0143] Step S1433-1, perform a convolution operation on the node state feature vector to extract the spatial correlation features between path nodes.

[0144] In this embodiment, in the 3D animated film transmission link, when the node state feature vector is obtained, this node state feature vector contains various information of the current transmission path nodes, such as a comprehensive representation of information such as the available bandwidth of the node, the number of hops of adjacent nodes, and the historical transmission packet loss rate. Thus, a convolution operation can be performed on this node state feature vector. Assume that the node state feature vector is a data structure with multiple dimensions. For example, in one dimension, it represents the available bandwidth of the node (in Mbps), and in another dimension, it represents the number of hops of adjacent nodes, etc. The convolution operation is calculated by sliding a specific convolution kernel on this node state feature vector.

[0145] The size and parameters of the convolution kernel are preset according to the extraction requirements of the spatial correlation features between path nodes. For example, the size of the convolution kernel is 3×3 (this is just a simple assumed size, and the actual situation may be more complex), and it slides on each dimension of the node state feature vector. For each sliding position, the elements in the convolution kernel are multiplied by the corresponding elements in the node state feature vector, and then the above products are added together to obtain a new value, which to a certain extent reflects the spatial correlation relationship between the node state features near this position.

[0146] By performing convolution operations on the entire node state feature vector, the spatial correlation features between path nodes are finally obtained. This spatial correlation feature can reflect the mutual influence relationship between different nodes in the spatial layout (here, the spatial layout refers to the positional relationship in the network topology). For example, if there is a certain correlation between the available bandwidth and the historical transmission packet loss rate of two adjacent nodes in the network topology, then this correlation will be reflected in the spatial correlation feature.

[0147] Step S1433-2: Align the spatial correlation feature with the historical path switching records in time series to generate path stability time series features.

[0148] The spatial correlation features obtained in this embodiment reflect the spatial relationship between nodes, while the historical path switching records contain relevant information about path switching in the past in the 3D animation film transmission link, such as switching time, switching reason, node states before and after switching, etc. In order to generate path stability time series features, it is necessary to align the spatial correlation feature with the historical path switching records in time series.

[0149] Assume that the spatial correlation features are obtained at a certain time interval (for example, features obtained by collecting data every 10 seconds), and the timestamps in the historical path switching records are also recorded with a certain time accuracy. First, it is necessary to find the corresponding relationship between the two in time. For example, the data collected at a certain time point in the spatial correlation feature should be matched with the records in the historical path switching records near this time point (considering factors such as possible collection time differences).

[0150] For each time point or time period, combine or correlate the relevant data in the spatial correlation feature with the data in the historical path switching records. For example, if the spatial correlation feature shows the bandwidth correlation between nodes at a certain moment, and the historical path switching records show the load change of a certain node at this moment, integrate these two pieces of information. By performing such operations on all time points or time periods, path stability time series features are finally generated. This path stability time series feature combines the spatial relationship between nodes and the historical path switching situation, and can better reflect the variation law of path stability over time.

[0151] Step S1433-3, call the pre-trained stability prediction sub-model to perform regression processing on the path stability time series features, and output the transmission stability score.

[0152] The pre-trained stability prediction sub-model is trained based on a large amount of historical 3D animation film transmission link data (including data such as different network topologies, node states, path switching situations, etc.). Input the generated path stability time series features into this stability prediction sub-model.

[0153] Regression processing is performed inside the model, and this regression processing process involves mathematical operations and the model structure. For example, the model may contain multiple hidden layers, and neurons in each hidden layer perform operations such as weighted summation and non-linear transformation on the input path stability time series features. Assume that the path stability time series features have multiple dimensional values, such as [10, 12, 15, 13, 11] (these are just example values). The neurons in the first layer of the model will perform weighted summation on the above values according to the pre-set weights. For example, the weights of the first neuron are [0.1, 0.2, 0.3, 0.2, 0.2], then the result of the weighted summation is 10×0.1 + 12×0.2 + 15×0.3 + 13×0.2 + 11×0.2 = 1 + 2.4 + 4.5 + 2.6 + 2.2 = 12.7. Then, through a non-linear transformation function (such as the ReLU function, which changes values less than 0 to 0 and keeps values greater than 0 unchanged), new values are obtained.

[0154] After being processed by multiple hidden layers, a value representing the transmission stability score is finally output. Assume that after being processed by the model, the output transmission stability score is 0.6. This transmission stability score represents the stability degree of the current 3D animation film transmission link after considering the spatial correlation features between path nodes and the historical path switching records. 0.6 indicates a certain degree of stability, but there are also some factors affecting stability.

[0155] Step S1433-4, construct a delay prediction equation based on the node available bandwidth and the adjacent node hop count, and calculate the delay guarantee score.

[0156] In the 3D animation film transmission link, the node available bandwidth and the adjacent node hop count are important factors affecting the delay. When constructing the delay prediction equation, the relationship between these two factors and their combined impact on the delay need to be considered.

[0157] It is assumed that the construction of the delay prediction equation is based on the principle that the delay is directly proportional to the number of hops of adjacent nodes and inversely proportional to the available bandwidth of the nodes. For example, let the delay be T, the available bandwidth of the node be B (Mbps), and the number of hops of adjacent nodes be H. A simple delay prediction equation can be constructed as: T = k × H / B (where k is a constant determined according to network characteristics and experience).

[0158] It is assumed that the available bandwidth B of the node is 5 Mbps, the number of hops H of adjacent nodes is 3, and the constant k is determined to be 10 according to historical data and network analysis. Then, according to the equation, the calculated delay T = 10 × 3 / 5 = 6 (the unit here can be milliseconds or other time units, defined according to the specific network environment and data).

[0159] Then, compare the calculated delay value with the pre-set delay standard value or reference value to generate a delay guarantee score. It is assumed that the pre-set delay standard value is 5 milliseconds and the calculated delay is 6 milliseconds. The closer the delay is to the standard value, the higher the delay guarantee score. A simple calculation method can be adopted, such as delay guarantee score = (standard value / calculated value) × 0.5 + 0.5 (this calculation method is just an example and can be adjusted according to actual needs). Then, the delay guarantee score = (5 / 6) × 0.5 + 0.5 = 0.4167 × 0.5 + 0.5 = 0.20835 + 0.5 = 0.70835. This delay guarantee score represents the degree of delay guarantee in the current node state.

[0160] Step S1433-5, perform normalization processing on the transmission stability score and the delay guarantee score so that they fall into the same numerical range.

[0161] The transmission stability score and the delay guarantee score may have different numerical ranges and meanings. For example, the transmission stability score may be between 0 and 1, while the delay guarantee score may be between 0 and 2. In order to perform a unified weighted aggregation operation on these two scores, normalization processing is required.

[0162] For the normalization processing of the delay guarantee score, assume that its original range is 0-2 and it needs to be normalized to between 0 and 1. If the original value of the delay guarantee score is 0.70835, the method of linear transformation is adopted. Calculate the normalized delay guarantee score = (original value - minimum value) / (maximum value - minimum value) = (0.70835 - 0) / (2 - 0) = 0.354175. In this way, the delay guarantee score is normalized to between 0 and 1 and is in the same numerical range as the transmission stability score.

[0163] In a possible implementation manner, step S150 includes:

[0164] Step S151: Analyze the encoding parameter adjustment instructions in the dynamic transmission optimization instruction set to generate the updated quantization parameter value for the encoder node and the motion estimation range limit threshold. Reconstruct the control parameter table of the encoder according to the updated quantization parameter value and the motion estimation range limit threshold.

[0165] In the 3D animation film transmission link, the dynamic transmission optimization instruction set contains a series of instructions for transmission optimization. When analyzing the encoding parameter adjustment instructions therein, assume that the encoding parameter adjustment instructions stipulate to adjust the quantization parameter and the motion estimation range of the encoder.

[0166] For example, the encoding parameter adjustment instruction indicates that the quantization parameter should be updated from the original 30 to 25. At the same time, the motion estimation range limit threshold also needs to be adjusted. Assume that the original motion estimation range limit threshold is 100 pixels, and according to the instruction, it should be adjusted to 80 pixels.

[0167] Reconstruct the control parameter table of the encoder according to the above-mentioned updated quantization parameter value and the motion estimation range limit threshold. The control parameter table of the encoder contains multiple encoding-related parameters, such as quantization parameters, motion estimation ranges, frame rate control parameters, etc. After updating the quantization parameter to 25 and the motion estimation range limit threshold to 80 pixels, corresponding adjustments are made to other relevant parameters in the control parameter table (such as encoding block size and other relevant parameters that may need to be adjusted according to the change of the quantization parameter) to ensure that the encoder can work properly under the new parameter settings and improve the encoding efficiency and transmission quality of the 3D animation film.

[0168] Step S152: Analyze the path switching instructions in the dynamic transmission optimization instruction set, update the next-hop address mapping table of the routing node, send a reconfiguration command containing the control parameter table to the encoder node, and send a path update command containing the next-hop address mapping table to the routing node.

[0169] When analyzing the path switching instructions in the dynamic transmission optimization instruction set, assume that the path switching instruction specifies to switch the transmission path from the current path to a new alternative path, which was determined to be a better path during the previous path selection process.

[0170] First, update the next-hop address mapping table of the routing node according to the path switching instruction. The next-hop address mapping table determines the forwarding direction of data packets at the routing node. For example, the original next-hop address was node A, and according to the path switching instruction, the next-hop address is updated to node B.

[0171] Then, send a reconfiguration command containing the reconstructed control parameter table to the encoder node, and this reconfiguration command will notify the encoder node to perform encoding operations according to the new control parameter table. At the same time, send a path update command containing the updated next-hop address mapping table to the routing node, so that the routing node can forward data packets according to the new path.

[0172] Step S153: Send status query requests to the encoder node and the routing node respectively, obtain the currently effective quantization parameter value and the next-hop address, and compare the difference between the currently effective quantization parameter value and the target value in the control parameter table to generate an encoder synchronization status flag.

[0173] In this embodiment, after sending a status query request to the encoder node, the encoder node will return the currently effective quantization parameter value. Assume that the currently effective quantization parameter value returned is 30 (possibly due to certain reasons, such as the instruction has not been fully effective or there are other interference factors).

[0174] Compare the currently effective quantization parameter value 30 with the target value 25 in the control parameter table. Since 30 is not equal to 25 and there is a difference, an encoder synchronization status flag of not synchronized is generated. This encoder synchronization status flag indicates that the current quantization parameter value of the encoder node is inconsistent with the target value that should be effective, and further adjustment operations are required.

[0175] Step S154: Check the consistency between the currently effective next-hop address and the target address in the path update command to generate a routing synchronization status flag.

[0176] In this embodiment, after sending a status query request to the routing node, obtain the currently effective next-hop address. Assume that the currently effective next-hop address is node A, while the target address in the path update command is node B.

[0177] Check the consistency between the currently effective next-hop address (node A) and the target address (node B). Since the two are different, a routing synchronization status flag of not synchronized is generated. This routing synchronization status flag indicates that the current next-hop address of the routing node is inconsistent with the target address that should be effective, and adjustment operations are also required.

[0178] Step S155: When the encoder synchronization status flag or the routing synchronization status flag is not synchronized, increment the retry counter and resend the corresponding command.

[0179] In this embodiment, since the encoder synchronization status flag or the routing synchronization status flag is not synchronized, this means that the encoder node and the routing node have not updated the parameters as expected. At this time, a retry counter starts to work. Assume that the initial value of the retry counter is 0.

[0180] When an unsynchronized situation is detected, increment the value of the retry counter by 1 to become 1. Then resend the corresponding commands, that is, resend the reconfiguration command containing the control parameter table to the encoder node, and resend the path update command containing the next-hop address mapping table to the routing node, hoping to enable the encoder node and the routing node to correctly update the parameters and reach the synchronized state by resending the commands.

[0181] Step S156, if the retry counter exceeds the preset threshold, generate an alarm log and trigger the manual intervention process.

[0182] For example, assume the preset threshold is 3. If the retry counter keeps increasing and reaches 3, it means that after multiple attempts, the encoder node and the routing node still cannot reach the synchronized state.

[0183] At this time, an alarm log can be generated correspondingly. This alarm log can record relevant information, such as the unsynchronized parameter information (quantization parameters of the encoder, next-hop address of the routing node, etc.), the number of retries, the time when the problem occurred, etc. At the same time, trigger the manual intervention process. The manual intervention process may include technicians manually checking the encoder node and the routing node to see if there are hardware failures, network connection problems or other abnormal situations, and then performing manual parameter adjustment or fault repair operations according to the specific situation to ensure that the 3D animation film transmission link can work properly.

[0184] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of a film transmission quality control system 100 based on AI intelligent analysis that can implement the idea of the present application provided by some embodiments of the present application. For example, the processor 120 can be used on the film transmission quality control system 100 based on AI intelligent analysis and is used to execute the functions in the present application.

[0185] The film transmission quality control system 100 based on AI intelligent analysis can be a general-purpose server or a special-purpose server, both of which can be used to implement the film transmission quality control method based on AI intelligent analysis of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0186] For example, the film transmission quality control system 100 based on AI intelligent analysis may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the film transmission quality control system 100 based on AI intelligent analysis may further include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to the above program instructions. The film transmission quality control system 100 based on AI intelligent analysis further includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0187] For ease of explanation, only one processor is described in the film transmission quality control system 100 based on AI intelligent analysis. However, it should be noted that the film transmission quality control system 100 in the present application may further include multiple processors. Therefore, the steps executed by one processor described in the present application may also be jointly executed or separately executed by multiple processors. For example, if the processor of the film transmission quality control system 100 based on AI intelligent analysis executes step A and step B, it should be understood that step A and step B may also be jointly executed by two different processors or separately executed in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0188] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the film transmission quality control method based on AI intelligent analysis as described above is implemented.

[0189] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A film transmission quality control method based on AI intelligent analysis, characterized in that: The method comprises: Acquire a real-time frame sequence data set in a target film transmission link, wherein the real-time frame sequence data set includes encoding parameters, transmission path node identifiers, and decoding delay indicators of a plurality of consecutive transmission frames; Performing transmission quality feature extraction processing on the real-time frame sequence data set to generate frame quality assessment features of each transmission frame and frame sequence dynamic features between adjacent transmission frames; Calling a pre-trained transmission quality analysis model, performing feature fusion processing on the frame quality assessment features and the frame sequence dynamic features, and generating a transmission quality attenuation prediction result of the target film transmission link; Generate a dynamic transmission optimization instruction set based on matching the transmission quality attenuation prediction result with a preset transmission parameter adjustment strategy set; The dynamic transmission optimization instruction set is fed back to the encoder node and the routing node corresponding to the target video transmission link to trigger the transmission parameter adaptive adjustment operation.

2. The film transmission quality control method based on AI intelligent analysis according to claim 1 is characterized in that: The performing of transmission quality feature extraction processing on the real-time frame sequence data set to generate frame quality assessment features of each transmission frame and frame sequence dynamic features between adjacent transmission frames includes: The coding parameters of each transmission frame are processed in blocks to obtain multiple coding parameter sub-block sets; Performing a dynamic feature extraction operation on the coding parameter sub-block set to generate inter-block quality fluctuation features and intra-block coding distortion features of each transmission frame; Aggregating the inter-block quality fluctuation characteristics and the intra-block coding distortion characteristics of the same transmission frame to generate a frame quality assessment feature of the transmission frame; The decoding delay index difference characteristics between adjacent transmission frames are extracted, and the path delay correlation analysis is performed in combination with the transmission path node identifier to generate the frame sequence dynamic characteristics; wherein the frame sequence dynamic characteristics include the inter-frame decoding delay correlation index and the path node load fluctuation correlation.

3. The film transmission quality control method based on AI intelligent analysis according to claim 2 is characterized in that: The extracting the decoding delay index difference characteristics between adjacent transmission frames and performing path delay correlation analysis in combination with the transmission path node identifier to generate the frame sequence dynamic characteristics includes: Perform sliding window statistics on the decoding delay indicators of N consecutive transmission frames to generate delay mean features and delay variance features; Matching a historical path delay record set according to the transmission path node identifier, extracting a load fluctuation feature of a current path node within a preset time period; Performing time alignment processing on the delay mean feature, the delay variance feature, and the load fluctuation feature to generate a delay-load joint analysis matrix; Calling a pre-trained delay correlation model to perform feature mapping on the delay-load joint analysis matrix, and outputting the inter-frame decoding delay correlation index; A path topology graph is constructed based on the transmission path node identifier, and the path node load fluctuation correlation is generated according to the real-time load data of each node in the path topology graph.

4. The film transmission quality control method based on AI intelligent analysis according to claim 2 is characterized in that: The calling of the pre-trained transmission quality analysis model, performing feature fusion processing on the frame quality assessment features and the frame sequence dynamic features, and generating a transmission quality attenuation prediction result of the target film transmission link, includes: Inputting the frame quality assessment feature into a first feature coding layer of the transmission quality analysis model to generate a first high-order quality feature; Inputting the inter-frame decoding delay correlation index and the path node load fluctuation correlation into the second feature coding layer of the transmission quality analysis model to generate a second high-order dynamic feature; Performing cross-modal attention weighted processing on the first high-order quality feature and the second high-order dynamic feature to generate a fused quality-dynamic feature vector; Performing similarity matching in a preset attenuation pattern library according to the fused quality-dynamic feature vector to determine the quality attenuation pattern category corresponding to the target film transmission link; A preset attenuation prediction parameter table is queried based on the quality attenuation mode category, and the transmission quality attenuation prediction result is output.

5. The film transmission quality control method based on AI intelligent analysis according to claim 4 is characterized in that: The performing cross-modal attention weighted processing on the first high-order quality feature and the second high-order dynamic feature to generate a fused quality-dynamic feature vector includes: Using the first high-order quality feature as a query vector and the second high-order dynamic feature as a key vector and a value vector; Calculating the dot product similarity between the query vector and the key vector to generate an initial attention weight distribution; Performing nonlinear activation processing on the initial attention weight distribution to generate normalized attention weights; Performing weighted summation on the value vector according to the normalized attention weight to generate a dynamic feature enhancement vector; The dynamic feature enhancement vector is residually connected with the first high-order quality feature to generate the fused quality-dynamic feature vector.

6. The film transmission quality control method based on AI intelligent analysis according to claim 1 is characterized in that: The generating a dynamic transmission optimization instruction set based on matching the transmission quality attenuation prediction result with a preset transmission parameter adjustment strategy set includes: Parsing a key attenuation factor set in the transmission quality attenuation prediction result, wherein the key attenuation factor set includes a coding complexity exceeding mark, a path node congestion level, and a decoding buffer overflow probability; According to the coding complexity exceeding mark, a preset coding parameter adjustment rule base is queried to generate a first optimization instruction subset; Invoking a dynamic routing decision model to generate a subset of alternative path switching instructions according to the congestion level of the path nodes; adjusting a buffer threshold parameter of a target decoder according to the decoding buffer overflow probability to generate a buffer control instruction subset; The first optimization instruction subset, the alternative path switching instruction subset and the buffer control instruction subset are aggregated to generate the dynamic transmission optimization instruction set.

7. The film transmission quality control method based on AI intelligent analysis according to claim 6 is characterized in that: The calling of the dynamic routing decision model according to the congestion level of the path node to generate a subset of alternative path switching instructions includes: Acquire real-time topology status data of the current transmission path node, wherein the real-time topology status data includes node available bandwidth, number of adjacent node hops, and historical transmission packet loss rate; Inputting the real-time topology state data into the feature extraction layer of the dynamic routing decision model to generate a node state feature vector; Calling the multi-path decision layer of the dynamic routing decision model to perform path score calculation on the node state feature vector to generate a transmission stability score and a delay guarantee score for each candidate path; Standardize the transmission stability score and the delay guarantee score to generate a standardized transmission stability score and a standardized delay guarantee score; According to the preset transmission stability weight coefficient and delay guarantee weight coefficient, the standardized transmission stability score and the standardized delay guarantee score are weighted and aggregated to generate a comprehensive optimization score for each alternative path; Based on the comprehensive optimization score, the alternative paths are sorted in descending order to generate an alternative path priority sequence, and the highest-ranked alternative path identifier in the alternative path priority sequence is extracted as a reference alternative path identifier, and the reference alternative path identifier is matched and verified with a preset path switching condition. If the verification passes, it is output as the optimal alternative path identifier, otherwise it triggers the operation of regenerating the alternative path priority sequence; Generate an alternative path switching instruction subset including the optimal alternative path identifier and the switching time window.

8. The film transmission quality control method based on AI intelligent analysis according to claim 7 is characterized in that: The multi-path decision layer calling the dynamic routing decision model performs path score calculation on the node state feature vector to generate a transmission stability score and a delay guarantee score for each candidate path, including: Performing a convolution operation on the node state feature vector to extract spatial correlation features between path nodes; Performing time series alignment of the spatial correlation features and the historical path switching records to generate path stability time series features; Calling a pre-trained stability prediction sub-model to perform regression processing on the path stability time series characteristics, and outputting the transmission stability score; Building a delay estimation equation based on the available bandwidth of the node and the number of hops of the adjacent nodes, and calculating the delay guarantee score; The transmission stability score and the delay guarantee score are standardized so that they fall into the same numerical range.

9. The film transmission quality control method based on AI intelligent analysis according to claim 1 is characterized in that: Feeding back the dynamic transmission optimization instruction set to the encoder node and routing node corresponding to the target video transmission link to trigger the transmission parameter adaptive adjustment operation includes: Parsing the encoding parameter adjustment instruction in the dynamic transmission optimization instruction set, generating a quantization parameter update value and a motion estimation range limit threshold of an encoder node, and reconstructing a control parameter table of the encoder according to the quantization parameter update value and the motion estimation range limit threshold; Parsing the path switching instruction in the dynamic transmission optimization instruction set, updating the next hop address mapping table of the routing node, sending a reconfiguration command including the control parameter table to the encoder node, and sending a path update command including the next hop address mapping table to the routing node; Sending a status query request to the encoder node and the routing node respectively, obtaining a currently effective quantization parameter value and a next hop address, and comparing the currently effective quantization parameter value with a target value in the control parameter table to generate an encoder synchronization status identifier; Performing a consistency check between the currently effective next hop address and the target address in the path update command to generate a routing synchronization status identifier; When the encoder synchronization state flag or the routing synchronization state flag is not synchronized, incrementing a retry counter and resending a corresponding command; If the retry counter exceeds a preset threshold, an alarm log is generated and a manual intervention process is triggered.

10. A film transmission quality control system based on AI intelligent analysis, characterized in that: The film transmission quality control system based on AI intelligent analysis includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the film transmission quality control method based on AI intelligent analysis as described in any one of claims 1 to 9 above.

Images