A high-multiple video compression method integrated with artificial intelligence and capable of transmission under a narrowband network

By employing AI-powered intelligent compression and reconstruction processes, the problem of poor video quality and lack of smoothness in narrowband IoT and long-distance wireless transmission has been solved, enabling high-definition video transmission under extremely low bandwidth conditions.

CN123123570APending Publication Date: 2026-11-17程伟胤 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611267940.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing video compression standards cannot meet the ultra-narrow bandwidth requirements of narrowband IoT and long-distance wireless transmission scenarios, resulting in problems such as long transmission distances, severe signal attenuation, poor image quality, and lack of smoothness.

Method used

The AI-powered intelligent compression process employs RGB-YUV color space conversion, a dynamic foreground change rate determination model, lightweight CNN+Transformer hybrid encoding, differential quantization encoding, and AI reconstruction enhancement at the receiver to achieve high-compression and reconstruction of the video.

Benefits of technology

It achieves high-definition and smooth video transmission under extremely low bandwidth, increases bandwidth utilization by more than a thousand times, and ensures stable image quality, thus solving the video transmission bottleneck under narrowband networks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123570A_ABST
    Figure CN123123570A_ABST
Patent Text Reader

Abstract

The application discloses a high-multiple video compression method integrated with artificial intelligence and capable of transmission under a narrowband network, and belongs to the technical field of video compression and artificial intelligence communication. In view of the technical difficulties that the existing traditional video coding relies on pixel-level full-quantity transmission, the compression ratio is limited, and cannot adapt to the kbps-level narrowband Internet of Things, the bandwidth is limited in the scene of long-distance wireless transmission, the application breaks through the bandwidth bottleneck of traditional video coding, and can stably transmit high-definition video under 1-50 kbps ultra-narrow bandwidth, and can balance the extreme compression ratio, core picture fidelity and video fluency, and is suitable for various low-bandwidth scenes such as field inspection, remote monitoring, satellite narrowband transmission, and has high engineering application value and innovation. The high-multiple video compression method integrated with artificial intelligence and capable of transmission under a narrowband network provided by the application is used for the video transmission scene with extremely limited bandwidth such as narrowband Internet of Things, long-distance wireless transmission, field satellite inspection and low-power consumption monitoring.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence video coding technology, and in particular relates to a high-compression video method that incorporates artificial intelligence and can be transmitted over narrowband networks. Background Technology

[0002] Wireless data transmission, with its advantages of flexible deployment, wide coverage, no need for wiring, and low overall cost, has been widely used in security monitoring, field inspection, traffic monitoring, satellite communication and other scenarios.

[0003] However, wireless transmission has inherent technical defects: the farther the transmission distance, the more severe the signal attenuation, environmental interference, and obstruction caused by obstacles. In order to ensure the success rate of transmission, it is necessary to adopt technologies such as redundant coding and spread spectrum transmission to further compress the effective transmission bandwidth.

[0004] Especially for narrowband IoT and long-distance wireless transmission scenarios, the effective bandwidth is only a few kbps to tens of kbps, while the bandwidth required for normal transmission of conventional 720P / 1080P video is hundreds of kbps to tens of Mbps, a difference of orders of magnitude. Existing mainstream video compression standards (H.264, H.265, H.266) are all based on pixel-level transform coding, inter-frame prediction, and entropy coding to achieve compression, which relies on the transmission of complete pixel streams. The compression ratio has a theoretical upper limit and cannot adapt to the needs of ultra-narrow bandwidth transmission.

[0005] Therefore, it is necessary to provide a new high-compression video method that incorporates artificial intelligence and can be transmitted over narrowband networks to solve the above-mentioned technical problems. Summary of the Invention

[0006] The technical problem solved by this invention is to provide a high-compression video method that integrates artificial intelligence and can be transmitted over narrowband networks for video transmission scenarios with extremely limited bandwidth, such as narrowband Internet of Things, long-distance wireless transmission, field satellite inspection, and low-power monitoring.

[0007] To solve the above-mentioned technical problems, the high-compression video method that incorporates artificial intelligence and can be transmitted over narrowband networks provided by the present invention includes several parts: AI intelligent compression processing flow at the sending end, low bit rate transmission flow over narrowband networks, and AI intelligent reconstruction and enhancement flow at the receiving end. The AI ​​intelligent compression processing flow at the sending end includes: performing AI noise reduction preprocessing and RGB-YUV color space conversion on the acquired raw video frames to remove image noise and reduce the color data volume; A dynamic foreground change rate determination model is constructed based on the three-frame difference algorithm and the OTSU adaptive threshold segmentation algorithm. The key frames and prediction frames are dynamically and adaptively selected according to the intensity of the foreground motion in the video. The prediction frames are directly discarded and only the key frames are retained for transmission. A lightweight CNN+Transformer hybrid coding AI model is used to accurately segment multiple regions of keyframes, including foreground, background, and micro-dynamic background residuals. Low-dimensional global semantic feature vectors are extracted from static background areas and normalized, differentiated quantized, and CABAC lossless arithmetic coding is performed to generate ultra-low bitrate background data streams. For the foreground areas of interest to users are compressed using high-rate video bitstreams to generate foreground data streams. The two data streams are combined and transmitted through a kbps-level narrowband wireless network. The narrowband network low bit rate transmission process is adapted to a bandwidth range of 1kbps~50kbps. It replaces the traditional full pixel transmission with semantic feature compression to avoid the technical defects of insufficient bandwidth in narrowband networks. The AI ​​intelligent reconstruction and enhancement process at the receiving end includes: de-decoding the combined bitstream, inverse quantizing to restore the background semantic feature vector, reconstructing the basic background image through a multi-level deconvolution upsampling decoder, and generating a natural dynamic background by combining micro-dynamic residual features; decoding to restore the high-definition foreground image, and using a hybrid Poisson fusion algorithm and LAB color space calibration to complete seamless stitching of the foreground and background; improving the image detail clarity through AI super-resolution upsampling, interpolating and generating discarded prediction frames based on optical flow estimation technology to complete the temporal continuity of the video, and finally outputting a high-definition, smooth, and color-consistent complete video; This method breaks through the limitations of traditional video encoding's pixel-level transmission by using a hybrid innovative architecture that combines AI semantic-level extreme background compression, traditional bitstream foreground fidelity compression, dynamic adaptive keyframe selection, and end-to-end AI-enhanced reconstruction. This enables stable transmission of high-fidelity video under narrowband and extremely low bandwidth conditions.

[0008] As a further aspect of the present invention, the specific implementation steps of the dynamic adaptive keyframe filtering are as follows: Select three consecutive time-series images , , The formula for calculating the bidirectional inter-frame absolute difference image is as follows: ; ; ; in, This is the difference image between the current frame and the previous frame. This is a difference image between the current frame and the next frame. This is the final difference image after fusion; In the difference image D t The Otsu's maximum inter-class variance algorithm is used to iterate through the grayscale thresholds and solve for the optimal segmentation threshold. The formula is as follows: ; in, The maximum number of gray levels in the image. Let T be the inter-class variance of foreground and background pixels corresponding to the threshold. The optimal segmentation threshold; The foreground motion region is obtained based on optimal threshold binarization segmentation, and the rate of change of the foreground region is calculated. The formula is as follows: ; in, for Frame binarization of foreground mask image, for Frame binarization of foreground mask image, This represents the total number of pixels in the image. Preset motion thresholds M1 (lower limit of rate of change) and M2 (upper limit of rate of change) with a value range of 0~1 are adaptively assigned according to the scenario: static monitoring scenario M1=0.015, M2=0.12; traffic road scenario M1=0.03, M2=0.18; handheld inspection scenario M1=0.05, M2=0.2; dense indoor scenario M1=0.04, M2=0.16; field satellite inspection scenario M1=0.01, M2=0.1; 2-5, dynamically determine frame type: When M2 is greater than M2, the current frame is directly determined as a keyframe; M1 < When M2 is less than or equal to 2, a keyframe is selected every 30 frames. When M1 is less than or equal to 1, a key frame is selected every 60 frames, and all other frames are judged as prediction frames and discarded directly.

[0009] As a further aspect of the present invention, the network structure and semantic feature extraction process of the lightweight CNN+Transformer hybrid encoder include: a shallow CNN texture extraction module: setting a four-layer progressive convolutional structure, The first layer converts 3 channels to 64 channels, uses 3×3 convolution, and has a stride of 2, reducing the resolution to half. The second layer converts 64 channels to 128 channels, uses 3×3 convolution, and has a stride of 2, reducing the resolution to 1 / 4. The third layer converts 128 channels to 256 channels, uses 3×3 convolution, and has a stride of 2, reducing the resolution to 1 / 8. The fourth layer converts 256 channels to 384 channels, performs a 3×3 convolution with a stride of 2, reduces the resolution to 1 / 16, and outputs the bottom-level spatial feature map F∈R. Mid-layer spatiotemporal fusion module: fuses the low-level features of the current frame with the historical features of the previous frame, models the slow changes in the scene and the micro-motions of objects, and outputs the mid-layer fusion feature F; High-level semantic projection module: It compresses the spatiotemporal dimension through global mean pooling, and generates the final global semantic feature vector z through two fully connected layers and the GELU activation function, as shown in the following formula: ; in, , The weight matrix is ​​the fully connected component. , For bias vectors, This is a global mean pooling operation. This is the GELU activation function.

[0010] As a further aspect of the present invention, the feature vector differential quantization encoding process includes: Perform global normalization on the extracted global semantic feature vector to scale the values ​​of each dimension to the [0,1] interval; Differentiated bit depth quantization: Ordinary static background feature vectors use 256-dimensional vectors and 8-bit depth quantization; micro-dynamic background residual feature vectors such as leaf swaying and water ripples use 64-dimensional vectors and 6-bit depth quantization. The quantized integer sequence is subjected to lossless compression using CABAC context-adaptive binary arithmetic coding to eliminate data redundancy and generate the final transmittable background bitstream.

[0011] As a further aspect of the present invention, the micro-dynamic background processing flow includes: independently extracting 64-dimensional low-dimensional residual semantic vectors for micro-dynamic textures in natural scenes, and transmitting them using 6-bit depth-specific encoding; after decoding at the receiving end, a dynamic perturbation map is generated, and the perturbation intensity coefficient λ is superimposed on the main background map to generate a dynamic background that fits the real scene, wherein the superposition formula is: ; in, To reconstruct the main background image, This is a micro-dynamic texture perturbation map. The disturbance intensity coefficient is in the range of 0.02 to 0.08. As a further aspect of the present invention, the refined background reconstruction process at the receiving end includes: Perform CABAC decoding and inverse quantization on the received background bitstream to restore a one-dimensional semantic feature vector; The one-dimensional feature vector is mapped to the original dimensional feature map through a fully connected layer, and the resolution is restored step by step through four levels of deconvolution upsampling: H / 16→H / 8 (384 channels→256 channels), H / 8→H / 4 (256 channels→128 channels), H / 4→H / 2 (128 channels→64 channels), H / 2→H (64 channels→3 channels), and the initial RGB background image is output. A first-order temporal moving average algorithm is used to eliminate inter-frame random jitter, with the smoothing coefficient α fixed at 0.8, as shown in the following formula: ; The pixel values ​​are cropped to the valid range of [0, 255] to output a stable and flicker-free final reconstructed background image.

[0012] As a further aspect of the present invention, the image fusion and color calibration process includes: The decoded foreground image and the reconstructed background image are initially stitched together. The hybrid Poisson fusion algorithm is used to solve the Poisson equation in the gradient domain to achieve a smooth gradient transition at the stitching boundary and eliminate stitching marks. Convert the RGB image to the LAB color space, unify the mean and variance of the L brightness channel, and correct the global brightness consistency. Fine-tune the average values ​​of the A and B chroma channels to eliminate temperature differences between foreground and background scenes and unify the color temperature, exposure, and contrast of the entire image. Convert the calibrated LAB image back to the RGB color space to complete the image fusion enhancement.

[0013] As a further aspect of the present invention, the AI ​​super-resolution enhancement process is as follows: a deep residual super-resolution network is used to perform upsampling of 2x in both horizontal and vertical directions on the fused complete image, and the image details and textures are restored through multi-scale feature reconstruction and sub-pixel convolution, thereby increasing the total pixel area of ​​the image by 4x and repairing the loss of details caused by compression transmission.

[0014] As a further aspect of the present invention, the video temporal continuity restoration process includes: Input two adjacent keyframes into a deep learning optical flow estimation model to predict the bidirectional motion vector field; Based on the motion vector field, displacement interpolation is performed on the key frame pixels to intelligently generate all transition prediction frames between two frames. The original keyframes and the interpolated predicted frames are reassembled in time sequence to fill in the frame sequence dropped by the sending end and restore the normal frame rate and smoothness of the video.

[0015] As a further aspect of the present invention, a closed-loop mechanism for model synchronization and background deviation correction is also included: The sending and receiving ends pre-store the same initial AI model. When switching scenes, the sending end compares multiple frames, corrects errors, and iteratively updates the model weights by combining the loss function and stochastic gradient descent optimizer to generate a new model that adapts to the scene. The transmitting end incrementally synchronizes the model parameters through a narrowband network to ensure model consistency between the transmitting and receiving ends; The receiver periodically compares and reconstructs the Euclidean distance between the background features and the transmitted semantic features. When the distance exceeds the threshold, the background deviation is immediately corrected using the latest keyframe features to prevent the cumulative distortion of the image over a long period of time. By employing a hybrid innovative architecture that combines AI semantic-level extreme background compression, traditional bitstream foreground fidelity compression, dynamic adaptive keyframe filtering, and end-to-end AI-enhanced reconstruction, the limitations of traditional video encoding pixel-level transmission are broken, enabling stable transmission of high-fidelity video under narrowband and extremely low bandwidth conditions.

[0016] This invention aims to overcome the shortcomings of existing traditional video compression algorithms, such as limited compression ratio, low bandwidth utilization, poor adaptability, poor image quality in narrowband transmission, and video choppyness. It provides a high-compression video method that incorporates artificial intelligence and can be transmitted in narrowband networks. Through AI semantics and traditional bitstream hybrid compression, dynamic keyframe adaptive selection, and end-to-end AI optimization, it achieves high-definition, smooth, and low-distortion video transmission in ultra-narrow bandwidth.

[0017] Compared with related technologies, the high-compression video method integrating artificial intelligence that can be transmitted over narrowband networks provided by this invention has the following beneficial effects: This invention breaks through the bandwidth bottleneck of traditional encoding, achieving extremely high compression ratios and perfectly adapting to narrowband scenarios. This invention abandons the traditional full-pixel transmission mode and innovatively adopts AI semantic-level compression technology to compress massive pixel background images into 256-dimensional / 64-dimensional low-dimensional semantic feature vectors. Combined with differentiated quantization encoding and CABAC lossless compression, the background data compression ratio is increased by more than a thousand times. Furthermore, a dynamic keyframe filtering mechanism significantly reduces the amount of redundant frames transmitted. The overall video transmission bandwidth can be compressed to 1~50kbps, reducing bandwidth usage by more than 90% compared to traditional H.265 encoding. This completely solves the industry pain point of being unable to transmit video in ultra-low bandwidth scenarios such as narrowband IoT, long-distance wireless transmission, and satellite inspection, with compression performance far exceeding existing traditional technologies.

[0018] This invention pioneers a three-tiered differentiated compression strategy for foreground, static background, and micro-dynamic background, overcoming the drawbacks of traditional globally uniform compression. High-rate bitrate compression is applied to foreground targets of core user interest, preserving target details and realism to the greatest extent possible. Extreme semantic compression is used for invalid static backgrounds to minimize bandwidth consumption. Micro-dynamic backgrounds such as swaying leaves and water ripples are separately encoded with low-dimensional residual features to avoid stiff and distorted background images. Precise bandwidth resource allocation ensures both clear image quality of core monitoring targets and a natural and realistic overall picture with extremely low bandwidth overhead, achieving optimal bandwidth utilization and video quality.

[0019] This invention abandons the traditional fixed-interval frame capture mode. Based on three-frame differential and OTSU adaptive thresholding algorithms, it calculates the foreground motion change rate in real time and dynamically adjusts the keyframe density according to the intensity of scene motion: keyframes are encrypted for scenes with intense motion to ensure image continuity; keyframes are sparse for static scenes to save bandwidth to the extreme. At the same time, a timed frame insertion fallback mechanism is set up to completely avoid background deviation at the receiving end caused by long periods without keyframes.

[0020] This invention constructs a comprehensive AI-enhanced restoration system for the receiving end. It achieves high-definition background reconstruction through multi-level deconvolutional decoding, completely eliminates foreground and background stitching artifacts through hybrid Poisson fusion, unifies global brightness and color temperature through LAB color space calibration, repairs detail loss through AI super-resolution, and fills in redundant discarded frames through optical flow estimation frame interpolation. Through multi-dimensional intelligent optimization, it completely compensates for the image quality loss and timing breaks caused by narrowband compression transmission, ultimately outputting a video with uniform color, smooth edges, rich details, and a stable frame rate, providing a viewing experience close to the original video. Attached Figure Description

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic diagram of the first overall principle architecture of the present invention; Figure 2 This is a schematic diagram of the second overall principle architecture of the present invention; Figure 3 This is a schematic diagram of a scene simulation provided by the present invention. Detailed Implementation

[0023] Please refer to the following: Figure 1 and Figure 3 ,in, Figure 1 This is a schematic diagram of the first overall principle architecture of the present invention; Figure 2 This is a schematic diagram of the second overall principle architecture of the present invention; Figure 3 This is a scene simulation diagram provided by the present invention. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks is implemented as follows:

[0024] Implementation Scenario 1 Parameter Settings This embodiment selects a field satellite inspection scenario. The original video resolution is 1920×1080, the frame rate is 25fps, and the transmission network is a narrowband satellite network with an effective bandwidth of 8~15kbps. The scene parameter configuration is as follows: M1=0.01, M2=0.1, micro-dynamic residual vector dimension is 64-dimensional with a bit depth of 6, ordinary background vector dimension is 256-dimensional with a bit depth of 8, temporal smoothing coefficient α=0.8, and disturbance intensity coefficient... =0.05.

[0025] Specific implementation steps of the sending end Step 1: Collect raw video frames from field inspections, remove image noise caused by strong light and sandstorms through an AI noise reduction network, convert the RGB color space to the YUV color space, and compress the chromaticity data volume.

[0026] Step 2: Activate the dynamic keyframe determination algorithm, continuously acquire three frames of images, and calculate using the bidirectional difference formula. , , The optimal segmentation threshold is obtained by using the OTSU algorithm to solve for the difference image. Calculate the rate of change of foreground motion The scene is mostly a static outdoor background with few targets. If the value is stable and less than 0.01, the system determines that one key frame is selected every 60 frames, and all other predicted frames are discarded.

[0027] Step 3: Input the keyframes into the AI ​​large model to segment the foreground areas such as inspection equipment and moving targets, the static background areas such as mountains and water, and the micro-dynamic background areas such as swaying leaves and water ripples.

[0028] Step 4: Extract background semantic features using a lightweight CNN+Transformer hybrid encoder. For static backgrounds, generate 256-dimensional feature vectors, and for micro-dynamic backgrounds, generate 64-dimensional residual feature vectors. After normalization, differential quantization, and lossless CABAC encoding, generate an ultra-low bitrate background data stream.

[0029] Step 5: Compress the foreground target using the H.265 high-ratio compression algorithm to ensure clear target outline and details, and generate a foreground data stream.

[0030] Step 6: The two streams are combined and encapsulated, and then transmitted to the receiving end via a 10kbps narrowband satellite network.

[0031] Specific implementation steps of the receiving end Step 1: The receiving end parses the combined bitstream, performs split decoding and inverse quantization to restore the 256-dimensional main background feature vector and the 64-dimensional micro-dynamic residual feature vector.

[0032] Step 2: Reconstruct a 1920×1080 high-definition main background image using a four-level deconvolution upsampling decoder, and generate a dynamic background with natural ripples and swaying leaves by combining residual features. Smooth the inter-frame jitter through temporal sliding average, and output a stable background image.

[0033] Step 3: Decode the foreground compressed bitstream to restore the high-definition foreground image of the inspection target.

[0034] Step 4: Combine foreground and background images, eliminate stitching boundaries using a hybrid Poisson fusion algorithm, convert the LAB color space to unify brightness, color temperature, and contrast, and correct color deviations.

[0035] Step 5: Upsample the image by 2x using an AI super-resolution network to restore details and textures and improve image clarity.

[0036] Step 6: Analyze the motion trajectory of adjacent keyframes using the optical flow estimation model, interpolate to generate all intermediate prediction frames, restore the video frame rate to 25fps, and restore the temporal continuity.

[0037] Step 7: Real-time verification of the error between the reconstructed background and the original features; automatic correction when the deviation exceeds the standard; synchronous monitoring of scene changes; and completion of model iteration and update.

[0038] Implementation effect verification

[0039] This embodiment successfully achieved stable 1080P video transmission with an ultra-low bandwidth of 10kbps, without stuttering, frame skipping, or significant distortion; the PSNR of the foreground inspection target was ≥32dB, and details were clearly discernible; the background image was natural and dynamic, without stiff flickering, and the colors were uniform and consistent; compared with the traditional H.265 compression method, the bandwidth usage was reduced by 92%, and the image quality was significantly improved, fully meeting the engineering application requirements of narrowband field inspection.

[0040] Example 2: Example of Narrowband Surveillance Video Compression and Transmission in Urban Traffic Circumstances Implementation scenario parameter settings This embodiment selects an urban traffic monitoring scene. The original video resolution is 1920×1080, the frame rate is 25fps, and the transmission network is a 5G narrowband IoT link with an effective bandwidth of 15~30kbps. The scene adaptation parameters are configured as follows: motion thresholds M1=0.03, M2=0.18, micro-dynamic residual vector dimension 64-dimensional, bit depth 6, ordinary background vector dimension 256-dimensional, bit depth 8, temporal smoothing coefficient α=0.8, and disturbance intensity coefficient λ=0.06. This scene contains high-frequency dynamic foreground elements such as vehicle movement and pedestrian movement, as well as micro-dynamic background elements such as swaying leaves and changes in road surface lighting. The motion complexity is moderate, which can fully verify the dynamic adaptive coding and differentiated compression performance of this invention.

[0041] Specific implementation steps of the sending end Step 1: The road monitoring camera continuously collects real-time video frames. The AI ​​noise reduction network removes image noise caused by changes in outdoor lighting, dust, rain and snow, completes the conversion from RGB color space to YUV color space, downsamples the chromaticity components, and initially reduces the video data volume to meet the requirements of narrowband transmission.

[0042] Step 2: Activate the dynamic keyframe intelligent filtering algorithm to capture three consecutive time-series video images in real time, and calculate the result using the bidirectional inter-frame difference formula. , Differential images are fused to generate the final differential image. The optimal segmentation threshold is adaptively solved using the Otsu's maximum inter-class variance algorithm. It accurately segments moving foreground elements such as vehicles and pedestrians from static background elements such as roads, green belts, and buildings. It also calculates the rate of change of motion in the foreground area in real time. In this scenario, due to the continuous movement of vehicles and pedestrians, It remains stable within the M1 and M2 interval.

[0043] Step 3: Based on scene threshold rules, the system automatically selects one keyframe every 30 frames, discarding all other intermediate predicted frames. This significantly reduces temporal redundancy while ensuring the continuity of dynamic footage. The selected keyframes are then input into a lightweight CNN+Transformer hybrid AI model to achieve accurate segmentation of multiple regions, independently dividing them into three categories: traffic target foreground, static road background, and micro-dynamic green plant background.

[0044] Step 4: Perform differentiated semantic compression for different regions: Extract 256-dimensional global semantic feature vectors from static backgrounds such as roads and buildings, and encode them after 8-bit deep quantization and normalization; extract 64-dimensional residual feature vectors from areas with swaying greenery and slight changes in light and shadow, and encode them using 6-bit lightweight quantization; synthesize the two types of background data into an ultra-low bit rate background data stream using CABAC lossless arithmetic encoding.

[0045] Step 5: For foreground targets that users are most concerned about, such as vehicles, pedestrians, and traffic signs, a high-ratio H.265 bitstream compression algorithm is used to retain core information such as target outlines, license plate details, and motion trajectories, generating a high-fidelity foreground data stream.

[0046] Step 6: Combine and encapsulate the background semantic bitstream and the foreground compressed bitstream, add timestamps and frame type verification identifiers, and complete real-time transmission through a 20kbps narrowband IoT link.

[0047] Specific implementation steps of the receiving end Step 1: The receiving end parses the combined bitstream in real time, and restores the 256-dimensional static background feature vector and the 64-dimensional micro-dynamic residual feature vector through CABAC decoding and inverse quantization, respectively, while decoding to restore the complete foreground target image.

[0048] Step 2: The background image resolution is restored step by step through a four-level deconvolution upsampling decoder to output a 1080P high-definition basic background image. Combined with micro-dynamic residual features, a dynamic background with natural lighting and shadows and green plant swaying effects is generated. The first-order temporal moving average algorithm is used to smooth the inter-frame jitter and eliminate the screen flickering problem that is prone to occur in outdoor scenes.

[0049] Step 3: Use the hybrid Poisson fusion algorithm to perform gradient domain fusion processing on the foreground and background images to completely eliminate the stitching gaps and color banding problems at the target edges; convert to LAB color space to unify the brightness, color temperature, and contrast of the entire image and correct the color difference caused by uneven outdoor lighting.

[0050] Step 4: Perform 2x bidirectional super-resolution reconstruction on the fused image using an AI super-resolution residual network to repair details such as license plate textures, vehicle outlines, and road markings lost during compression, thereby improving image clarity.

[0051] Step 5: Use the optical flow estimation model to analyze the vehicle and pedestrian motion vector fields of adjacent keyframes, intelligently interpolate to generate all discarded intermediate prediction frames, stabilize the video frame rate to 25fps, make up for the temporal continuity, and eliminate vehicle driving stuttering and screen jump problems.

[0052] Step 6: Activate the closed-loop correction mechanism. The receiver periodically compares the Euclidean distance between the reconstructed background and the original semantic features to correct minor background deviations in real time. The transmitter adaptively fine-tunes the model parameters according to changes in traffic density to ensure long-term stable operation in complex traffic scenarios.

[0053] Implementation effect verification This embodiment achieves stable real-time transmission of 1080P traffic video with a narrow bandwidth of 20kbps. Dynamic vehicles and pedestrians are clearly identifiable without ghosting or blurring; the background is natural and realistic, and the subtle dynamic effects of greenery and lighting are smooth without stiffness or distortion; the video maintains a stable frame rate throughout, without stuttering or frame skipping. Compared to traditional H.265 encoding, bandwidth usage is reduced by 88%, and target recognition availability is significantly improved, fully meeting the real-time transmission and intelligent analysis requirements of urban narrowband IoT traffic monitoring.

[0054] Example 3: Low-power video compression and transmission example for indoor dense scene surveillance Implementation scenario parameter settings This embodiment selects a dense indoor office monitoring scenario. The original video resolution is 1920×1080, the frame rate is 25fps, and the transmission network is a low-power narrowband WiFi link with an effective bandwidth of 5~12kbps. The scenario adaptation parameters are configured as follows: motion threshold M1=0.04, M2=0.16, micro-dynamic residual vector dimension 64-dimensional, bit depth 6, ordinary background vector dimension 256-dimensional, bit depth 8, temporal smoothing coefficient α=0.8, and disturbance intensity coefficient λ=0.04. This scenario contains dense small-amplitude dynamics such as people walking and objects moving, as well as weak micro-dynamics such as curtains swaying and air conditioning airflow disturbances, adapting to the low-bandwidth, high-stability indoor monitoring requirements.

[0055] Specific implementation steps of the sending end Step 1: The indoor camera captures high-definition video frames, which are then processed by an AI lightweight noise reduction network to remove indoor lighting noise and image artifacts. The RGB-YUV color space is converted, and redundant chromaticity data is compressed to meet the requirements of low-power narrowband transmission pre-compression.

[0056] Step 2: Execute the dynamic frame filtering algorithm to acquire three consecutive frames of indoor video images. Use a two-way inter-frame difference algorithm to accurately capture areas of slight movement of people and objects, and generate a fused difference image. The optimal segmentation threshold T* is solved by using the OTSU adaptive threshold algorithm without parameters, which can accurately distinguish between the moving foreground of people and the static background of walls, furniture and ground.

[0057] Step 3: Real-time calculation of the rate of change of indoor foreground motion Small, frequent movements of people indoors In the low to medium range, the system selects a key frame every 30 frames according to the rules, filtering out a large number of repetitive static frames and slightly redundant dynamic frames, so as to save low-power bandwidth resources to the maximum extent.

[0058] Step 4: The AI ​​hybrid encoder performs region segmentation on keyframes, distinguishing between foreground elements with people in motion, static interior furnishings, and dynamic background elements such as curtains and gentle breezes. Static walls and furniture backgrounds are extracted with 256-dimensional semantic feature vectors for extreme compression; curtains and lightweight ornaments with subtle dynamic regions are extracted with 64-dimensional residual features to preserve subtle dynamic effects; all background features are quantized, normalized, and losslessly encoded using CABAC to generate an ultra-small background bitstream.

[0059] Step 5: For the core foreground areas such as people and moving objects, use high-fidelity, high-rate bitrate compression to preserve human outlines, movement details, and object features, ensuring the core image quality requirements of indoor security monitoring.

[0060] Step 6: After the two streams are combined and encapsulated, continuous and stable transmission is achieved through an 8kbps low-power narrowband WiFi link.

[0061] Specific implementation steps of the receiving end Step 1: The receiving end decodes the bitstream in multiple paths to restore the background semantic feature vector and the foreground high-definition image data, and recovers the complete feature information through inverse quantization.

[0062] Step 2: Multi-level deconvolution upsampling reconstructs a 1080P indoor high-definition background, and superimposes micro-dynamic residual features to achieve a realistic effect of natural swaying of curtains and subtle changes in air texture. The temporal smoothing algorithm eliminates the image jitter caused by indoor light flicker.

[0063] Step 3: Use the hybrid Poisson fusion algorithm to achieve seamless stitching of foreground and background, eliminating stitching edge defects caused by complex interior furnishings; use LAB color calibration to unify the color temperature and brightness of indoor lights, solving the problems of excessive brightness and color cast in some areas of the image.

[0064] Step 4: AI super-resolution network repairs indoor detail textures, optimizes the clarity of details on desktops, walls, and furnishings, and compensates for the loss of detail caused by low bitrate compression.

[0065] Step 5: The optical flow interpolation algorithm accurately generates intermediate frames of human movement, fills in the dropped prediction frames, and makes the human movement smooth and natural without any stuttering or lag, restoring the standard 25fps video frame rate.

[0066] Step 6: The system continuously performs model synchronization and deviation correction, adapts to scene changes such as indoor personnel entering and exiting, and changes in the placement of items, automatically iterates and optimizes AI model parameters, and prevents long-term monitoring images from accumulating distortion.

[0067] Implementation effect verification This embodiment achieves long-term stable transmission of 1080P indoor video with an ultra-low power consumption and narrow bandwidth of 8kbps. Personnel movements are smooth, details are clear, and there is no blurring or ghosting. Indoor background images are realistic and natural, with soft micro-dynamic effects and no stiffness or rigidity. There is no flickering, stuttering, or color distortion throughout the entire process. Compared to traditional video encoding technologies, bandwidth resource consumption is reduced by 93%, and power consumption is significantly reduced. It is perfectly suited for indoor security monitoring scenarios requiring low power consumption, ultra-narrow bandwidth, and high stability, demonstrating strong scenario adaptability and engineering practicality.

[0068] The working principle of the high-compression video compression method incorporating artificial intelligence that can be transmitted over narrowband networks provided by this invention is as follows: After the camera captures raw RGB video frames, a deep learning AI noise reduction network is first used to remove salt-and-pepper noise and Gaussian noise caused by low light and environmental interference, thus purifying the raw image data. Then, a color space conversion is performed, replacing the RGB color space with the YUV color space. Taking advantage of the human eye's lower sensitivity to color than to brightness, the U and V chromaticity components are downsampled, initially reducing the data volume by about 50%, laying the foundation for subsequent ultra-high compression.

[0069] To eliminate temporal redundancy in video, this invention employs a motion-driven dynamic keyframe determination algorithm that is fully adaptive and has no fixed threshold. The first step is to capture three consecutive frames of time-series images in real time, eliminate noise errors from unidirectional difference using a bidirectional inter-frame difference algorithm, calculate the fused difference image, and accurately capture the motion areas of the image. The second step is to use the OTSU maximum inter-class variance adaptive thresholding algorithm to automatically traverse all grayscale thresholds from 0 to L-1 and solve for the optimal segmentation threshold that maximizes the inter-class variance of the foreground and background, thus accurately segmenting the moving foreground and static background without the need for manual parameter tuning. The third step is to statistically analyze the proportion of moving pixels in the foreground, calculate the real-time foreground change rate, and combine the M1 and M2 threshold ranges for five full-scene adaptations to dynamically determine the frame type and accurately distinguish between effective keyframes and redundant predicted frames. The fourth step is to discard all predicted frames and retain only key frames for subsequent compressed transmission, thereby significantly reducing the amount of data transmitted from a temporal perspective.

[0070] The selected high-quality keyframes are input into a pre-trained AI model, which possesses scene-adaptive learning capabilities: when matching existing scenes in the model library, it directly calls pre-trained weights to quickly segment the foreground, background, and micro-dynamic regions; when no matching scene exists, it automatically generates a custom scene model through multi-frame comparison, deduplication and error correction, and iterative optimization of the loss function, without the need for manual annotation. Ultimately, this achieves accurate partitioning of the foreground, static background, and micro-dynamic textured background that users focus on, providing a basis for differentiated compression.

[0071] Differentiated compression strategies are employed for different segments: The static background region is mapped to a 256-dimensional low-dimensional semantic feature vector by a lightweight CNN+Transformer hybrid encoder, which performs four convolutional shallow feature extraction layers, mid-layer spatiotemporal feature fusion, and high-level semantic projection. The micro-dynamic background is extracted separately as a 64-dimensional residual feature vector to accurately represent subtle texture changes. Subsequently, the data scale is unified by normalization, and the number of bits is reduced by differential bit depth quantization. Finally, CABAC context-adaptive arithmetic coding is used to eliminate data redundancy, generating an ultra-small volume background bitstream.

[0072] The foreground area uses a mature high-ratio video coding algorithm to maximize the compression bit rate within a controllable distortion range, ensuring the integrity of the target outline, details, and motion trajectory, and achieving core image quality fidelity.

[0073] By combining and encapsulating the background semantic bitstream and the foreground compressed bitstream, and adding identification information such as timestamps, frame types, and checksums, it is adapted to 1~50kbps narrowband wireless network transmission, solving the problem that traditional video cannot be transmitted under ultra-low bandwidth.

[0074] The receiving end parses the combined bitstream and decodes the separate streams to restore the semantic feature vectors and foreground images. After inverse quantization and decoding, the background features are upsampled through four levels of deconvolution to restore the image resolution step by step, and natural dynamic backgrounds are generated by combining micro-dynamic residual features. Inter-frame flickering and jitter are eliminated by first-order temporal moving average to ensure the stability of the background image.

[0075] The reconstructed background and decoded foreground are initially stitched together. A hybrid Poisson fusion algorithm is used to smoothly merge the stitching boundaries in the gradient domain, completely eliminating stitching gaps and edge discontinuities. Further conversion to the LAB color space unifies global brightness, color temperature, and contrast, resolving issues of color fragmentation between foreground and background and uneven brightness, achieving visual consistency across the entire image.

[0076] An AI super-resolution network is used to perform a 2x bidirectional upsampling of the entire image, restoring texture details lost during compressed transmission and improving image clarity. Finally, optical flow estimation technology is used to predict inter-frame motion trajectories, and intelligent interpolation is used to generate all predicted frames discarded by the sending end, completing the video temporal sequence, restoring the standard frame rate, and completely resolving video stuttering and frame skipping issues.

[0077] The system has self-iterative optimization capabilities. The sending end monitors scene changes in real time, automatically updates model parameters and incrementally synchronizes them to the receiving end to ensure model consistency between the sending and receiving ends. The receiving end periodically checks the background reconstruction error. If the error exceeds the threshold, it automatically corrects the deviation using the latest key frame features, forming a complete closed loop of "acquisition-compression-transmission-reconstruction-verification-correction" to ensure that the system can run stably and without distortion for a long time.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments, or they can be used directly or indirectly, without departing from the principles and spirit of the invention. In other related technical fields, the scope of the invention is defined by the appended claims and their equivalents, and they are similarly included within the scope of patent protection of the invention.

Claims

1. A high-compression video method incorporating artificial intelligence for transmission over narrowband networks, characterized in that, include: The process consists of several parts: AI-powered intelligent compression processing at the sending end, low-bit-rate transmission process in narrowband networks, and AI-powered intelligent reconstruction and enhancement process at the receiving end. The AI ​​intelligent compression processing flow at the sending end includes: performing AI noise reduction preprocessing and RGB-YUV color space conversion on the acquired original video frames to remove image noise and reduce the color data volume; A dynamic foreground change rate determination model is constructed based on the three-frame difference algorithm and the OTSU adaptive threshold segmentation algorithm. The key frames and prediction frames are dynamically and adaptively selected according to the intensity of the foreground motion in the video. The prediction frames are directly discarded and only the key frames are retained for transmission. A lightweight CNN+Transformer hybrid coding AI model is used to accurately segment multiple regions of keyframes, including foreground, background, and micro-dynamic background residuals. Low-dimensional global semantic feature vectors are extracted from static background regions and normalized, differentiated quantized, and CABAC lossless arithmetic coding is performed to generate ultra-low bitrate background data streams. For the foreground regions of interest to the user, high-rate video bitstream compression is used to generate foreground data streams. The two data streams are combined and transmitted through a kbps-level narrowband wireless network.

2. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1, characterized in that, in, The narrowband network low bit rate transmission process is adapted to a bandwidth range of 1kbps~50kbps. It replaces the traditional full pixel transmission with semantic feature compression to avoid the technical defects of insufficient bandwidth in narrowband networks. The specific implementation steps of the dynamic adaptive keyframe filtering are as follows: Select three consecutive time-series images , , The formula for calculating the bidirectional inter-frame absolute difference image is as follows: ; ; ; in, This is the difference image between the current frame and the previous frame. This is a difference image between the current frame and the next frame. This is the final difference image after fusion; In difference image The Otsu's maximum inter-class variance algorithm is used to iterate through the grayscale thresholds and solve for the optimal segmentation threshold. The formula is as follows: ; in, The maximum number of gray levels in the image. Let T be the inter-class variance of foreground and background pixels corresponding to the threshold. The optimal segmentation threshold; The foreground motion region is obtained based on optimal threshold binarization segmentation, and the rate of change of the foreground region is calculated. The formula is as follows: ; in, for Frame binarization of foreground mask image, for Frame binarization of foreground mask image, This represents the total number of pixels in the image. Preset motion threshold values ​​M1 (lower limit of rate of change) and M2 (upper limit of rate of change) within the range of 0 to 1, which are adaptively assigned according to the scene: In the static monitoring scenario, M1=0.015 and M2=0.12; In the traffic and road scenario, M1=0.03 and M2=0.

18. In the handheld inspection scenario, M1=0.05 and M2=0.

2. In dense indoor scenes, M1=0.04 and M2=0.16; In the field satellite inspection scenario, M1=0.01 and M2=0.1; Dynamically determine frame type: When M2 is greater than M2, the current frame is directly determined as a keyframe; M1 < When M2 is less than or equal to 2, a keyframe is selected every 30 frames. When M1 is less than or equal to 1, a key frame is selected every 60 frames, and all other frames are judged as prediction frames and discarded directly.

3. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1, characterized in that, in, The AI ​​intelligent reconstruction and enhancement process at the receiving end includes: de-decoding the combined bitstream, inverse quantizing to restore the background semantic feature vector, reconstructing the basic background image through a multi-level deconvolution upsampling decoder, and generating a natural dynamic background by combining micro-dynamic residual features; decoding to restore the high-definition foreground image, and using a hybrid Poisson fusion algorithm and LAB color space calibration to complete seamless stitching of the foreground and background; improving the image detail clarity through AI super-resolution upsampling, interpolating and generating discarded prediction frames based on optical flow estimation technology to complete the temporal continuity of the video, and finally outputting a high-definition, smooth, and color-consistent complete video; The lightweight CNN+Transformer hybrid encoder's network structure and semantic feature extraction process include: a shallow CNN texture extraction module: setting up a four-layer progressive convolutional structure. The first layer converts 3 channels to 64 channels, uses 3×3 convolution, and has a stride of 2, reducing the resolution to half. The second layer converts 64 channels to 128 channels, uses 3×3 convolution, and has a stride of 2, reducing the resolution to 1 / 4. The third layer converts 128 channels to 256 channels, uses 3×3 convolution, and has a stride of 2, reducing the resolution to 1 / 8. The fourth layer converts 256 channels to 384 channels, performs a 3×3 convolution with a stride of 2, reduces the resolution to 1 / 16, and outputs the bottom-level spatial feature map F∈R. Mid-layer spatiotemporal fusion module: fuses the low-level features of the current frame with the historical features of the previous frame, models the slow changes in the scene and the micro-motions of objects, and outputs the mid-layer fusion feature F; High-level semantic projection module: Compresses the spatiotemporal dimension through global mean pooling, and generates the final global semantic feature vector through two fully connected layers and the GELU activation function. The formula is as follows: ; in, , The weight matrix is ​​the fully connected component. , For bias vectors, This is a global mean pooling operation. This is the GELU activation function.

4. A high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1 or 3, characterized in that, The feature vector differential quantization encoding process includes: Perform global normalization on the extracted global semantic feature vector to scale the values ​​of each dimension to the [0,1] interval; Differentiated bit depth quantization: Ordinary static background feature vectors use 256-dimensional vectors and 8-bit depth quantization; micro-dynamic background residual feature vectors such as leaf swaying and water ripples use 64-dimensional vectors and 6-bit depth quantization. The quantized integer sequence is subjected to lossless compression using CABAC context-adaptive binary arithmetic coding to eliminate data redundancy and generate the final transmittable background bitstream.

5. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1, characterized in that, The micro-dynamic background processing flow includes: independently extracting 64-dimensional low-dimensional residual semantic vectors for subtle dynamic textures in natural scenes, and transmitting them using 6-bit depth-coded individual encoding; after decoding at the receiving end, a dynamic perturbation map is generated, which is then superimposed on the main background map using a perturbation intensity coefficient λ to generate a dynamic background that fits the real scene. The superposition formula is as follows: ; in, To reconstruct the main background image, This is a micro-dynamic texture perturbation map. The disturbance intensity coefficient is in the range of 0.02 to 0.

08.

6. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1, characterized in that, The refined background reconstruction process at the receiving end includes: Perform CABAC decoding and inverse quantization on the received background bitstream to restore a one-dimensional semantic feature vector; The one-dimensional feature vector is mapped to the original dimensional feature map through a fully connected layer, and the resolution is restored step by step through four levels of deconvolution upsampling: H / 16→H / 8 (384 channels→256 channels), H / 8→H / 4 (256 channels→128 channels), H / 4→H / 2 (128 channels→64 channels), H / 2→H (64 channels→3 channels), and the initial RGB background image is output. A first-order temporal moving average algorithm is used to eliminate inter-frame random jitter, with the smoothing coefficient α fixed at 0.8, as shown in the following formula: ; The pixel values ​​are cropped to the valid range of [0, 255] to output a stable and flicker-free final reconstructed background image.

7. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1, characterized in that, The image fusion and color calibration process includes: The decoded foreground image and the reconstructed background image are initially stitched together. The hybrid Poisson fusion algorithm is used to solve the Poisson equation in the gradient domain to achieve a smooth gradient transition at the stitching boundary and eliminate stitching marks. Convert the RGB image to the LAB color space, unify the mean and variance of the L brightness channel, and correct the global brightness consistency. Fine-tune the average values ​​of the A and B chroma channels to eliminate temperature differences between foreground and background scenes and unify the color temperature, exposure, and contrast of the entire image. Convert the calibrated LAB image back to the RGB color space to complete the image fusion enhancement.

8. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1, characterized in that, The AI ​​super-resolution enhancement process is as follows: a deep residual super-resolution network is used to perform upsampling of 2x in both horizontal and vertical directions on the fused complete image. The image details and textures are restored through multi-scale feature reconstruction and sub-pixel convolution, increasing the total pixel area of ​​the image by 4x and repairing the loss of details caused by compression and transmission.

9. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1, characterized in that, The video temporal continuity restoration process includes: Input two adjacent keyframes into a deep learning optical flow estimation model to predict the bidirectional motion vector field; Based on the motion vector field, displacement interpolation is performed on the key frame pixels to intelligently generate all transition prediction frames between two frames. The original keyframes and the interpolated predicted frames are reassembled in time sequence to fill in the frame sequence dropped by the sending end and restore the normal frame rate and smoothness of the video.

10. The high-compression video method incorporating artificial intelligence for transmission over narrowband networks according to claim 1, characterized in that, It also includes a closed-loop mechanism for model synchronization and background bias correction: The sending and receiving ends pre-store the same initial AI model. When switching scenes, the sending end compares multiple frames, corrects errors, and iteratively updates the model weights by combining the loss function and stochastic gradient descent optimizer to generate a new model that adapts to the scene. The transmitting end incrementally synchronizes the model parameters through a narrowband network to ensure model consistency between the transmitting and receiving ends; The receiver periodically compares the Euclidean distance between the reconstructed background features and the transmitted semantic features. When the distance exceeds the threshold, the background deviation is immediately corrected using the latest keyframe features to prevent the cumulative distortion of the image over a long period of time.