Multimedia data compression method and device for narrowband transmission

Abstract

The application relates to the technical field of data compression, and discloses a multimedia data compression method and device for narrowband transmission. The method comprises the following steps: performing double-threshold contour extraction on an input image to obtain main contour data and secondary contour data; performing mixed-precision chain code coding on the main contour data and the secondary contour data to obtain compressed contour chain codes; performing parameterization decomposition and superframe coding on an input speech signal to obtain compressed speech parameters; allocating a transmission bandwidth according to the compressed speech parameters, and packing the compressed contour chain codes and the compressed speech parameters based on the transmission bandwidth to obtain a target transmission packet sequence. The application ensures the time synchronization and transmission reliability of image and speech data, solves the problems of multi-modal data asynchronous transmission and high bit error rate in a narrowband environment, and guarantees correct reconstruction and playing at a receiving end.

Description

Technical Field

[0001] This invention relates to the field of data compression technology, and in particular to a multimedia data compression method and apparatus for narrowband transmission. Background Technology

[0002] In extreme communication environments such as maritime rescue, polar exploration, and field operations, satellite communication systems are constrained by power limitations, channel fading, and equipment costs, with available bandwidth typically strictly limited to a narrow band range of 2-10 kbps. Under these extremely low bandwidth conditions, when simultaneous transmission of image and voice information is required to support emergency communication and remote collaboration, there is a lack of adaptability to the narrow band environment, and voice clarity is severely degraded at low bit rates. Existing technologies process and transmit image and voice data independently, lacking a joint optimization mechanism based on content semantics. Summary of the Invention

[0003] This invention provides a multimedia data compression method and apparatus for narrowband transmission. This invention ensures the time synchronization and transmission reliability of image and voice data, solves the problems of asynchronous transmission of multimodal data and high bit error rate in narrowband environment, and guarantees the correct reconstruction and playback at the receiving end.

[0004] In a first aspect, the present invention provides a multimedia data compression method for narrowband transmission, the multimedia data compression method for narrowband transmission comprising: Double-threshold contour extraction is performed on the input image to obtain main contour data and secondary contour data; The main contour data and the secondary contour data are encoded using mixed-precision chain code to obtain compressed contour chain code. The input speech signal is parametrically decomposed and superframe encoded to obtain compressed speech parameters; The transmission bandwidth is allocated according to the compressed speech parameters, and the compressed profile chain code and the compressed speech parameters are packaged based on the transmission bandwidth to obtain the target transmission packet sequence.

[0005] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, the step of performing dual-threshold contour extraction on the input image to obtain primary contour data and secondary contour data includes: The input image is preprocessed by Gaussian filtering to obtain a preprocessed image, and the gradient magnitude of the preprocessed image is calculated using the Sobel operator. Set a main contour threshold and a secondary contour threshold, classify pixels whose gradient magnitude is greater than the main contour threshold into the main contour point set, and classify pixels whose gradient magnitude is between the secondary contour threshold and the main contour threshold into the secondary contour point set. Calculate the first Euclidean distance between adjacent contour points in the main contour point set, and mark the first break position when the first Euclidean distance exceeds the target pixel; at the same time, calculate the second Euclidean distance between adjacent contour points in the secondary contour point set, and mark the second break position when the second Euclidean distance exceeds the target pixel. Based on the first fracture position, curve fitting is performed on the set of main contour points to obtain main contour data, and based on the second fracture position, curve fitting is performed on the set of secondary contour points to obtain secondary contour data.

[0006] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the step of performing mixed-precision chain code encoding on the main contour data and the secondary contour data to obtain compressed contour chain code includes: Calculate the corner response value, local information entropy, spatial distance, and semantic weight value of each contour point based on the main contour data and the secondary contour data; The importance score of each contour point is obtained by weighted summing of the corner response value, the local information entropy, the spatial distance and the semantic weight value. The mean and standard deviation of each contour point are calculated, and the cluster radius is calculated based on the mean and standard deviation. Contour points whose importance score is less than the score threshold and whose spatial distance is less than the clustering radius are grouped into the same cluster to obtain feature point clusters; Based on the feature point clusters, the main contour data and the secondary contour data are encoded using mixed-precision chain code to obtain compressed contour chain code.

[0007] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the step of grouping contour points whose importance scores are less than a score threshold and whose spatial distance is less than the clustering radius into the same cluster to obtain a feature point cluster includes: Calculate the minimum number of samples, traverse each contour point and count the number of neighboring contour points within the cluster radius. When the number of neighboring contour points is greater than or equal to the minimum number of samples, mark the core point. Select the core point as the cluster seed, and search for contour points whose importance score within the cluster radius is less than the score threshold to form an initial cluster group. Check whether the contour points in the initial cluster group are core points. If they are core points, continue to search for contour points in the neighborhood that meet the importance score less than the score threshold and add them to the initial cluster group to obtain the target cluster group. Feature point clusters are constructed by selecting the contour points with the highest importance scores from the target cluster groups.

[0008] In conjunction with the first aspect, in a fourth implementation of the first aspect of the present invention, the step of performing mixed-precision chain code encoding on the main contour data and the secondary contour data based on the feature point cluster to obtain compressed contour chain code includes: For each contour point in the main contour data and the secondary contour data, the local curvature value and the normalized bandwidth coefficient are calculated respectively, and the importance weight is set according to the feature point cluster. The chain code precision factor is calculated based on the local curvature value, the normalized bandwidth coefficient, and the importance weight. When the chain code precision factor is less than the precision threshold, the target encoding mode is selected as the 4-direction Freeman encoding mode; when the chain code precision factor is greater than or equal to the precision threshold, the target encoding mode is selected as the 8-direction Freeman encoding mode. The first three directions of each contour point are read in combination with the target encoding pattern, and a prediction result is generated based on the first three directions; The prediction result is compared with the actual chain code direction to obtain the prediction deviation, and a compressed profile chain code is generated based on the prediction deviation.

[0009] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present invention, the step of reading the first three directions of each contour point in conjunction with the target encoding pattern and generating a prediction result based on the first three directions includes: Each contour point is encoded based on the target encoding pattern to obtain a compressed contour chain code sequence; The first three directions of each contour point are extracted from the compressed contour chain code sequence and recorded as the first direction, the second direction and the third direction. A third-order Markov state vector is constructed based on the first direction, the second direction and the third direction. Check whether the first direction, the second direction, and the third direction in the third-order Markov state vector are completely identical. If the first three directions are completely identical, set the directions corresponding to the first three directions as the first prediction probability, set the remaining seven candidate directions as the second prediction probability, and generate the prediction result based on the first prediction probability and the second prediction probability. When there is a directional change in the third-order Markov state vector, the turning angle between adjacent directions is calculated, the turning probability is calculated based on the turning angle, and a prediction result is generated.

[0010] In conjunction with the first aspect, in the sixth implementation of the first aspect of the present invention, the step of parametric decomposition and superframe coding of the input speech signal to obtain compressed speech parameters includes: The input speech signal is segmented into frames to obtain a speech frame sequence, and the fundamental frequency parameter sequence of the speech frame sequence is extracted. Linear predictive coding analysis is performed on the speech frame sequence to obtain formant parameter combinations and spectral envelope coefficients; The fundamental frequency change of N consecutive speech frames in the fundamental frequency parameter sequence is detected. When the fundamental frequency change is less than the change threshold, these N speech frames are merged into a superframe unit to obtain superframe fundamental frequency encoded data. The formant parameter combination is mapped to a codebook index through nearest neighbor search, and the superframe fundamental frequency coding data, the codebook index, the spectral envelope coefficients, and the silence segment coding combination are combined into compressed speech parameters.

[0011] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, the step of allocating transmission bandwidth according to the compressed speech parameters and packaging the compressed contour chain code and the compressed speech parameters based on the transmission bandwidth to obtain a target transmission packet sequence includes: Hidden Markov Model is used to perform phoneme recognition and priority classification on the compressed speech parameters to obtain speech content priority; The bandwidth allocation coefficient is calculated based on the priority of the voice content to obtain the transmission bandwidth; Identify indicative words in the compressed speech parameters, and combine the indicative words to divide the compressed contour chain code into core packs and enhancement packs, and divide the compressed speech parameters into basic packs and extension packs; Generate a data packet combination based on the core package, the enhancement package, the base package, and the extension package; Based on the transmission bandwidth, a transmission strategy is allocated to each data packet combination to obtain the target transmission packet sequence.

[0012] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, the multimedia data compression method for narrowband transmission further includes: Establish a time synchronization mapping table for voice packets and image packets in the target transmission packet sequence; Data transmission is performed based on the time synchronization mapping table, and the channel bit error rate is calculated by statistically analyzing the acknowledgment information fed back by the receiving end. Based on the channel bit error rate matching error correction protection strategy, a narrowband synchronous transmission data stream is generated based on the error correction protection strategy.

[0013] Secondly, the present invention provides a multimedia data compression apparatus for narrowband transmission, the multimedia data compression apparatus for narrowband transmission comprising: The dual-threshold contour extraction module is used to extract the main contour data and the secondary contour data from the input image using dual-threshold contour extraction. A mixed-precision chain code encoding module is used to perform mixed-precision chain code encoding on the main contour data and the secondary contour data to obtain a compressed contour chain code. The parametric decomposition and superframe coding module is used to perform parametric decomposition and superframe coding on the input speech signal to obtain compressed speech parameters; The packaging module is used to allocate transmission bandwidth according to the compressed speech parameters, and package the compressed profile chain code and the compressed speech parameters based on the transmission bandwidth to obtain the target transmission packet sequence.

[0014] The technical solution provided by this invention achieves a flexible compression strategy based on bandwidth conditions through a dual-threshold contour hierarchical extraction mechanism, avoiding the information loss problem of traditional single-threshold methods. It also improves the accuracy and robustness of keypoint recognition by integrating multi-dimensional features. The hybrid-precision Freeman chain code encoding strategy overcomes the limitations of fixed-precision encoding, dynamically adjusting the encoding mode according to contour complexity and bandwidth conditions, and significantly reducing encoding redundancy by utilizing contour direction continuity. Through superframe merging and vector quantization, it achieves extremely high compression ratios while ensuring intelligibility, overcoming the limitations of independent image and speech processing, ensuring reliable transmission in narrowband environments, and solving the technical problem of multimedia data transmission under extreme bandwidth conditions of 2-10kbps. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a multimedia data compression method for narrowband transmission provided in an embodiment of this application; Figure 2 This is a schematic block diagram of a multimedia data compression device for narrowband transmission provided in an embodiment of this application. Detailed Implementation

[0017] 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, not all, of the embodiments of the present invention. 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.

[0018] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the described order. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change based on the actual situation.

[0019] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0020] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0021] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features described herein can be combined with each other.

[0022] Please see Figure 1 , Figure 1 A flowchart illustrating the multimedia data compression method for narrowband transmission provided in this application embodiment is shown below. Figure 1 As shown, the multimedia data compression method for narrowband transmission provided in this application includes: Step S100: Perform dual-threshold contour extraction on the input image to obtain main contour data and secondary contour data; Specifically, noise suppression and edge enhancement are performed on the input raw image. A Gaussian filter is applied to the original image by setting the parameters of the Gaussian kernel function. A 3×3 Gaussian kernel with a standard deviation σ of 0.8 is selected to obtain a preprocessed image with continuous spatial distribution and smooth details. The Sobel operator is used to calculate gradients on the preprocessed image, obtaining the gradient components of each pixel in the horizontal and vertical directions, and constructing a gradient magnitude image to reflect the local edge intensity distribution. A primary contour threshold and a secondary contour threshold are set for contour grading. The primary contour threshold distinguishes high-contrast structural boundaries, while the secondary contour threshold identifies medium-intensity detail edges. Pixels with gradient magnitudes greater than the primary threshold are classified as primary contour points, while pixels with gradient magnitudes between the secondary and primary thresholds are classified as secondary contour points. Connectivity analysis is performed on both the main contour point set and the secondary contour point set. The Euclidean distance between any pair of spatially adjacent contour points is calculated. If the first Euclidean distance between a pair of main contour points exceeds a set breakage threshold pixel distance, it is identified as the first breakage location and marked as a potential structural discontinuity point. Similarly, if the second Euclidean distance between a pair of secondary contour points exceeds a target pixel threshold, it is also marked as the second breakage location to identify discontinuous regions of detailed structures in the image. Cubic B-spline curve fitting based on the minimum energy criterion is performed on the breakage segments in the main contour point set containing the first breakage location. Curvature smoothing is achieved by minimizing the integral of the squared second derivative of the fitted curve, generating continuous main contour data. For the secondary contour point set containing the second breakage location, the same fitting strategy is used to reconstruct the complete contour curve, thus obtaining secondary contour data with continuity and compression characteristics, forming the image contour representation.

[0023] Step S200: Perform mixed-precision chain code encoding on the main contour data and the secondary contour data to obtain the compressed contour chain code; Specifically, corner response values, local information entropy, spatial distance, and semantic weight values ​​are calculated for each contour point based on the main contour data and secondary contour data. The corner response value measures the degree of abrupt changes in image structure; the response function value is obtained after constructing a structure tensor using the Harris algorithm to reflect the intensity of edge corners. Local information entropy calculates the uncertainty of information by statistically analyzing the probability density function of gray-level distribution within the neighborhood of the contour point. Spatial distance measures the shortest Euclidean distance between the current point and its neighboring contour points to measure its independence within the contour region. Semantic weight values ​​are assigned based on the semantic category of the image region where the contour is located; for example, contour points in face or text regions are given higher weights to enhance perceptual priority. The corner response values, local information entropy, spatial distance, and semantic weight values ​​are weighted and summed using fixed weight coefficients to obtain the importance score for each contour point. Statistical analysis is performed on the importance scores of all contour points to extract the mean and standard deviation. A dynamic clustering radius is constructed based on the mean and standard deviation, allowing the clustering range to adapt to image complexity. During clustering, all contour points with importance scores below a set threshold and spatial distances from the cluster center less than the cluster radius are grouped into the same cluster, forming multiple feature point clusters. Each feature point cluster represents several contour structures of similar importance. Based on these feature point clusters, mixed-precision chain code encoding is performed on the main contour data and secondary contour data. Eight-directional chain code is used in regions containing high-importance representative points to preserve detailed contour precision, while four-directional chain code is used in regions with flat edges or lower importance to save bits. This achieves a contour compression mechanism that adaptively selects encoding precision based on semantic structure and image geometric complexity, generating compressed contour chain code.

[0024] Step S300: Perform parametric decomposition and superframe coding on the input speech signal to obtain compressed speech parameters; Specifically, the input speech signal is processed by dividing it into frames with a fixed frame length of 20 milliseconds. Using an 8kHz sampling rate, each frame contains 160 sampling points, forming an ordered speech frame sequence. Fundamental frequency parameters are extracted frame by frame from the speech frame sequence. Fundamental frequency extraction employs autocorrelation function analysis supplemented by dynamic programming to improve the smoothness of the fundamental frequency trajectory, ensuring quantization within the frequency range of 80 to 400Hz. Linear predictive coding analysis is performed on the speech frame sequence to extract the parameters of the first three formants, namely the combinations F1, F2, and F3, which respectively include frequency and bandwidth values. Simultaneously, the linear prediction coefficients are compressed into a spectral envelope feature vector using discrete cosine transform, obtaining representative spectral envelope coefficients for spectral contour reconstruction. Inter-frame stability detection is performed throughout the fundamental frequency parameter sequence. When the fundamental frequency change of N consecutive speech frames is less than a preset change threshold, the fundamental frequency parameter sequence segment is determined to be in a stable speech state. These N speech frames are merged into a superframe unit, and only the first frame records the complete fundamental frequency parameters. At the same time, the differential change between the first and second frames is recorded to achieve differential compression, thus forming the superframe fundamental frequency encoded data. The vector quantization codebook obtained by training all formant parameter combinations using the LBG algorithm is used for nearest neighbor search and mapped to a fixed-length codebook index value to significantly reduce bit requirements. The obtained superframe fundamental frequency encoded data, codebook index, spectral envelope coefficients, and silence segment information are combined and encapsulated into a unified data structure. The silence segment is identified using energy detection and zero-crossing rate determination methods and is compressed using the start time and duration, thus forming the compressed speech parameter data.

[0025] Step S400: Allocate transmission bandwidth according to compressed speech parameters, and package compressed contour chain code and compressed speech parameters based on transmission bandwidth to obtain target transmission packet sequence.

[0026] Specifically, a speech content parsing mechanism based on a Hidden Markov Model (HMM) is employed to perform phoneme recognition on the obtained compressed speech parameters. By constructing a state transition probability matrix and an observation probability model, phoneme combinations constituting semantic content are identified from the speech frames. The recognition results are then matched with a predefined keyword dictionary to determine the urgency of the content expressed by the current speech segment, establishing a speech content priority level label. A transmission bandwidth allocation coefficient is calculated based on the response strategy corresponding to different priority levels. For example, when emergency words such as "help" or "fire" are identified, a higher proportion of bandwidth is allocated to the speech channel, while bandwidth is allocated to image chaining to enhance visual information fidelity when no key semantics are present. This yields dynamic transmission bandwidth parameters. The algorithm analyzes whether keywords in compressed speech parameters contain indicative components. Once indicative words are identified, the compressed contour chain code synchronized with the speech sequence in the current image frame undergoes semantic focusing processing. The compressed contour chain code is then divided into core packets and enhancement packets based on content importance. The core packet contains the main contour and key feature point data to ensure structural recognition, while the enhancement packet contains secondary contours and detail repair information. The compressed speech parameters are hierarchically divided into basic packets and extended packets. The basic packet includes the fundamental frequency and first formant parameters, while the extended packet includes higher-order formants and spectral envelope coefficients. Data packet combinations are constructed according to the priority of each packet type, giving core and basic packets higher scheduling priority. A bandwidth mapping scheduling strategy is applied to the data packet combinations based on the transmission bandwidth coefficient, allocating resources such as transmission order, rate, and error correction protection level to each combination to form the target transmission packet sequence.

[0027] In one specific embodiment, the process of performing step S100 may specifically include the following steps: The input image is preprocessed by Gaussian filtering to obtain a preprocessed image, and the gradient magnitude of the preprocessed image is calculated using the Sobel operator. Set a primary contour threshold and a secondary contour threshold. Pixels with gradient magnitudes greater than the primary contour threshold are classified into the primary contour point set, and pixels with gradient magnitudes between the secondary contour threshold and the primary contour threshold are classified into the secondary contour point set. Calculate the first Euclidean distance between adjacent contour points in the main contour point set. When the first Euclidean distance exceeds the target pixel, mark it as the first break position. At the same time, calculate the second Euclidean distance between adjacent contour points in the secondary contour point set. When the second Euclidean distance exceeds the target pixel, mark it as the second break position. Curve fitting is performed on the set of main contour points based on the first fracture location to obtain the main contour data, and curve fitting is performed on the set of secondary contour points based on the second fracture location to obtain the secondary contour data.

[0028] Specifically, a Gaussian filter is applied to the input image to reduce high-frequency noise interference and achieve smooth, continuous edge transitions. A 3×3 Gaussian kernel is used to perform a two-dimensional convolution operation on the image, with the kernel parameter σ set to 0.8. This removes isolated noise points while preserving the overall contour features of the image, resulting in a smoothly transitioned preprocessed image. Based on the preprocessed image, the Sobel gradient operator is introduced to perform edge enhancement operations in the horizontal and vertical directions. Local gradient vectors are constructed from the convolution responses along the X and Y axes, and their L2 norms are calculated. A real-valued gradient magnitude image is obtained at each pixel location, reflecting the intensity changes of the local structure of the image. Edge regions of different intensities in the image are processed hierarchically, with two static gradient thresholds set: a primary contour threshold and a secondary contour threshold, specifically set to 80 and 30, respectively, based on the gradient distribution statistics of the image under typical narrowband compression conditions. For any pixel in the image, if its gradient magnitude is greater than the main contour threshold, it is considered to be in a high-intensity edge structure region and has geometric significance in constituting the main contour of the image, and the pixel is included in the main contour point set. If the gradient magnitude of a pixel falls between the secondary threshold and the main threshold, it indicates that it belongs to a weak boundary, a blurred transition region, or a texture boundary region, and it is classified into the secondary contour point set. For pixels with gradient magnitudes lower than the secondary contour threshold, since their edge features are weak and their contribution is low, they are directly discarded to reduce the propagation and processing overhead of redundant data. Connectivity detection and structural integrity repair are performed on the main contour point set and the secondary contour point set, respectively. Point-by-point adjacency analysis is performed on the main contour point set. The spatial Euclidean distance between two adjacent contour points is calculated. When the first Euclidean distance exceeds the set target pixel threshold (e.g., 5 pixels), it is determined that there is a structural break, which is recorded as the first break position. The same method is applied to the secondary contour point set. If the second Euclidean distance between any two adjacent secondary contour points also exceeds the target pixel threshold, it is marked as the second break position, thus completing the localization and breakpoint identification of all contour break segments. Curve fitting was performed on the main contour point set containing the first fracture location and the secondary contour point set containing the second fracture location, respectively. A cubic B-spline function model based on minimum energy constraints was used for fitting. During curve construction, an energy functional was constructed and its value minimized to ensure a natural transition at both ends of the fracture segment, avoiding unnatural abrupt changes and achieving flexible completion of the contour structure and boundary consistency correction. In the fitting of the main contour, points with high curvature changes and main structural boundaries were preferentially retained to ensure that the image skeleton information was not weakened; while in the fitting of the secondary contour, the fitting interval of discontinuous segments was compressed to reduce unnecessary detail overhead. Finally, the main contour data and secondary contour data were obtained.

[0029] The process involves curve fitting of the main contour point set based on the first fracture location to obtain the main contour data, and curve fitting of the secondary contour point set based on the second fracture location to obtain the secondary contour data. This includes: analyzing the fracture type at both the first and second fracture locations, classifying fractures into three types based on fracture length: short-distance fractures, medium-distance fractures, and long-distance fractures. Fractures shorter than 3 pixels are marked as short-distance fractures, those between 3 and 8 pixels are marked as medium-distance fractures, and those longer than 8 pixels are marked as long-distance fractures, resulting in a fracture type classification. Based on this classification, appropriate repair strategies are selected: linear interpolation is used for short-distance fractures, quadratic Bézier curve fitting for medium-distance fractures, and cubic B-spline curve fitting for long-distance fractures. The control point coordinates and weight parameters of the fitted curves are calculated to obtain the repair curves. The parameter set is optimized by minimizing the energy function, which consists of three components: curve smoothness, endpoint constraint, and curvature continuity. The optimal parameter combination is solved using the Lagrange multiplier method to ensure a natural connection between the repaired curve and the original contour, resulting in an optimized repaired curve. The sampling point density of the optimized repaired curve is adjusted by dynamically determining the sampling interval based on the local curvature changes of the contour. The sampling density is increased in areas with large curvature changes and decreased in areas with small curvature changes. The discrete point sequence obtained from the sampling is inserted into the original contour point set to obtain the repaired continuous contour. The geometric consistency of the repaired continuous contour is verified by checking the tangent continuity and curvature continuity between the repaired segment and adjacent contour segments. When the continuity index is lower than the preset threshold, the repair parameters are readjusted and the repair process is repeated until the continuity requirements are met, resulting in the final main contour data and secondary contour data.

[0030] In one specific embodiment, the process of performing step S200 may specifically include the following steps: Calculate the corner response value, local information entropy, spatial distance, and semantic weight value of each contour point based on the main contour data and the secondary contour data; The importance score of each contour point is obtained by weighting and summing the corner response value, local information entropy, spatial distance and semantic weight value. The mean and standard deviation of each contour point are calculated, and the cluster radius is calculated based on the mean and standard deviation. Contour points whose importance scores are less than the score threshold and whose spatial distance is less than the cluster radius are grouped into the same cluster to obtain feature point clusters; Compressed contour chain codes are obtained by performing mixed-precision chain code encoding on the main contour data and secondary contour data based on feature point clusters.

[0031] Specifically, corner response values, local information entropy, spatial distance indices, and semantic weight values ​​are calculated for all contour points in the main contour data and secondary contour data. The corner response values ​​are constructed using the Harris corner detection algorithm. By defining a structure tensor matrix sensitive to image grayscale changes and combining it with the response function formula, corner intensity is calculated to identify locations of abrupt local edge changes. Higher corner intensity values ​​indicate that the contour point is more likely to be located in a geometrically salient region. Local information entropy reflects the statistical complexity of grayscale information surrounding the contour point, using the normalization of the pixel grayscale histogram within the 3×3 or 5×5 neighborhood of the contour point. Based on probability distribution, the information entropy formula is used to calculate the information content; a higher entropy value indicates richer information. The spatial distance index is calculated by taking the minimum Euclidean distance between a contour point and all other points in its contour set, thus measuring the spatial independence of the contour point and eliminating redundant points with high local clustering and lacking structural representativeness. Semantic weights are assigned based on the semantic region to which the contour point belongs in the entire image; higher weights are given to semantically significant regions such as faces, license plates, and text. For example, a face region is assigned 2.0, a text region 1.8, and a background or ordinary region 1.0. The corner response values, local information entropy, spatial distance, and semantic weights are normalized and then weighted and summed according to set weight coefficients to calculate a unique importance score for each contour point. After obtaining the importance scores of all contour points, statistical analysis is performed on the score sequence, calculating the mean μ and standard deviation σ of the score sequence to dynamically construct a density-adaptive cluster radius ε, set as ε = μ + 2σ. This allows the clustering algorithm to automatically adapt the clustering range according to the importance distribution of image contour points, avoiding over-aggregation or sparse splitting. During contour point clustering, the association of contour points is determined based on two constraints: the difference in importance scores between two points cannot exceed a set score threshold, and the spatial distance between the two points must be less than the cluster radius ε determined by the aforementioned statistical values. Only contour points that simultaneously meet both conditions are assigned to the same cluster, thus dividing the entire main contour data and secondary contour data into several feature point clusters with similar structures and importance characteristics. For each feature point cluster, the representative point with the highest score is selected as the encoding reference point, and the local contour structure represented by the encoding reference point is considered a critical region that cannot be lost in image encoding. During the encoding process, a precision scheduling mechanism based on feature point clusters is adopted. For contour segments containing representative points, they are automatically identified as high-fidelity regions and encoded using standard 8-direction Freeman chain codes to capture all directional changes, including diagonal directions, thereby restoring boundary contour details. For contour segments with low feature point scores, slow structural changes, or located in ordinary regions, 4-direction simplified chain codes are used to encode only the horizontal and vertical directions. By reducing the bit width of each contour point, the compression ratio is improved, which is suitable for bandwidth-constrained scenarios, resulting in compressed contour chain codes.

[0032] In one specific embodiment, the process of grouping contour points with importance scores less than a score threshold and spatial distances less than the cluster radius into the same cluster to obtain feature point clusters can specifically include the following steps: Calculate the minimum number of samples, traverse each contour point and count the number of neighboring contour points within the cluster radius. When the number of neighboring contour points is greater than or equal to the minimum number of samples, mark the core point. Select the core point as the cluster seed, and search for contour points whose importance score within the cluster radius is less than the score threshold to form the initial cluster group. Check whether the contour points in the initial cluster group are core points. If they are core points, continue to search for contour points in the neighborhood that meet the importance score less than the score threshold and add them to the initial cluster group to obtain the target cluster group. Feature point clusters are constructed by selecting the contour points with the highest importance scores from the target cluster groups.

[0033] Specifically, the minimum number of samples used in the density clustering process is calculated. This minimum number of samples is adaptively set based on the logarithmic function of the total number of contour points, for example, using the formula M= log2(N) +2, where N is the number of merged main contour points and secondary contour points. This calculation method dynamically adjusts the clustering sensitivity according to the image complexity, thereby avoiding excessive fragmentation of clustering when the number of points is large or the generation of pseudo-clustering structures when the number of points is sparse. All contour points are traversed point by point. For each contour point, the Euclidean distance between it and all other contour points is calculated, and the number of neighboring contour points contained within the given clustering radius ε (calculated from the mean and standard deviation of the importance score) is counted. If the number of neighbors is greater than or equal to the aforementioned minimum sample number, the current contour point is determined to be a structure center with significant local density, that is, it is marked as a core point. Each core point is used as the initial seed point for the clustering process. A neighborhood contour point search based on importance score constraints is performed within the radius. Starting from the core point, all contour points are retrieved within its ε-neighborhood, and the score of each retrieved point is evaluated. If its importance score is less than a set threshold, the retrieved point is considered to have feature similarity to the current core point, meeting the clustering homogeneity requirement, and is included in the current initial cluster group, forming an initial cluster unit containing a core point and its nearest low-score neighbors. A recursive clustering expansion mechanism is introduced: for each non-core point in the initial cluster group, it is checked whether it also meets the core point condition. If so, the same neighborhood low-score contour point search is performed again for the core point, and all points meeting the requirements are merged into the current cluster group. The target cluster group is constructed through iterative expansion. The contour point with the highest importance score in the target cluster group is selected as the representative point of that cluster, thus constructing a feature point cluster set.

[0034] In one specific embodiment, the process of performing mixed-precision chain code encoding on the main contour data and secondary contour data based on feature point clusters to obtain compressed contour chain code can specifically include the following steps: For each contour point in the main contour data and the secondary contour data, the local curvature value and the normalized bandwidth coefficient are calculated respectively, and the importance weight is set according to the feature point cluster. Chain code precision factor is calculated based on local curvature value, normalized bandwidth coefficient, and importance weight; When the chain code precision factor is less than the precision threshold, the target encoding mode is selected as the 4-direction Freeman encoding mode; when the chain code precision factor is greater than or equal to the precision threshold, the target encoding mode is selected as the 8-direction Freeman encoding mode. Combine the target encoding pattern to read the first three directions of each contour point, and generate prediction results based on the first three directions; The prediction bias is obtained by comparing the actual chain code direction with the prediction result, and a compressed profile chain code is generated based on the prediction bias.

[0035] Specifically, the local curvature value is calculated for each contour point to measure the curvature of the contour surrounding the point. The deformation intensity is inferred by observing the boundary change trend formed by several pixels before and after the contour point. If the contour change at a point is drastic, it is considered a high curvature point, indicating a complex morphology and significant structure in its region. Simultaneously, the current bandwidth resources of the system are assessed. The currently available bandwidth is normalized relative to the set maximum bandwidth limit to obtain a parameter representing communication capability. A higher value indicates more abundant transmission capability, allowing the encoder to use higher image fidelity; conversely, a lower value indicates a tight transmission load, prioritizing information compression to ensure basic communication needs. Combining the semantic hierarchy of the image structure, each contour point is assigned a corresponding weight value. Contour points belonging to key feature point clusters or semantically focused areas of the image (such as faces, text, or operation prompts) are given higher weights to emphasize their important role in structural reconstruction and semantic understanding, while points in ordinary areas maintain a basic weight. The local curvature value, normalized bandwidth coefficient, and importance weight are fused to generate a comprehensive discriminant value reflecting the encoding demand intensity of each contour point, namely the chain code precision factor. A higher chain code precision factor indicates that the current point is in a region with severe deformation, good transmission conditions, and significant semantics, thus requiring a higher precision representation during encoding. Conversely, a lower chain code precision factor indicates that the contour point is in a flat boundary or a region with less important structures, and that current system communication resources are limited, necessitating a more compact encoding mode to save bit resources. A fixed precision threshold is set. When the chain code precision factor of a point is less than the threshold, four-directional Freeman encoding is selected. Four-directional Freeman encoding only allows four basic directions: east, south, west, and north. It has a simple encoding structure, occupies very few bits, and is suitable for flat regions where compression efficiency is paramount. When the chain code precision factor is not less than the threshold, it indicates that the current point requires higher expressive power, and the encoding is switched to eight-directional Freeman encoding, supporting eight directions including the diagonals, suitable for complex shapes or semantically critical regions. Based on the target encoding pattern, a Markov prediction mechanism is introduced to compress and optimize the encoding process. The three previous historical encoded directions of the current contour point are read, and a prediction rule is constructed based on the assumption of continuity in the chain code sequence. When the three directions are completely consistent, that direction is used as the predicted value for the current direction, with a high empirically set prediction accuracy. When there is a direction reversal, a direction probability model is constructed based on the reversal angle to model the impact of the reversal pattern on the direction prediction. After predicting the direction of the current contour point through the above mechanism, it is compared with the true direction. If the prediction is correct, only one flag bit is needed, reducing bit usage; if the prediction is incorrect, an error flag bit is used with an additional error correction value. The error correction value is differentially encoded based on the direction offset to obtain the minimum encoding difference between the true and predicted directions. The prediction deviation codes of all contour points are sequentially concatenated to generate the compressed contour chain code.

[0036] In one specific embodiment, the process of performing the steps of reading the first three directions of each contour point in combination with the target encoding mode and generating prediction results based on the first three directions can specifically include the following steps: Each contour point is encoded based on the target encoding pattern to obtain a compressed contour chain code sequence; The first three directions of each contour point are extracted from the compressed contour chain code sequence and recorded as the first direction, the second direction and the third direction. A third-order Markov state vector is constructed based on the first direction, the second direction and the third direction. Check whether the first direction, second direction and third direction in the third-order Markov state vector are completely the same. If the first three directions are completely the same, set the directions corresponding to the first three directions as the first prediction probability, set the remaining seven candidate directions as the second prediction probability, and generate the prediction result based on the first prediction probability and the second prediction probability. When there is a directional change in the third-order Markov state vector, the turning angle between adjacent directions is calculated, the turning probability is calculated based on the turning angle, and the prediction result is generated.

[0037] Specifically, each contour point is encoded point-by-point under the determined target encoding mode. The target encoding mode is determined by the chain code precision factor. If the region to which the contour point belongs is a region with a simple structure, low semantic weight, or strict bandwidth constraints, a 4-direction encoding mode is adopted, using only the four basic directions of east, south, west, and north to express the contour direction. If the contour segment has significant geometric changes or high semantic importance, an 8-direction encoding mode is enabled, adding diagonal directions to improve directional resolution, thereby obtaining a more refined contour encoding result. Through the encoding process, the actual direction of each contour point is converted into a directional code, and a compressed contour chain code sequence is constructed to describe the coherent structural path of the main and secondary contours in the image. The first three directions of any contour point in its encoding order are extracted from the compressed contour chain code sequence and labeled as the first direction, the second direction, and the third direction, respectively. A third-order Markov state vector is constructed based on this to model the local directional evolution features of the contour point. The system determines whether the three directions contained in the state vector are completely consistent. If the first, second, and third directions are the same, it indicates that the contour segment exhibits a linear extension in space, and the continuous directions have not undergone abrupt changes. This conforms to the common edge straight line region characteristics in chain code sequences. In this case, the continuous direction is taken as the most likely prediction direction and set as the first prediction probability direction, i.e., it is given the highest prediction confidence. All other candidate directions that are not equal to this direction (the other three in the 4-direction mode and the other seven in the 8-direction mode) are uniformly classified into the second prediction probability range, which represents the set of directions with a lower probability of suboptimal. This forms a prediction candidate set based on probability hierarchy. On this basis, prediction results are generated through random selection, maximum likelihood, or a set strategy to provide a direction prediction reference for the current contour point. If a change in direction is detected in the third-order Markov state vector, meaning that the three historical directions are not completely consistent, it indicates that there is a change in direction, change in contour curvature, or boundary reconstruction in the current contour segment. At this time, the turning angle between any two adjacent directions in the state vector is calculated. The turning angle reflects the trend and bending magnitude of the contour direction in the local area. The more severe the turning, the greater the uncertainty of the current direction. Based on the turning angle, all possible candidate directions are analyzed, and a probability mapping mechanism for direction prediction is constructed. Directions with smaller turning angles are assigned higher turning probabilities, while the probabilities corresponding to directions with larger deviations are set to lower values, reflecting the tendency of edge continuity. The predicted direction is randomly selected from all directions according to the turning probability weight.

[0038] In one specific embodiment, the process of performing step S300 may specifically include the following steps: The input speech signal is segmented into frames to obtain a speech frame sequence, and the fundamental frequency parameter sequence of the speech frame sequence is extracted. Linear predictive coding analysis is performed on the speech frame sequence to obtain the formant parameter combination and spectral envelope coefficient; The fundamental frequency change of N consecutive speech frames in the fundamental frequency parameter sequence is detected. When the fundamental frequency change is less than the change threshold, these N speech frames are merged into a superframe unit to obtain superframe fundamental frequency encoded data. The formant parameter combination is mapped to the codebook index through nearest neighbor search, and the superframe fundamental frequency coding data, codebook index, spectral envelope coefficients and silence segment coding combination are combined into compressed speech parameters.

[0039] Specifically, the input speech signal is segmented in the time domain. By setting a fixed frame length, the continuous speech signal is divided into a sequence of equal-length speech frames, with a frame length of 20 milliseconds. Combined with a sampling rate of 8 kHz, each frame contains 160 sampling points, forming a framed dataset with a temporal structure. Fundamental frequency extraction is performed on each frame. Fundamental frequency parameters are key features reflecting pitch and vocal periodicity in speech perception. An autocorrelation function analysis method combined with an intra-frame peak detection strategy is used to identify the fundamental frequency position. A dynamic path search method is employed to eliminate instantaneous jumps and erroneous estimations, achieving temporal continuity extraction of the fundamental frequency sequence and generating a fundamental frequency parameter sequence. Linear predictive coding is applied to the speech frame sequence to extract formant parameters and spectral envelope features that reflect the speech spectral structure. The linear predictive coding obtains the linear prediction coefficients for each frame by constructing an intra-frame short-time linear prediction model. Then, the formant tracking algorithm is used to identify the combination of the frequencies and bandwidths of the first three formants as a compressed representation of the speech resonance features. At the same time, the spectral envelope information is extracted by applying discrete cosine transform to the prediction coefficients, and the first few dimensions of DCT coefficients are selected as spectral shape descriptors to achieve a compressed representation of the spectral shape of the speech frame. A stability test is performed on the fundamental frequency parameter sequence to identify whether the variation range of several consecutive frames (let's say N frames) at the fundamental frequency level is stable. Specifically, if the absolute value of the fundamental frequency parameter difference within consecutive frames is consistently less than a set threshold, the speech segment is considered to be in a pitch-stable stage. Therefore, these N frames are merged into a superframe unit. During encoding, only the complete fundamental frequency value is recorded for the first frame, while subsequent frames record the fundamental frequency difference relative to the previous frame as differential information. This uses a smaller number of bits to represent the similar structure between multiple frames, thereby compressing the transmission bandwidth required for the fundamental frequency sequence and obtaining superframe fundamental frequency encoded data with intra-frame continuity and inter-frame compactness. Simultaneously, for the formant parameter combinations obtained in each frame, a vector quantization strategy is introduced. By using a fixed-size codebook trained from historical speech data as the parameter mapping space, a nearest neighbor search operation is performed on the formant feature vector of the current frame, selecting the codebook vector closest to the current parameter combination and recording its index position as a compressed representation. Silence detection is performed on each frame. Frame segments with energy below a set threshold and an extremely low zero-crossing rate are identified as silent intervals. Silence segments are recorded using a compressed representation of the start timestamp and duration, avoiding redundant encoding of frames without speech information. All the compressed speech features, including superframe fundamental frequency coding data, formant parameter codebook indexes, spectral envelope coefficients, and silence segment coding results, are encapsulated in a unified data structure to form compressed speech parameter outputs for low-bitrate speech transmission scenarios.

[0040] The process involves establishing a formant parameter codebook with 256 entries. Formant parameter combinations are mapped to an 8-bit codebook index using nearest neighbor search. This includes: collecting formant parameter data from a large number of speech samples; extracting the frequency and bandwidth information of the first, second, and third formants; constructing a high-dimensional feature vector space; normalizing the feature vectors to eliminate the influence of dimensions to obtain a standardized formant feature dataset; using an improved LBG algorithm to perform cluster analysis on the standardized formant feature dataset; initializing 256 cluster centers; minimizing the sum of squared intra-cluster distances through iterative optimization; setting the convergence condition as a distortion change of less than 0.01% between two consecutive iterations to obtain an initial vector quantization codebook; and weighting the initial vector quantization codebook based on the semantic importance of the speech content, further optimizing the allocation of formant patterns containing vowel phonemes. A semantically weighted codebook is obtained by assigning lower weights to formant patterns containing consonant phonemes and readjusting the codebook center position using a weighted k-means algorithm. An adaptive codebook update mechanism is established, statistically analyzing the usage frequency of each codeword during actual transmission. When the usage frequency of a codeword is less than 10% of the average frequency, it is marked as a redundant codeword. When the minimum distance between a new formant pattern and an existing codeword is greater than an adaptive threshold, a new codeword is added. The effectiveness of the codebook is maintained through codeword replacement and addition / deletion operations, resulting in an adaptively optimized codebook. After preprocessing the formant parameter combinations, a nearest neighbor search is performed in the adaptively optimized codebook to calculate the Euclidean distance between the input feature vector and each codeword. The codeword with the smallest distance is selected as the quantization result. The position index of the codeword in the codebook is converted into an 8-bit binary code, resulting in a compressed formant parameter index.

[0041] In this embodiment, indicative words in compressed speech parameters are identified, and the compressed contour chain code is divided into core and enhancement packages based on these indicative words. The compressed speech parameters are divided into basic and extended packages. This includes: performing semantic content analysis on the compressed speech parameters; using a deep learning-based speech recognition model to extract lexical information and semantic features from the speech; establishing a mapping table between words and spatial orientation; identifying speech segments containing locative words, indicative words, and action words to obtain semantically annotated speech sequences; determining corresponding image regions of interest based on indicative words in the semantically annotated speech sequences; activating the weight of the central region of the image at the corresponding time when spatial indicative words such as "here" and "over there" are identified; activating the weight of the image region in the corresponding direction when locative words such as "above" and "below" are identified; establishing a dynamic mapping relationship between speech semantics and image spatial regions to obtain a multimodal semantic association matrix; and re-evaluating the importance of the compressed contour chain code based on the multimodal semantic association matrix, placing the segments located within the speech indicative regions... The importance weight of contour data is increased to 3.0, the weight of contour data in relevant regions is set to 2.0, and the weight of contour data in irrelevant regions is kept at 1.0. Contour data is divided into core packets and enhancement packets according to their weights, resulting in semantically driven image data grading. A precise temporal synchronization mechanism is established, assigning a timestamp accurate to the millisecond level to each speech phoneme. Image frames are segmented according to the speech phoneme boundaries to ensure that each image segment is strictly aligned with the corresponding speech phoneme in time. The mismatch between image frame rate and speech frame rate is handled by linear interpolation, resulting in a precisely synchronized multimodal data stream. A cross-modal consistency verification mechanism is constructed to check the semantic consistency between the speech indication content and the corresponding image region. When a mismatch is found between the speech description and the image content, the transmission priority of that time period is reduced. When a high match is found, the transmission priority is increased. The semantically driven image data grading and the partitioning strategy of the basic packet and the extension packet are dynamically adjusted through consistency scoring, resulting in a consistency-optimized multimodal data packet combination.

[0042] In one specific embodiment, the process of performing step S400 may specifically include the following steps: Hidden Markov Model is used to perform phoneme recognition and priority classification on compressed speech parameters to obtain speech content priority; The bandwidth allocation coefficient is calculated based on the priority of the voice content to obtain the transmission bandwidth; Identify indicative words in compressed speech parameters, and combine indicative words to divide compressed contour chain codes into core packets and enhancement packets, and divide compressed speech parameters into basic packets and extension packets; Generate data package combinations based on the core package, enhancement package, basic package, and extension package; Based on the transmission bandwidth, a transmission strategy is allocated to each data packet combination to obtain the target transmission packet sequence.

[0043] Specifically, a speech content recognition process based on a Hidden Markov Model (HMM) is adopted to perform structural analysis and semantic reconstruction of compressed speech parameters. By constructing an acoustic model and a state transition probability matrix, compressed speech parameters are mapped to corresponding phoneme sequences, and phoneme-vocabulary matching is performed in conjunction with a language model to reconstruct speech content at the temporal level. Fundamental frequency information, formant features, and spectral envelope coefficients contained in the compressed speech data are used as observation vectors. A maximum likelihood path decoding algorithm is used to analyze the state transition paths of speech frames, and a keyword dictionary is combined for precise matching and semantic judgment. If keywords defined as high urgency, such as "help," "fire," "someone," and "danger," are detected in the identified phoneme sequences, the speech content is marked as the highest priority. If prompting phrases such as "please check," "attention," and "report" are identified, they are classified as medium priority content, while ordinary declarative sentences are classified as the lowest priority content, thus constructing speech content priority labels. Based on the above recognition results, a bandwidth allocation coefficient for scheduling transmission resources between audio and video channels is calculated. The bandwidth allocation coefficient is dynamically adjusted according to the speech content priority; the higher the priority, the larger the proportion of bandwidth allocated to the speech channel. For example, when emergency help keywords are detected in the speech, the transmission weight of the speech information is automatically increased, allocating most of the available transmission bandwidth to the speech parameter channel to ensure the clarity and timeliness of the help information. When the content only contains neutral descriptions or non-critical information, the bandwidth is allocated in a balanced manner or tilted towards the image contour channel, thereby optimizing the efficiency of transmission bandwidth utilization and obtaining the target transmission bandwidth allocation value for the current time period. The system performs semantic refinement processing on the keywords in the compressed speech parameters, identifying whether they contain indicative words with image-directive characteristics, such as spatial reference phrases, and matches such words with the time labels of the image contours, thereby increasing the transmission priority of image data within the corresponding time period. In this process, the compressed contour chain code is structurally decomposed, defining the content containing key feature points and main contour segments as the core packet to ensure the integrity of the image structure; while the chain code data containing secondary details, texture edges, or low weights are defined as enhancement packets for additional transmission when bandwidth is sufficient. Features constituting the semantic backbone of compressed speech parameters, such as fundamental frequency, first formant, and silence structure, are defined as basic packets. Higher-order formant combinations and spectral envelope information are used as extension packets to supplement audio quality enhancement data when the image channel is under low pressure. Based on the combination and encapsulation of core packets, enhancement packets, basic packets, and extension packets in both image and speech dimensions, a multimodal data packet set is constructed, and a multi-level transmission queue is established according to priority. Based on the transmission bandwidth allocation ratio, rate scheduling and error correction level settings are applied to the data packet combinations, dynamically allocating bit rate, time slice, and error correction protection level to generate the target transmission packet sequence within the available bandwidth of the overall network environment.

[0044] In one specific embodiment, the multimedia data compression method for narrowband transmission further includes the following steps: Establish a time synchronization mapping table between voice packets and image packets in the target transmission packet sequence; Data transmission is performed based on a time synchronization mapping table, and the channel bit error rate is calculated by analyzing the acknowledgment information fed back from the receiving end. A channel bit error rate matching error correction protection strategy is used, and a narrowband synchronous transmission data stream is generated based on the error correction protection strategy.

[0045] Specifically, during the target transmission packet sequence construction phase, each voice packet and each image packet is assigned an independent timestamp field, and the time management module performs frame-level alignment of all compressed voice parameters and compressed image contour chain codes. Voice packets use fixed frame rate encoding at 20 milliseconds per frame, therefore each voice packet contains an absolute timestamp in milliseconds. Image packets generate timestamps based on the image frame acquisition time and contour extraction batch time. By comparing the image frame time and the voice frame time, a correspondence or approximate mapping relationship between image packets and voice packets is established, and this correspondence is stored as a time synchronization mapping table. Data transmission is based on the time synchronization mapping table, and a timestamp and sequence identifier field is embedded in each data packet in conjunction with the transmission protocol. A sending window and acknowledgment mechanism are configured for each type of data packet to ensure that data packets arrive at the receiving end on time and in order in the narrowband channel. During transmission, acknowledgment feedback information from the receiving end is continuously monitored, including whether the data packet with the corresponding timestamp was successfully received, whether content errors were detected, and whether the error correction module was triggered for reconstruction. The transmission error rate of the communication link within a certain time window is calculated based on the acknowledgment feedback information. The transmission bit error rate (BER) is estimated as the proportion of erroneous data packets to the total number of transmitted packets, and smoothed by incorporating the time range of the feedback window to form a dynamic BER indicator for the current channel state. The channel BER is input into the error correction strategy matching module, which matches the corresponding level of error correction protection strategy based on the BER range. When the BER is below the set lower limit, the lightest error detection mechanism, such as parity checking, is selected, providing only minimal overhead and basic error detection capabilities. When the BER is within the intermediate tolerance range, a medium-level error correction strategy is activated, such as an encoding scheme capable of correcting single-bit errors and detecting double-bit errors, balancing protection capabilities with bandwidth usage. When the BER exceeds the set upper limit, indicating a deteriorating channel environment and increased risk of data loss, the error correction level is automatically increased, employing advanced encoding mechanisms with strong error correction capabilities, such as Hamming codes or BCH codes, to ensure the integrity of critical voice data and image master contour chain codes. Based on the error correction protection strategy, according to the semantic importance, synchronization mapping relationship and transmission priority of each type of data packet, corresponding error correction parameters are configured for the voice base packet, image core packet, voice extension packet and image enhancement packet, respectively. All data packets are repackaged, and error correction redundant data is added. Then, they are scheduled and sorted according to the time synchronization mapping table to output the narrowband synchronous transmission data stream.

[0046] The multimedia data compression method for narrowband transmission also includes a real-time channel quality adaptive adjustment step: During the real-time transmission of the target transmission packet sequence, the receiver feedback signal is monitored, the number of acknowledgment packets and error packets is counted, and the real-time bit error rate and signal-to-noise ratio (SNR) parameters within the current transmission window are calculated to obtain the channel quality assessment result; the compression depth parameter is adjusted based on the SNR value in the channel quality assessment result; when the SNR is greater than 15 dB, the compression ratio of the compression profile chain code and the compression speech parameter is reduced to ensure transmission quality; when the SNR is less than 10 dB, the compression ratio is increased to ensure transmission success rate, resulting in an adaptive compression parameter configuration; the error correction coding strength level is selected based on the real-time bit error rate in the channel quality assessment result. When the bit rate is less than 0.01%, parity check codes are selected to reduce encoding overhead; when the bit error rate is between 0.01% and 0.1%, BCH codes are selected to provide moderate protection; and when the bit error rate is greater than or equal to 0.1%, Hamming codes are selected to provide the strongest protection, resulting in a dynamic error correction protection strategy. Adaptive compression parameter configuration is applied to the generation process of subsequent transmission packets, while the dynamic error correction protection strategy is applied to the encoding protection of data packets, achieving real-time adaptive adjustment of compression depth and error correction strength. A channel quality prediction model is established, predicting the channel state of the next window based on the bit error rate variation trend and signal-to-noise ratio fluctuation law of multiple consecutive transmission windows, and adjusting the adaptive compression parameter configuration and dynamic error correction protection strategy in advance, resulting in a predictive adaptive transmission control mechanism.

[0047] Please see Figure 2 , Figure 2 A schematic block diagram of the structure of a multimedia data compression apparatus for narrowband transmission provided in the embodiments of this application, as shown below. Figure 2 As shown, a multimedia data compression device for narrowband transmission includes: The dual-threshold contour extraction module 210 is used to perform dual-threshold contour extraction on the input image to obtain main contour data and secondary contour data; The mixed-precision chain code encoding module 220 is used to encode the main contour data and the secondary contour data in a mixed-precision chain code to obtain a compressed contour chain code. The parametric decomposition and superframe coding module 230 is used to perform parametric decomposition and superframe coding on the input speech signal to obtain compressed speech parameters; Packaging module 240 is used to allocate transmission bandwidth according to compressed speech parameters, and to package compressed contour chain code and compressed speech parameters based on transmission bandwidth to obtain target transmission packet sequence.

[0048] Through the collaborative efforts of the aforementioned components, and by setting primary and secondary contour thresholds, the transmission contour level can be flexibly selected based on real-time bandwidth conditions. Priority is given to transmitting the primary contour when bandwidth is limited, avoiding the information loss problem inherent in traditional single-threshold methods during extreme compression. Contour continuity is ensured through break detection and curve fitting. The importance of contour points is evaluated across four dimensions: corner response value, local information entropy, spatial independence distance, and semantic weight. Compared to single feature descriptors, this approach offers higher accuracy and robustness. Adaptive density clustering effectively reduces redundant feature points, providing high-quality keypoint constraints for subsequent encoding. Dynamic selection of 4- or 8-directional encoding modes based on local contour geometry and real-time bandwidth conditions overcomes the limitations of traditional fixed-precision encoding, maximizing compression efficiency while preserving important contour details, making it particularly adaptable to dynamic changes in narrowband transmission environments. Predictive encoding utilizing the continuity features of contour directions significantly reduces encoding redundancy, significantly reducing bit overhead compared to direct encoding methods, improving transmission efficiency under extremely low bandwidth conditions, and maintaining the integrity of contour information. By employing superframe merging coding and vector quantization techniques, an extremely high compression ratio is achieved while maintaining speech intelligibility. Silence detection further optimizes transmission efficiency, making it more adaptable to the strict constraints of narrowband transmission compared to traditional fixed-parameter encoders. Based on semantic analysis of speech content, the transmission bandwidth for image and speech is dynamically allocated, overcoming the limitations of traditional independent processing and ensuring priority transmission of critical information in emergency situations, thus achieving intelligent collaborative transmission of multimedia data. Unified timestamps and adaptive error correction coding ensure time synchronization and transmission reliability of image and speech data, solving the problems of asynchronous transmission of multimodal data and high bit error rates in narrowband environments, guaranteeing correct reconstruction and playback at the receiving end.

[0049] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0050] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multimedia data compression method for narrowband transmission, characterized in that, include: Double-threshold contour extraction is performed on the input image to obtain main contour data and secondary contour data; The main contour data and the secondary contour data are encoded using mixed-precision chain code to obtain compressed contour chain code. The input speech signal is parametrically decomposed and superframe encoded to obtain compressed speech parameters; The transmission bandwidth is allocated according to the compressed speech parameters, and the compressed profile chain code and the compressed speech parameters are packaged based on the transmission bandwidth to obtain the target transmission packet sequence.

2. The multimedia data compression method for narrowband transmission according to claim 1, characterized in that, The process of performing dual-threshold contour extraction on the input image to obtain primary contour data and secondary contour data includes: The input image is preprocessed by Gaussian filtering to obtain a preprocessed image, and the gradient magnitude of the preprocessed image is calculated using the Sobel operator. Set a main contour threshold and a secondary contour threshold, classify pixels whose gradient magnitude is greater than the main contour threshold into the main contour point set, and classify pixels whose gradient magnitude is between the secondary contour threshold and the main contour threshold into the secondary contour point set. Calculate the first Euclidean distance between adjacent contour points in the main contour point set, and mark the first break position when the first Euclidean distance exceeds the target pixel; at the same time, calculate the second Euclidean distance between adjacent contour points in the secondary contour point set, and mark the second break position when the second Euclidean distance exceeds the target pixel. Based on the first fracture position, curve fitting is performed on the set of main contour points to obtain main contour data, and based on the second fracture position, curve fitting is performed on the set of secondary contour points to obtain secondary contour data.

3. The multimedia data compression method for narrowband transmission according to claim 1, characterized in that, The step of performing mixed-precision chain code encoding on the main contour data and the secondary contour data to obtain compressed contour chain code includes: Calculate the corner response value, local information entropy, spatial distance, and semantic weight value of each contour point based on the main contour data and the secondary contour data; The importance score of each contour point is obtained by weighted summing of the corner response value, the local information entropy, the spatial distance and the semantic weight value. The mean and standard deviation of each contour point are calculated, and the cluster radius is calculated based on the mean and standard deviation. Contour points whose importance score is less than the score threshold and whose spatial distance is less than the clustering radius are grouped into the same cluster to obtain feature point clusters; Based on the feature point clusters, the main contour data and the secondary contour data are encoded using mixed-precision chain code to obtain compressed contour chain code.

4. The multimedia data compression method for narrowband transmission according to claim 3, characterized in that, The step of grouping contour points whose importance score is less than the score threshold and whose spatial distance is less than the cluster radius into the same cluster to obtain feature point clusters includes: Calculate the minimum number of samples, traverse each contour point and count the number of neighboring contour points within the cluster radius. When the number of neighboring contour points is greater than or equal to the minimum number of samples, mark the core point. Select the core point as the cluster seed, and search for contour points whose importance score within the cluster radius is less than the score threshold to form an initial cluster group. Check whether the contour points in the initial cluster group are core points. If they are core points, continue to search for contour points in the neighborhood that meet the importance score less than the score threshold and add them to the initial cluster group to obtain the target cluster group. Feature point clusters are constructed by selecting the contour points with the highest importance scores from the target cluster groups.

5. The multimedia data compression method for narrowband transmission according to claim 4, characterized in that, The step of performing mixed-precision chain code encoding on the main contour data and the secondary contour data based on the feature point cluster to obtain compressed contour chain code includes: For each contour point in the main contour data and the secondary contour data, the local curvature value and the normalized bandwidth coefficient are calculated respectively, and the importance weight is set according to the feature point cluster. The chain code precision factor is calculated based on the local curvature value, the normalized bandwidth coefficient, and the importance weight. When the chain code precision factor is less than the precision threshold, the target encoding mode is selected as the 4-direction Freeman encoding mode; when the chain code precision factor is greater than or equal to the precision threshold, the target encoding mode is selected as the 8-direction Freeman encoding mode. The first three directions of each contour point are read in combination with the target encoding pattern, and a prediction result is generated based on the first three directions; The prediction result is compared with the actual chain code direction to obtain the prediction deviation, and a compressed profile chain code is generated based on the prediction deviation.

6. The multimedia data compression method for narrowband transmission according to claim 5, characterized in that, The step of reading the first three directions of each contour point in conjunction with the target encoding pattern and generating a prediction result based on the first three directions includes: Each contour point is encoded based on the target encoding pattern to obtain a compressed contour chain code sequence; The first three directions of each contour point are extracted from the compressed contour chain code sequence and recorded as the first direction, the second direction and the third direction. A third-order Markov state vector is constructed based on the first direction, the second direction and the third direction. Check whether the first direction, the second direction, and the third direction in the third-order Markov state vector are completely identical. If the first three directions are completely identical, set the directions corresponding to the first three directions as the first prediction probability, set the remaining seven candidate directions as the second prediction probability, and generate the prediction result based on the first prediction probability and the second prediction probability. When there is a directional change in the third-order Markov state vector, the turning angle between adjacent directions is calculated, the turning probability is calculated based on the turning angle, and a prediction result is generated.

7. The multimedia data compression method for narrowband transmission according to claim 1, characterized in that, The process of parametric decomposition and superframe coding of the input speech signal to obtain compressed speech parameters includes: The input speech signal is segmented into frames to obtain a speech frame sequence, and the fundamental frequency parameter sequence of the speech frame sequence is extracted. Linear predictive coding analysis is performed on the speech frame sequence to obtain formant parameter combinations and spectral envelope coefficients; The fundamental frequency change of N consecutive speech frames in the fundamental frequency parameter sequence is detected. When the fundamental frequency change is less than the change threshold, these N speech frames are merged into a superframe unit to obtain superframe fundamental frequency encoded data. The formant parameter combination is mapped to a codebook index through nearest neighbor search, and the superframe fundamental frequency coding data, the codebook index, the spectral envelope coefficients, and the silence segment coding combination are combined into compressed speech parameters.

8. The multimedia data compression method for narrowband transmission according to claim 1, characterized in that, The step of allocating transmission bandwidth according to the compressed speech parameters, and packaging the compressed contour chain code and the compressed speech parameters based on the transmission bandwidth to obtain the target transmission packet sequence includes: Hidden Markov Model is used to perform phoneme recognition and priority classification on the compressed speech parameters to obtain speech content priority; The bandwidth allocation coefficient is calculated based on the priority of the voice content to obtain the transmission bandwidth; Identify indicative words in the compressed speech parameters, and combine the indicative words to divide the compressed contour chain code into core packs and enhancement packs, and divide the compressed speech parameters into basic packs and extension packs; Generate a data packet combination based on the core package, the enhancement package, the base package, and the extension package; Based on the transmission bandwidth, a transmission strategy is allocated to each data packet combination to obtain the target transmission packet sequence.

9. The multimedia data compression method for narrowband transmission according to claim 1 or 8, characterized in that, The multimedia data compression method for narrowband transmission further includes: Establish a time synchronization mapping table for voice packets and image packets in the target transmission packet sequence; Data transmission is performed based on the time synchronization mapping table, and the channel bit error rate is calculated by statistically analyzing the acknowledgment information fed back by the receiving end. Based on the channel bit error rate matching error correction protection strategy, a narrowband synchronous transmission data stream is generated based on the error correction protection strategy.

10. A multimedia data compression device for narrowband transmission, characterized in that, A multimedia data compression method for narrowband transmission as described in any one of claims 1-9, comprising: The dual-threshold contour extraction module is used to extract the main contour data and the secondary contour data from the input image using dual-threshold contour extraction. A mixed-precision chain code encoding module is used to perform mixed-precision chain code encoding on the main contour data and the secondary contour data to obtain a compressed contour chain code. The parametric decomposition and superframe coding module is used to perform parametric decomposition and superframe coding on the input speech signal to obtain compressed speech parameters; The packaging module is used to allocate transmission bandwidth according to the compressed speech parameters, and package the compressed profile chain code and the compressed speech parameters based on the transmission bandwidth to obtain the target transmission packet sequence.

Images