Narrowband environment video code rate optimization system and method based on convolutional neural network

Abstract

The invention discloses a narrowband environment video code rate optimization system and method based on a convolutional neural network, and belongs to the technical field of video processing. The method comprises the following steps: after a video frame sequence is obtained by configuring a video transmission synchronization time sequence, setting the grouping number of a grouping convolution CNN model according to the number of feature types, and generating an initialized input feature map; establishing a pixel two-dimensional coordinate system, and dynamically adjusting the convolution weight by combining the real-time bandwidth and the transmission delay to obtain a local enhanced pixel value; constructing a pixel state matrix, and calculating the video frame sequence similarity of the current and historical synchronous transmission stages; the average similarity of the windows is calculated through sliding window grouping, a trend function is constructed, the scale of the sliding windows is optimized, the picture change trend is judged according to the slope of the trend function, and the code rate compression mode of the next stage is indicated. Dynamic balance of video transmission image quality and transmission efficiency in a narrowband environment is realized, the problems of poor adaptability and adjustment lag of a traditional scheme are solved, and the method is suitable for a narrowband Internet of Things scene.

Description

Technical Field

[0001] This invention relates to the field of video processing technology, specifically to a narrowband environment video bitrate optimization system and method based on convolutional neural networks. Background Technology

[0002] With the popularization of technologies such as the Internet of Things (IoT) and remote monitoring, video transmission is increasingly used in narrowband environments, such as forest fire prevention monitoring in remote mountainous areas, rural security monitoring, and data transmission from outdoor IoT devices. In these scenarios, network bandwidth resources are limited and fluctuate drastically, resulting in high transmission latency. Traditional video bitrate optimization solutions are insufficient to meet actual needs and present numerous technical bottlenecks.

[0003] In existing technologies, narrowband video bitrate control often employs fixed bitrate or simple dynamic bitrate adjustment strategies, failing to fully consider the dynamic adaptation of video frame feature differences to narrowband transmission conditions. On one hand, traditional solutions often use a single convolutional model for feature extraction, without grouping features by type, resulting in low feature extraction efficiency and an inability to accurately distinguish the impact of different features on image quality, leading to unreasonable bitrate allocation. On the other hand, existing bitrate adjustments rely heavily on passive feedback from real-time transmission status, without considering the similarity trends of historical transmission data for forward prediction, resulting in a lag in bitrate adjustments: when bandwidth suddenly decreases, the bitrate cannot be reduced in advance, causing transmission stuttering; when the image content changes continuously, the bitrate cannot be increased in time, resulting in blurred image quality. Furthermore, traditional similarity calculations only apply to single frames or local segments, without using sliding window grouping to calculate average similarity and optimize trend stability, leading to inaccurate judgments of image content change trends and a lack of scientific rigor in the bitrate adjustment strategy. Meanwhile, existing solutions do not incorporate a narrowband adaptation factor to dynamically adjust convolution weights, failing to effectively adapt to real-time fluctuations in bandwidth and latency. This makes it difficult to balance transmission stability and image quality integrity in narrowband environments. These issues severely restrict the user experience of video transmission in narrowband environments. Summary of the Invention

[0004] The purpose of this invention is to provide a narrowband video bitrate optimization system and method based on convolutional neural networks to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for optimizing video bitrate in narrowband environments based on convolutional neural networks, the method comprising the following steps: Step S1: Configure the synchronization timing of video transmission, obtain the video frame sequence within the synchronized transmission segment, set the number of groups for the convolutional neural network feature extraction model according to the number of video frame feature types, and generate initial input feature maps corresponding to each type of feature for each video frame. Step S2: Establish a two-dimensional coordinate system for video frame pixels, set the convolution kernel sliding offset, perform intra-group convolution operation on each initial input feature map, and dynamically adjust the convolution weights in combination with real-time bandwidth and transmission delay to obtain the local enhanced pixel value of each pixel. Step S3: Evaluate the impact of different feature types on image quality based on local enhanced pixel values ​​to construct the pixel state matrix of the video frame; Step S4: Integrate the pixel state matrices of all video frames within the same synchronous transmission segment to form a sample set, take the synchronous transmission segment as the current stage, and calculate the similarity between the current stage and the video frame sequences of all historical stages. Step S5: Set the sliding window scale, group the historical stages, and calculate the average window similarity between each group of historical stages and the current stage; Step S6: Based on the sliding window number and the corresponding average window similarity, construct a trend function, adjust the sliding window scale to optimize trend stability, and judge the video content change trend according to the trend function characteristics to indicate the video bitrate compression method in the next stage.

[0006] Preferably, the specific implementation process of step S1 includes: The synchronization timing of video transmission is configured uniformly to form a synchronized transmission timing set, denoted as . and obtain synchronous transmission segments. The video frame sequence within, where, This represents the nth synchronous transmission segment consisting of two adjacent synchronous transmission nodes, where N represents the total number of synchronous transmission segments. A CNN feature extraction model using grouped convolution is adopted, with the number of feature types in video frames as the number of groups G in the CNN feature extraction model; For synchronous transmission segments The m-th video frame in the video frame sequence Video frames are generated under the guidance of video frame feature types. There are G initial input feature maps, where one initial input feature map is generated for each type of video frame feature.

[0007] Preferably, the specific implementation process of step S2 includes: Establish a two-dimensional coordinate system based on the macroblock pixel size in the video frame, and represent the pixel as... ,in, The x-axis value, The vertical axis value; Set the horizontal and vertical sliding offsets of the convolution kernel, respectively. and , forming sliding pixels ; The g-th initial input feature map is initialized at pixel point. The pixel value at that location is denoted as ; Perform intra-group convolution on the g-th initialized input feature map to obtain pixel points. Local enhancement pixel value at ,in, This indicates that the convolution kernel moves at the pixel point. The weight of the position, This represents the narrowband adaptation factor, used for dynamically adjusting the weights. The value, In response to video frames Real-time bandwidth at that time Indicates response to video frames Narrowband transmission delay is determined by real-time bandwidth. and narrowband transmission delay Linear quantization narrowband fit factor , , and All are linear coefficients after training. Represents pixels With sliding pixels Weight bias between them.

[0008] Preferably, the specific implementation process of step S3 includes: In the case where an initial input feature map is generated corresponding to a certain video frame feature type, based on the video frame... Local enhancement pixel values Evaluate the image quality impact of the feature type of the g-th video frame. Where max{} and min{} are the maximum and minimum value functions, respectively; Construct the pixel state matrix of the video frame, where the row index of the pixel state matrix is ​​the pixel number. The column index is the video frame feature type. If image quality affects If the value is greater than or equal to the preset image quality impact threshold, then the pixel at row index [number] is [number]. The column index is a video frame feature type. The corresponding matrix position is set to 1 if the image quality impact is... If the value is less than the preset image quality impact threshold, then the pixel at the row index is... The column index is a video frame feature type. Set the corresponding matrix position to 0 to obtain the video frame. pixel state matrix .

[0009] Preferably, the specific implementation process of step S4 includes: Obtain Synchronous Transmission Segments The pixel state matrix of each video frame in the video frame sequence constitutes the segment state matrix sample set, denoted as . ,in, Indicates synchronous transmission segment The total number of video frames contained in the video frame sequence; If the synchronous transmission segment As the current stage of bitrate optimization, the segments will be transmitted synchronously. All previous synchronous transmission segments are considered as historical phases; Analyze the similarity of video frame sequences between the current stage and any historical stage. ,in, This represents the h-th video frame. Represents video frames The pixel state matrix, This represents the e-th synchronous transmission segment. The corresponding generated fragment state matrix sample set, Indicates synchronous transmission segment The total number of video frames contained in the video frame sequence, e < n. For counting functions, Represents the pixel state matrix With pixel state matrix The number of 1s contained in the result of an inter-Boolean logical AND operation. Represents the pixel state matrix With pixel state matrix The number of 1s contained after a Boolean logical OR operation, where P is a preset similarity threshold. If the condition is met... Then let If not satisfied Then let .

[0010] Preferably, the specific implementation process of step S5 includes: Synchronous transmission segments As the current stage of rate optimization, based on the order of synchronous transmission segments in the synchronous transmission time sequence set, the synchronous transmission segments are... The preceding synchronous transmission segments serve as a historical phase for bitrate optimization. A sliding window for the preset synchronous transmission segment is defined, with a sliding window size of d. Based on the sliding window scale d, generate the historical stage grouping set of the j-th sliding window. Used to record historical stages, and ; Calculate historical stage grouping set Average similarity between windows of all historical stages and the current stage ,in, Representing historical stage grouping sets The total number of synchronous transmission segments included.

[0011] Preferably, the specific implementation process of step S6 includes: Two-dimensional data points are constructed based on the sliding window index and the average similarity of the windows. The similarity trend function is obtained. ,in, The slope The intercept; J represents the total number of sliding windows; Based on similarity trend function Assess trend stability , The overall average similarity of all sliding windows; Adjusting the sliding window scale d value outputs a trend stability. Maximize the sliding window size d and apply it to the similarity trend function. Update; Based on the updated similarity trend function ,like This indicates that the earlier historical stages are similar to the current stage, and the content shows a "retrospective similarity" trend. It predicts that the next stage will use a low bitrate for video compression. If the recent historical phase is similar to the current phase, and the content of the images is highly continuous, it is predicted that a high bitrate method will be used for video compression in the next phase after the current phase.

[0012] Secondly, the present invention provides a system for video bitrate optimization in narrowband environments. This system includes a storage device and a processor. The storage device includes a program for a narrowband video bitrate optimization method based on a convolutional neural network. When the processor executes the program for the narrowband video bitrate optimization method based on a convolutional neural network, it performs the following steps: Configure the synchronization timing of video transmission, obtain the video frame sequence within the synchronized transmission segment, set the number of groups for the convolutional neural network feature extraction model according to the number of video frame feature types, and generate initial input feature maps corresponding to each type of feature for each video frame. A two-dimensional coordinate system for video frame pixels is established, the sliding offset of the convolution kernel is set, intra-group convolution operation is performed on each initial input feature map, and the convolution weights are dynamically adjusted in combination with real-time bandwidth and transmission delay to obtain the local enhanced pixel value of each pixel. The impact of different feature types on image quality is evaluated based on locally enhanced pixel values ​​in order to construct the pixel state matrix of video frames; The pixel state matrices of all video frames within the same synchronous transmission segment are integrated to form a sample set. This synchronous transmission segment is taken as the current stage, and the similarity between the current stage and the video frame sequences of all historical stages is calculated. Set the sliding window scale, group the historical stages, and calculate the average window similarity between each group of historical stages and the current stage; Based on the sliding window number and the corresponding average window similarity, a trend function is constructed, the sliding window scale is adjusted to optimize trend stability, and the trend of video content change is judged according to the characteristics of the trend function to indicate the video bitrate compression method in the next stage.

[0013] Compared with the prior art, the beneficial effects achieved by the present invention are: Existing technologies mostly use fixed convolution weights, which cannot cope with the dynamic fluctuations of narrowband bandwidth and latency. This invention dynamically adjusts the weights through a narrowband adaptation factor, so that the feature extraction process matches the narrowband state in real time. This invention uses a grouped convolutional CNN model combined with a narrowband adaptation factor to dynamically adjust the convolution weights. It not only ensures the extraction accuracy of different feature types through group processing, but also achieves dynamic adaptation of convolution weights through linear quantization of real-time bandwidth and transmission latency, so as to maintain stable feature extraction performance even in narrowband fluctuation scenarios. Traditional solutions fail to quantify the impact of different features on image quality, leading to blind bitrate allocation. This invention achieves precise quantification of image quality impact through a pixel state matrix. To avoid image quality distortion in critical areas and to consider the redundancy of bitrate in non-critical areas, a 0-1 matrix is ​​used to accurately mark critical image quality areas, enabling differentiated bitrate allocation. Existing technologies rely on real-time feedback for passive bitrate adjustment, which has a lag. This invention analyzes historical similarity trends using a sliding window and trend function to predict changes in the image in advance. By adjusting the scale of the sliding window to maximize trend stability, it achieves forward prediction of image change trends, transforming bitrate adjustment from "passive feedback" to "active prediction," which helps improve the response speed of bitrate adjustment. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0015] Figure 1 This is a schematic diagram illustrating the steps of the narrowband environment video bitrate optimization method based on convolutional neural networks of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 In this first embodiment, a method for optimizing video bitrate in narrowband environments based on convolutional neural networks is provided. This embodiment is applied to a forest fire monitoring scenario in remote mountainous areas. This scenario is a typical narrowband environment: the bandwidth is stable at 1-2 Mbps, and the transmission delay fluctuates between 50-150 ms. Video frame feature types may include brightness, contrast, edge, texture, color, etc. The method includes the following steps: Step S1: Configure the synchronization timing of video transmission, obtain the video frame sequence within the synchronized transmission segment, set the number of groups for the convolutional neural network feature extraction model according to the number of video frame feature types, and generate initial input feature maps corresponding to each type of feature for each video frame. Specifically, the synchronization timing of video transmission is configured uniformly to form a synchronized transmission timing set, denoted as... and obtain synchronous transmission segments. The video frame sequence within, where, This represents the nth synchronous transmission segment consisting of two adjacent synchronous transmission nodes, where N represents the total number of synchronous transmission segments. A CNN feature extraction model using grouped convolution is adopted, with the number of feature types in video frames as the number of groups G in the CNN feature extraction model; For synchronous transmission segments The m-th video frame in the video frame sequence Video frames are generated under the guidance of video frame feature types. There are G initial input feature maps, where one initial input feature map is generated for each type of video frame feature.

[0018] Step S2: Establish a two-dimensional coordinate system for video frame pixels, set the convolution kernel sliding offset, perform intra-group convolution operation on each initial input feature map, and dynamically adjust the convolution weights in combination with real-time bandwidth and transmission delay to obtain the local enhanced pixel value of each pixel. Specifically, a two-dimensional coordinate system based on the macroblock pixel size in a video frame is used to represent pixels as follows: ,in, The x-axis value, The vertical axis value; Set the horizontal and vertical sliding offsets of the convolution kernel, respectively. and , forming sliding pixels ; The g-th initial input feature map is initialized at pixel point. The pixel value at that location is denoted as ; Perform intra-group convolution on the g-th initialized input feature map to obtain pixel points. Local enhancement pixel value at ,in, This indicates that the convolution kernel moves at the pixel point. The weight of the position, This represents the narrowband adaptation factor, used for dynamically adjusting the weights. The value, In response to video frames Real-time bandwidth at that time Indicates response to video frames Narrowband transmission delay is determined by real-time bandwidth. and narrowband transmission delay Linear quantization narrowband fit factor , , and All are linear coefficients after training. Represents pixels With sliding pixels Weight bias between; For example, establish a two-dimensional coordinate system for macroblock pixels, where pixel (x, y) covers a video frame resolution of 1920×1080; and use the narrowband adaptation factor γᵍ(B) m D m )=0.3B m -0.2D m +0.5 dynamically adjusts the convolution kernel weights; when the real-time bandwidth B m =1.2Mbps, latency D m When B = 100ms, γᵍ = 0.3×1.2 - 0.2×100 + 0.5 = -19.14, effectively reducing the convolution weights and adapting to low-bandwidth transmission; when B m =1.8Mbps, D m =60ms, γᵍ=0.3×1.8-0.2×60+0.5=-11.06, the weight should be increased appropriately to retain more details.

[0019] Step S3: Evaluate the impact of different feature types on image quality based on local enhanced pixel values ​​to construct the pixel state matrix of the video frame; Specifically, when an initial input feature map is generated corresponding to a certain video frame feature type, based on the video frame... Local enhancement pixel values Evaluate the image quality impact of the feature type of the g-th video frame. Where max{} and min{} are the maximum and minimum value functions, respectively; Construct the pixel state matrix of the video frame, where the row index of the pixel state matrix is ​​the pixel number. The column index is the video frame feature type. If image quality affects If the value is greater than or equal to the preset image quality impact threshold, then the pixel at row index [number] is [number]. The column index is a video frame feature type. The corresponding matrix position is set to 1 if the image quality impact is... If the value is less than the preset image quality impact threshold, then the pixel at the row index is... The column index is a video frame feature type. Set the corresponding matrix position to 0 to obtain the video frame. pixel state matrix .

[0020] Step S4: Integrate the pixel state matrices of all video frames within the same synchronous transmission segment to form a sample set, take the synchronous transmission segment as the current stage, and calculate the similarity between the current stage and the video frame sequences of all historical stages. Specifically, acquiring synchronous transmission segments The pixel state matrix of each video frame in the video frame sequence constitutes the segment state matrix sample set, denoted as . ,in, Indicates synchronous transmission segment The total number of video frames contained in the video frame sequence; If the synchronous transmission segment As the current stage of bitrate optimization, the segments will be transmitted synchronously. All previous synchronous transmission segments are considered as historical phases; Analyze the similarity of video frame sequences between the current stage and any historical stage. ,in, This represents the h-th video frame. Represents video frames The pixel state matrix, This represents the e-th synchronous transmission segment. The corresponding generated fragment state matrix sample set, Indicates synchronous transmission segment The total number of video frames contained in the video frame sequence, e < n. For counting functions, Represents the pixel state matrix With pixel state matrix The number of 1s contained in the result of an inter-Boolean logical AND operation. Represents the pixel state matrix With pixel state matrix The number of 1s contained after a Boolean logical OR operation, where P is a preset similarity threshold. If the condition is met... Then let If not satisfied Then let .

[0021] Step S5: Set the sliding window scale, group the historical stages, and calculate the average window similarity between each group of historical stages and the current stage; Specifically, the segments will be transmitted synchronously. As the current stage of rate optimization, based on the order of synchronous transmission segments in the synchronous transmission time sequence set, the synchronous transmission segments are... The preceding synchronous transmission segments serve as a historical phase for bitrate optimization. A sliding window for the preset synchronous transmission segment is defined, with a sliding window size of d. Based on the sliding window scale d, generate the historical stage grouping set of the j-th sliding window. Used to record historical stages, and ; Calculate historical stage grouping set Average similarity between windows of all historical stages and the current stage ,in, Representing historical stage grouping sets The total number of synchronous transmission segments included.

[0022] Step S6: Based on the sliding window number and the corresponding average window similarity, construct a trend function, adjust the sliding window scale to optimize trend stability, and judge the video content change trend according to the trend function characteristics to indicate the video bitrate compression method in the next stage. Specifically, two-dimensional data points are constructed based on the sliding window index and the average window similarity. The similarity trend function is obtained. ,in, The slope The intercept; J represents the total number of sliding windows; Based on similarity trend function Assess trend stability , The overall average similarity of all sliding windows; Adjusting the sliding window scale d value outputs a trend stability. Maximize the sliding window size d and apply it to the similarity trend function. Update; Based on the updated similarity trend function ,like This indicates that the earlier historical stages are similar to the current stage, and the content shows a "retrospective similarity" trend. It predicts that the next stage will use a low bitrate for video compression. If the recent historical phase is similar to the current phase, and the content of the images is highly continuous, it is predicted that high bitrate video compression will be used in the next phase after the current phase. It should be noted that in the similarity trend function, the smaller j is, the closer the historical stages included in the window are to the current stage (closer in time sequence); the larger j is, the further the historical stages included in the window are from the current stage (greater in time sequence). When k>0, j and They are positively correlated; the larger j is (the further the window is from the current stage), the better. The larger the value (the higher the average similarity), the more the content of the image shows a "retrospective similarity" trend; When k < 0, j and They are negatively correlated; the smaller j is (the closer the window is to the current stage). The larger the value (the higher the average similarity), the stronger the recent continuity of the image content. The sign of k is determined by the consistency between the "direction of change of j" and the "direction of change of average similarity". If the two change in the same direction, then k > 0. The direction of change of j is directly related to the "temporal distance between the window and the current stage" (the larger j is, the farther the distance).

[0023] In this second embodiment: a narrowband environment video bitrate optimization system based on convolutional neural networks is provided. This system includes a storage device and a processor. The storage device includes a program for a narrowband environment video bitrate optimization method based on convolutional neural networks. When the processor executes the program for the narrowband environment video bitrate optimization method based on convolutional neural networks, it performs the following steps: Configure the synchronization timing of video transmission, obtain the video frame sequence within the synchronized transmission segment, set the number of groups for the convolutional neural network feature extraction model according to the number of video frame feature types, and generate initial input feature maps corresponding to each type of feature for each video frame. A two-dimensional coordinate system for video frame pixels is established, the sliding offset of the convolution kernel is set, intra-group convolution operation is performed on each initial input feature map, and the convolution weights are dynamically adjusted in combination with real-time bandwidth and transmission delay to obtain the local enhanced pixel value of each pixel. The impact of different feature types on image quality is evaluated based on locally enhanced pixel values ​​in order to construct the pixel state matrix of video frames; The pixel state matrices of all video frames within the same synchronous transmission segment are integrated to form a sample set. This synchronous transmission segment is taken as the current stage, and the similarity between the current stage and the video frame sequences of all historical stages is calculated. Set the sliding window scale, group the historical stages, and calculate the average window similarity between each group of historical stages and the current stage; Based on the sliding window number and the corresponding average window similarity, a trend function is constructed, the sliding window scale is adjusted to optimize trend stability, and the trend of video content change is judged according to the characteristics of the trend function to indicate the video bitrate compression method in the next stage.

[0024] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0025] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for narrowband environment video bitrate optimization based on convolutional neural network, characterized in that, The method includes the following steps: Step S1: Configure the synchronization timing of video transmission, obtain the video frame sequence within the synchronized transmission segment, set the number of groups for the convolutional neural network feature extraction model according to the number of video frame feature types, and generate initial input feature maps corresponding to each type of feature for each video frame. Step S2: Establish a two-dimensional coordinate system for video frame pixels, set the convolution kernel sliding offset, perform intra-group convolution operation on each initial input feature map, and dynamically adjust the convolution weights in combination with real-time bandwidth and transmission delay to obtain the local enhanced pixel value of each pixel. Step S3: Evaluate the impact of different feature types on image quality based on local enhanced pixel values ​​to construct the pixel state matrix of the video frame; Step S4: Integrate the pixel state matrices of all video frames within the same synchronous transmission segment to form a sample set, take the synchronous transmission segment as the current stage, and calculate the similarity between the current stage and the video frame sequences of all historical stages. Step S5: Set the sliding window scale, group the historical stages, and calculate the average window similarity between each group of historical stages and the current stage; Step S6: Based on the sliding window number and the corresponding average window similarity, construct a trend function, adjust the sliding window scale to optimize trend stability, and judge the video content change trend according to the trend function characteristics to indicate the video bitrate compression method in the next stage.

2. The method for optimizing video bitrate in narrowband environments based on convolutional neural networks according to claim 1, characterized in that, The specific implementation process of step S1 includes: The synchronization timing of video transmission is configured uniformly to form a synchronized transmission timing set, denoted as . and obtain synchronous transmission segments. The video frame sequence within, where, This represents the nth synchronous transmission segment consisting of two adjacent synchronous transmission nodes, where N represents the total number of synchronous transmission segments. A CNN feature extraction model using grouped convolution is adopted, with the number of feature types in video frames as the number of groups G in the CNN feature extraction model; For synchronous transmission segments The m-th video frame in the video frame sequence Video frames are generated under the guidance of video frame feature types. There are G initial input feature maps, where one initial input feature map is generated for each type of video frame feature.

3. The method for optimizing video bitrate in narrowband environments based on convolutional neural networks according to claim 2, characterized in that, The specific implementation process of step S2 includes: Establish a two-dimensional coordinate system based on the macroblock pixel size in the video frame, and represent the pixel as... ,in, The x-axis value, The vertical axis value; Set the horizontal and vertical sliding offsets of the convolution kernel, respectively. and , forming sliding pixels ; The g-th initial input feature map is initialized at pixel point. The pixel value at that location is denoted as ; Perform intra-group convolution on the g-th initialized input feature map to obtain pixel points. Local enhancement pixel value at ,in, This indicates that the convolution kernel moves at the pixel point. The weight of the position, This represents the narrowband adaptation factor, used for dynamically adjusting the weights. The value, In response to video frames Real-time bandwidth at that time Indicates response to video frames Narrowband transmission delay is determined by real-time bandwidth. and narrowband transmission delay Linear quantization narrowband fit factor , , and All are linear coefficients after training. Represents pixels With sliding pixels Weight bias between them.

4. The method for optimizing video bitrate in narrowband environments based on convolutional neural networks according to claim 3, characterized in that, The specific implementation process of step S3 includes: In the case where an initial input feature map is generated corresponding to a certain video frame feature type, based on the video frame... Local enhancement pixel values Evaluate the image quality impact of the feature type of the g-th video frame. Where max{} and min{} are the maximum and minimum value functions, respectively; Construct the pixel state matrix of the video frame, where the row index of the pixel state matrix is ​​the pixel number. The column index is the video frame feature type. If image quality affects If the value is greater than or equal to the preset image quality impact threshold, then the pixel at row index [number] is [number]. The column index is a video frame feature type. The corresponding matrix position is set to 1 if the image quality impact is... If the value is less than the preset image quality impact threshold, then the pixel at the row index is... The column index is a video frame feature type. Set the corresponding matrix position to 0 to obtain the video frame. pixel state matrix .

5. The method for optimizing video bitrate in narrowband environments based on convolutional neural networks according to claim 4, characterized in that, The specific implementation process of step S4 includes: Obtain Synchronous Transmission Segments The pixel state matrix of each video frame in the video frame sequence constitutes the segment state matrix sample set, denoted as . ,in, Indicates synchronous transmission segment The total number of video frames contained in the video frame sequence; If the synchronous transmission segment As the current stage of bitrate optimization, the segments will be transmitted synchronously. All previous synchronous transmission segments are considered as historical phases; Analyze the similarity of video frame sequences between the current stage and any historical stage. ,in, This represents the h-th video frame. Represents video frames The pixel state matrix, This represents the e-th synchronous transmission segment. The corresponding generated fragment state matrix sample set, Indicates synchronous transmission segment The total number of video frames contained in the video frame sequence, e < n. For counting functions, Represents the pixel state matrix With pixel state matrix The number of 1s contained in the result of an inter-Boolean logical AND operation. Represents the pixel state matrix With pixel state matrix The number of 1s contained after a Boolean logical OR operation, where P is a preset similarity threshold. If the condition is met... Then let If not satisfied Then let .

6. The method for optimizing video bitrate in narrowband environments based on convolutional neural networks according to claim 5, characterized in that, The specific implementation process of step S5 includes: Synchronous transmission segments As the current stage of rate optimization, based on the order of synchronous transmission segments in the synchronous transmission time sequence set, the synchronous transmission segments are... The preceding synchronous transmission segments serve as a historical phase for bitrate optimization. A sliding window for the preset synchronous transmission segment is defined, with a sliding window size of d. Based on the sliding window scale d, generate the historical stage grouping set of the j-th sliding window. Used to record historical stages, and ; Calculate historical stage grouping set Average similarity between windows of all historical stages and the current stage ,in, Representing historical stage grouping sets The total number of synchronous transmission segments included.

7. The method for optimizing video bitrate in narrowband environments based on convolutional neural networks according to claim 6, characterized in that, The specific implementation process of step S6 includes: Two-dimensional data points are constructed based on the sliding window index and the average similarity of the windows. The similarity trend function is obtained. ,in, The slope The intercept; J represents the total number of sliding windows; Based on similarity trend function Assess trend stability , The overall average similarity of all sliding windows; Adjusting the sliding window scale d value outputs a trend stability. Maximize the sliding window size d and apply it to the similarity trend function. Update; Based on the updated similarity trend function ,like This indicates that the earlier historical stages are similar to the current stage, and the content shows a "retrospective similarity" trend. It predicts that the next stage will use a low bitrate for video compression. If the recent historical phase is similar to the current phase, and the content of the images is highly continuous, it is predicted that a high bitrate method will be used for video compression in the next phase after the current phase.

8. An optimization system for implementing the narrowband environment video bitrate optimization method based on convolutional neural networks as described in claim 1, characterized in that, The optimization system includes a storage device and a processor. The storage device includes a program for a narrowband video bitrate optimization method based on a convolutional neural network. When the program for the narrowband video bitrate optimization method based on a convolutional neural network is executed by the processor, the following steps are implemented: Configure the synchronization timing of video transmission, obtain the video frame sequence within the synchronized transmission segment, set the number of groups for the convolutional neural network feature extraction model according to the number of video frame feature types, and generate initial input feature maps corresponding to each type of feature for each video frame. A two-dimensional coordinate system for video frame pixels is established, the sliding offset of the convolution kernel is set, intra-group convolution operation is performed on each initial input feature map, and the convolution weights are dynamically adjusted in combination with real-time bandwidth and transmission delay to obtain the local enhanced pixel value of each pixel. The impact of different feature types on image quality is evaluated based on locally enhanced pixel values ​​in order to construct the pixel state matrix of video frames; The pixel state matrices of all video frames within the same synchronous transmission segment are integrated to form a sample set. This synchronous transmission segment is taken as the current stage, and the similarity between the current stage and the video frame sequences of all historical stages is calculated. Set the sliding window scale, group the historical stages, and calculate the average window similarity between each group of historical stages and the current stage; Based on the sliding window number and the corresponding average window similarity, a trend function is constructed, the sliding window scale is adjusted to optimize trend stability, and the trend of video content change is judged according to the characteristics of the trend function to indicate the video bitrate compression method in the next stage.

Images