Self-adaptive image transmission planning method and system based on image content and downlink transmission bandwidth

Abstract

The invention discloses a self-adaptive image transmission planning method and system based on image content and downlink transmission bandwidth, and belongs to the technical field of image processing and task planning. The method comprises the following steps: obtaining the size and resolution of an original image; a region-of-interest image and a non-region-of-interest image are generated by extracting a region-of-interest in an original image, and the position of the region-of-interest in the original image is acquired; wherein the resolution of the region-of-interest image is the resolution of the original image, and the resolution of the non-region-of-interest image is generated based on a downlink transmission bandwidth; and transmitting the region-of-interest image, the non-region-of-interest image, the size of the original image and the position of the region-of-interest in the original image to a second device via a downlink, so that the second device recovers the original image. According to the invention, the transmission pressure of the ground station can be effectively reduced, and redundant data are significantly reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing and task planning, and in particular relates to an adaptive image transmission planning method and system based on image content and downlink transmission bandwidth. Background Art

[0002] The development of low-cost satellite manufacturing and remote sensing technology has significantly improved the resolution and range of Earth imaging, generating massive amounts of data. However, transmitting this data in real time is challenging due to limitations in on-orbit storage and downlink bandwidth. High-resolution satellites typically operate in low-Earth orbits with only a few minutes of transmission windows during each pass, limiting the ability to quickly transmit imagery without losing important information. Therefore, an efficient transmission planning scheme is needed to reduce the amount of data and lower transmission costs.

[0003] Modern remote sensing services are usually completed by multiple autonomous satellites in collaboration. Image transmission planning becomes complicated due to factors such as changing requirements, coordinated capture, image processing and dynamic transmission speed. These complexities often exceed the computing power of a single satellite. Spatial redundancy in high-resolution images enables autonomous satellites to use image recognition technology to prioritize the transmission of key image parts. The main challenge lies in the reasonable allocation of transmission resolution between the area of ​​interest and the background area.

[0004] Image recognition technology can help identify regions of interest in remote sensing images. Although machine learning and deep learning have made breakthroughs in image recognition, they rely on a large number of training samples and accurate annotations, and complex models have high computational requirements and are difficult to run in real time on satellites. Some studies use low-level image features to simulate visual attention mechanisms for saliency analysis to extract regions of interest; others use predefined rules and adaptive fusion of different algorithms to optimize region of interest extraction, but these methods can be very time-consuming when processing multi-scale images and are not suitable for use on satellites. Summary of the invention

[0005] The present invention proposes an adaptive image transmission planning method and system based on image content and downlink transmission bandwidth, which can effectively alleviate the transmission pressure of ground stations and significantly reduce redundant data.

[0006] To achieve the above objectives, the technical solution of the present invention includes the following contents.

[0007] An adaptive image transmission planning method based on image content and downlink transmission bandwidth is applied to a first device, the method comprising:

[0008] Get the size and resolution of the original image;

[0009] Generate an image of the region of interest and an image of the region of non-interest by extracting the region of interest in the original image, and obtain the position of the region of interest in the original image; wherein the resolution of the image of the region of interest is the resolution of the original image, and the resolution of the image of the region of non-interest is generated based on the downlink transmission bandwidth;

[0010] The image of the region of interest, the image of the non-region of interest, the size of the original image, and the position of the region of interest in the original image are transmitted to the second device via a downlink, so that the second device restores the original image.

[0011] Furthermore, the step of extracting the region of interest in the original image, generating the region of interest image and the region of non-interest image, and obtaining the position of the region of interest in the original image includes:

[0012] Preprocess the original image;

[0013] Performing multi-scale processing on the pre-processed original image using a multi-scale Gaussian smoothing pyramid to obtain a multi-scale image of the original image;

[0014] Based on the multi-scale image, generating a mixed saliency map of the original image;

[0015] Screening and removing redundancy of regions of interest in the mixed saliency map;

[0016] Separating the region of interest from the background of the original image to obtain an image of the region of interest and a position of the region of interest in the original image;

[0017] The image of the non-interest region is generated by performing a full black mask processing on the tile position of the interest region in the original image, and calculating the resolution of the image of the non-interest region based on the downlink transmission bandwidth.

[0018] Furthermore, the preprocessing of the original image includes:

[0019] Perform adaptive downsampling preprocessing according to the size of the original image;

[0020] Based on the HSV color space of the adaptive downsampling preprocessing result, the shadow area is marked, and the HSV value of the shadow area is replaced by the average HSV value of the adaptive downsampling preprocessing result.

[0021] Further, based on the multi-scale image, generating a mixed saliency map of the original image includes:

[0022] Based on a global contrast saliency calculation method, a multi-scale saliency map pyramid corresponding to the multi-scale image is calculated; wherein the global contrast saliency calculation method includes: an LC algorithm;

[0023] Using image entropy as a metric, evaluating the saliency of each scale saliency map in the multi-scale saliency map pyramid, and performing weighted processing on the saliency map of each scale based on the saliency to generate a saliency map of the original image;

[0024] After edges and fine structures in the original image are identified using the Canny edge detection algorithm, a mixed saliency map is generated based on the edge detection result and the saliency map of the original image.

[0025] Furthermore, the regions of interest are screened and de-redundant in the mixed saliency map, including:

[0026] Binarizing the mixed saliency map to obtain a region of interest in the mixed saliency map;

[0027] Based on the YOLOv5 model, a semantic segmentation result of the original image is obtained, and based on the image entropy and inverse moment of the semantic segmentation result, the region of interest in the mixed saliency map that does not conform to the eigenvalue distribution is removed to obtain a first screening result of the region of interest;

[0028] Using the vegetation index based on the red-green-blue model, the first screening result of the area of ​​interest is screened to obtain a second screening result of the area of ​​interest;

[0029] According to the size of the region of interest and the position between the regions of interest, the regions of interest in the redundancy of the second screening result of the regions of interest are deleted.

[0030] Further, the resolution of the non-interest region image is generated based on the downlink transmission bandwidth, including:

[0031] Define the transmission resolution level of the image of non-interest areas;

[0032] Determine the time step for predicting downlink transmission bandwidth based on the satellite's image acquisition frequency and orbital dynamics;

[0033] Obtaining the downlink transmission bandwidth at several historical moments, and using a regression fitting model to predict the downlink transmission bandwidth at the next moment; wherein the regression fitting model includes: a LightGBM model;

[0034] Determine whether the downlink transmission bandwidth at the next moment is a sudden increase in bandwidth;

[0035] In the case that the downlink transmission bandwidth at the next moment is not a sudden increase in bandwidth, based on the downlink transmission bandwidth at the next moment, the time step of predicting the downlink transmission bandwidth and the data amount of the original image, calculating the maximum available resolution, and determining the resolution of the non-interest region image in the transmission resolution level according to the maximum available resolution;

[0036] In the case that the downlink transmission bandwidth at the next moment is a sudden increase in bandwidth, the resolution of the non-interesting image transmitted last time is used as the resolution of the non-interesting image.

[0037] Furthermore, the obtaining of downlink transmission bandwidths at a plurality of historical moments and using a regression fitting model to predict the downlink transmission bandwidth at the next moment includes:

[0038] Calculate the rolling mean and rolling standard deviation of the downlink transmission bandwidth at several historical moments;

[0039] Perform Z-score normalization on the downlink transmission bandwidth at several historical moments and the corresponding rolling mean and rolling standard deviation;

[0040] The standardized data is input into a trained regression fitting model to obtain the downlink transmission bandwidth at the next moment; wherein the parameters in the trained regression fitting model are obtained by grid search optimization.

[0041] Further, the determining whether the downlink transmission bandwidth at the next moment is a sudden increase in bandwidth includes:

[0042] Generate a time series according to the downlink transmission bandwidth at a historical moment and the downlink transmission bandwidth at a next moment;

[0043] The time series data is divided into a plurality of windows of fixed size by a sliding window, and the standard deviation in each window is calculated;

[0044] If the standard deviation of the window where the downlink transmission bandwidth at the next moment is located and the standard deviation of other windows is greater than the set threshold, and the bandwidth increase and decrease trend is reversed, it is determined that the downlink transmission bandwidth at the next moment is a bandwidth sudden increase.

[0045] Further, the determining the resolution of the non-interest region image in the transmission resolution level according to the available maximum resolution includes:

[0046] Among the transmission resolution levels, obtaining a level that is not greater than the available maximum resolution;

[0047] Calculating an excitation factor for each level not greater than the maximum resolution available;

[0048] According to the maximum value of the excitation factor, a resolution level is determined in a level greater than the maximum available resolution to obtain the resolution of the image of the non-interest region.

[0049] An adaptive image transmission planning system based on image content and downlink transmission bandwidth, the system comprising:

[0050] The first device is used to obtain the size and resolution of the original image; generate an image of the region of interest and an image of the region of non-interest by extracting the region of interest in the original image, and obtain the position of the region of interest in the original image; wherein the resolution of the image of the region of interest is the resolution of the original image, and the resolution of the image of the region of non-interest is generated based on the downlink transmission bandwidth; and transmit the size of the image of the region of interest, the image of the region of non-interest, the original image, and the position of the region of interest in the original image to the second device via the downlink;

[0051] The second device is used to restore the original image based on the image of the region of interest, the image of the non-region of interest, the size of the original image, and the region of interest in the original image.

[0052] Compared with the prior art, the present invention uses multi-scale hybrid saliency mapping combined with semantic refinement to quickly and accurately extract regions of interest, and downsamples regions of non-interest to adapt to the dynamic changes in downlink bandwidth, achieving the best balance between data volume and real-time transmission speed, thereby effectively alleviating the transmission pressure of the ground station and significantly reducing redundant data. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The invention is a flowchart of an adaptive image transmission planning method based on image content and downlink transmission bandwidth.

[0054] Figure 2 The present invention is a schematic diagram of the process of extracting regions of interest in an adaptive image transmission planning method based on image content and downlink transmission bandwidth.

[0055] Figure 3 The invention relates to a downlink transmission bandwidth prediction result and a corresponding non-interest area transmission resolution allocation result in an adaptive image transmission planning method based on image content and downlink transmission bandwidth.

[0056] Figure 4 These are three examples of fused and stitched images that are transmitted to the ground end through an adaptive image transmission planning method based on image content and downlink transmission bandwidth. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below through its specific embodiments in conjunction with the accompanying drawings.

[0058] The present invention proposes an adaptive image transmission planning method based on image content and downlink transmission bandwidth, such as Figure 1 As shown, the steps include steps 1 to 3.

[0059] Step 1: Get the size and resolution of the original image.

[0060] In an embodiment of the present invention, the original image is a high-resolution earth image taken by a satellite.

[0061] Step 2: Generate an area of ​​interest image and a non-area of ​​interest image by extracting the area of ​​interest in the original image, and obtain the position of the area of ​​interest in the original image.

[0062] like Figure 2 As shown, the generation of the region of interest image and the non-region of interest image includes the following process.

[0063] (1) Image preprocessing.

[0064] (1-1) Adaptive downsampling.

[0065] Adaptive downsampling preprocessing is performed according to the size of the original image to speed up the subsequent saliency map calculation. For large images, the image size is reduced to increase the processing speed.

[0066] (1-2) Shadow area detection.

[0067] The shadow area in the output is marked (1-1) using an algorithm based on the HSV color space and replaced with the average HSV value to avoid the shadow being misidentified as the area of ​​interest. The shadow area has a larger hue value; the scattered light in the shadow area mainly comes from the blue-violet light with a shorter wavelength, so it has a high saturation value; the sunlight in the shadow area is blocked, resulting in a low brightness value.

[0068] (2) Multi-scale image pyramid.

[0069] (2-1) Multi-scale image processing.

[0070] A multi-scale Gaussian smoothing pyramid is used to perform multi-scale processing on the image output by (1-2). Each image obtains a multi-dimensional output of multiple sizes, that is, a multi-scale image pyramid is output for each image to adapt to the diversity of target scales, improve detection accuracy, and maintain processing speed.

[0071] (3) Mixed saliency map.

[0072] (3-1) Multi-scale saliency map pyramid.

[0073] The saliency map of the multi-scale image output by (2-1) is calculated using the global contrast saliency calculation method to obtain a multi-scale saliency map pyramid. For example, the LC algorithm is used, which performs saliency evaluation based on the grayscale contrast of pixels. The specific formula is as follows:

[0074]

[0075] Among them, S(p) is the significance value of pixel p, I is the remote sensing image, and d(p,q) is the distance between pixels p and q.

[0076] (3-2) Adaptive weighting.

[0077] In view of the intra-class differences caused by multi-scale and target diversity, the present invention introduces image entropy as a metric to evaluate the significance of the LC feature map generated at each scale. Then, a multi-scale adaptive weighting method is used to improve the detection accuracy and maintain a high processing speed. The calculation formula of image entropy is as follows:

[0078]

[0079] Among them, p i is the probability of a specific gray value appearing in the image, which can be directly obtained from the grayscale histogram.

[0080] (3-3) Fusion edge detection generates a hybrid saliency map.

[0081] The Canny edge detection algorithm is used to accurately identify edges and subtle structures in the image through multi-scale Gaussian smoothing, gradient detection, and non-maximum suppression. The edge detection result is combined with the LC saliency map output by (3-2) to generate a hybrid saliency map. This combination can accurately locate useful boundaries while preserving the contrast of important areas.

[0082] (4) Region of interest extraction.

[0083] Although the extraction of regions of interest based on low-level features is fast, the results may not be consistent with the high-level semantics expected by users (such as buildings, objects, scenes, etc.). In order to achieve fast complex semantic detection under limited computing resources, this paper proposes a screening method based on prior semantic constraints, which uses the segmentation results of the YOLOv5 model as a supervision constraint to further screen the regions of interest.

[0084] (4-1) Binarization processing.

[0085] The mixed saliency map is binarized, and the region of interest is obtained according to the binarization result.

[0086] (4-2) Semantic Constraint Filtering

[0087] The original image is sent to the YOLOv5 model for semantic segmentation. The image entropy and inverse moment difference in the segmentation result of the YOLOv5 model are analyzed to form a feature value threshold range. The threshold is used to further filter the region of interest obtained from the mixed saliency map with semantic constraints, and remove traffic signs, site boundaries, and uninteresting buildings that do not meet the semantic requirements. The image entropy and inverse moment difference metrics have low computational complexity and meet the real-time requirements of on-orbit autonomous processing.

[0088] Image entropy is used to measure image complexity, and its definition is similar to the multi-scale adaptive feature fusion part mentioned above. The inverse difference moment represents the uniformity of image texture and is used to measure the change of local texture. The calculation formula of IDM is as follows:

[0089]

[0090] Where N is the number of gray levels in the image, and P(i,j) is the value of position (i,j) in the gray level co-occurrence matrix, which represents the probability of gray values ​​i and j co-occurring. The gray level co-occurrence matrix is ​​generated by counting the spatial relationship between the gray values ​​of pixels in the image to quantify the texture information of the image. The method first sets the relative distance and direction of the gray value pairs of pixels in the image (such as 0 degrees, 45 degrees, 90 degrees or 135 degrees), counts the co-occurrence frequency of the gray value pairs that satisfy the spatial relationship, and generates a matrix. The matrix element (i,j) represents the number of co-occurrences of gray values ​​i and j under the specified direction and distance conditions. By normalizing the matrix, it is converted into a probability matrix to extract texture features such as energy, contrast, entropy and homogeneity, so as to achieve effective description and analysis of image texture.

[0091] (4-3) Vegetation Index

[0092] Using the vegetation index based on the red-green-blue (RGB) model, the vegetation pixels in the results of (4-2) are extracted and these plant patches are filtered out to further reduce the false positives caused by vegetation such as lawns and tree canopies.

[0093] (5) Remove redundant regions of interest

[0094] (5-1) Noise filtering.

[0095] If the size of the region of interest is too small, for example, the total number of pixels is less than 20, it is regarded as noise and deleted.

[0096] (5-2) Remove duplicates.

[0097] After completing the above ROI screening, the small rectangles that are completely contained in the larger rectangle are also removed to avoid redundant segmentation. The final set of ROIs will be used for subsequent resolution allocation and adaptive image transmission.

[0098] (6) Region of interest segmentation.

[0099] Separate the useful region of interest from the background, and save the original image size and the position of the rectangular block of the region of interest in the original image. Perform a full black mask on the block position of the region of interest in the original image to generate an image of the non-region of interest.

[0100] Step 3: Transmit the image of the region of interest, the image of the non-region of interest, the size of the original image, and the position of the region of interest in the original image to the second device via a downlink, so that the second device can restore the original image.

[0101] (1) Define the transmission resolution level of non-interest area tiles.

[0102] Four resolution levels are defined, namely L (lowest resolution), M1, M2 and H (original image resolution), where L represents the lowest acceptable resolution for the imaging task and H represents the highest resolution.

[0103] (2) Determine the time step for each bandwidth prediction.

[0104] The transmission frame rate f is determined based on the frequency of satellite image capture and orbit dynamics. After the frame rate f is determined, the time step of each bandwidth prediction is dt = 1 / f.

[0105] (3) Downlink transmission bandwidth prediction.

[0106] (3-1) Select a regression fitting model for bandwidth prediction, such as the LightGBM model. LightGBM is a strong prediction model based on a gradient boosting framework, which is constructed by gradually adding weak learners (usually decision trees).

[0107] (3-2) The downlink transmission bandwidth values ​​of the current 20 historical moments are used as the input features of the current prediction, and the rolling mean (RM) and rolling standard deviation (RSD) are used as additional features.

[0108] (3-3) Z-score normalization is performed and grid search is used to optimize the parameters of the prediction model used in (9-1) to select the best prediction performance settings.

[0109] (3-4) During the task execution, the trained prediction model used in (9-3) is used to predict the bandwidth b(t+dt) at the next moment in real time.

[0110] (4) Determine whether the bandwidth predicted at the next moment is a sudden increase bandwidth.

[0111] If the predicted bandwidth surge is considered temporary, the current resolution is kept unchanged. Among them, the bandwidth abnormal fluctuation detection algorithm analyzes the bandwidth fluctuation of the satellite downlink through the sliding window method. The algorithm divides the bandwidth data into multiple fixed-size windows according to the time series and calculates the standard deviation in each window. When the difference between the current bandwidth and the window standard deviation exceeds the set threshold, the algorithm determines it as an abnormal fluctuation. In this way, the algorithm can monitor the sudden change of bandwidth in real time and identify abnormal fluctuations in time. In order to avoid misjudging the continuous bandwidth trend change as abnormal, the algorithm introduces a trend detection mechanism to identify the normal increase and decrease trend of bandwidth. If the bandwidth suddenly changes and changes continuously, the algorithm regards it as a normal trend. Only when the direction of bandwidth change reverses, it will re-judge whether there is an abnormal fluctuation. This mechanism combines the sliding window and standard deviation methods to ensure that the algorithm can adapt to normal fluctuations and sensitively detect sudden abnormal fluctuations.

[0112] (5) Determine the maximum resolution.

[0113] When the bandwidth at the next moment is not a sudden increase, the maximum available resolution r is calculated based on the predicted bandwidth b(t+dt), the time step dt and the amount of original image data. max .

[0114] (6) Adjust the resolution.

[0115] Select less than or equal to r max To ensure that the image can be fully transmitted under the predicted bandwidth. Specifically, if r max ∈[a,b), where a,b∈[L,M1,M2,H], and select candidate resolutions based on a. For example, when a=M2, three resolution levels L, M1 and M2 can be selected as candidate resolutions. Then, a custom excitation factor (IF) is calculated for each candidate resolution. The optimal resolution is selected based on the maximum value of IF. IF considers the resolution quality score and the resolution switching difference as a measure of the amount of information and transmission smoothness, and its calculation formula is as follows:

[0116] IF(t)=Score[r(t)] 2 -μ|Score[r(t)]-Score[r(t-1)]| 2

[0117] Where r(t) represents the resolution, Score[r(t)] is the resolution quality score, and μ is a constant.

[0118] In one embodiment, if r max<M1, the resolution can be directly set to the lowest acceptable resolution L.

[0119] Figure 3 It is the prediction result of the downlink transmission bandwidth and the corresponding transmission resolution allocation result of the non - interested region in an adaptive image transmission planning method based on image content and downlink transmission bandwidth. From the data of historical and current moments, the red curve is the actual downlink transmission bandwidth value, and the blue curve is the bandwidth prediction result using the LightGBM algorithm. It can be seen that the prediction effect is excellent and there is no obvious time - delay impact, which proves the feasibility and reliability of the prediction algorithm. From the prediction of future moments, the black curve is the bandwidth prediction result of the LightGBM algorithm at future moments. Since there is no true - value reference at future moments, there is no curve representing the actual value. The green dashed line is the smoothed trend result after removing jitter from the LightGBM algorithm prediction result.

[0120] (7) Image transmission.

[0121] For the region of interest, high - resolution transmission is always retained to ensure the integrity of image semantic information. For the non - interested region, the above - mentioned adaptive resolution adjustment method is used to dynamically adjust the resolution according to the real - time bandwidth change, reducing the amount of transmitted data.

[0122] (8) Image stitching and synthesis at the ground end.

[0123] Using the original image size, the non - interested region tiles at the ground end are upsampled to the original resolution size, and then, using the position of the rectangular block of the region of interest in the original image, the high - resolution region - of - interest tiles are stitched with the processed non - interested region tiles.

[0124] Figure 4 They are three instances of fused and stitched pictures transmitted to the ground end through an adaptive image transmission planning method based on image content and downlink transmission bandwidth. It can be seen from the figure that the part of the target region of interest is clear at the original high resolution, and the unimportant background region is at the downsampled low resolution.

[0125] Those skilled in the art will readily think of other embodiments of the present disclosure after considering the specification and practicing the present disclosure. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include the common general knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and embodiments are only regarded as exemplary, and the present disclosure is not limited to the exact structures already described and shown in the drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. An adaptive image transmission planning method based on image content and downlink transmission bandwidth, characterized in that: Applied to a first device, the method comprises: Get the size and resolution of the original image; Generate an image of the region of interest and an image of the region of non-interest by extracting the region of interest in the original image, and obtain the position of the region of interest in the original image; wherein the resolution of the image of the region of interest is the resolution of the original image, and the resolution of the image of the region of non-interest is generated based on the downlink transmission bandwidth; The image of the region of interest, the image of the non-region of interest, the size of the original image, and the position of the region of interest in the original image are transmitted to the second device via a downlink, so that the second device restores the original image.

2. The method according to claim 1, characterized in that The step of extracting the region of interest from the original image, generating a region of interest image and a region of non-interest image, and obtaining the position of the region of interest in the original image includes: Preprocess the original image; Performing multi-scale processing on the pre-processed original image using a multi-scale Gaussian smoothing pyramid to obtain a multi-scale image of the original image; Based on the multi-scale image, generating a mixed saliency map of the original image; Screening and removing redundancy of regions of interest in the mixed saliency map; Separating the region of interest from the background of the original image to obtain an image of the region of interest and a position of the region of interest in the original image; The image of the non-interest region is generated by performing a full black mask processing on the tile position of the interest region in the original image, and calculating the resolution of the image of the non-interest region based on the downlink transmission bandwidth.

3. The method according to claim 2, characterized in that The preprocessing of the original image comprises: Perform adaptive downsampling preprocessing according to the size of the original image; Based on the HSV color space of the adaptive downsampling preprocessing result, the shadow area is marked, and the HSV value of the shadow area is replaced by the average HSV value of the adaptive downsampling preprocessing result.

4. The method according to claim 2, characterized in that: Based on the multi-scale image, generating a mixed saliency map of the original image, comprising: Based on a global contrast saliency calculation method, a multi-scale saliency map pyramid corresponding to the multi-scale image is calculated; wherein the global contrast saliency calculation method includes: an LC algorithm; Using image entropy as a metric, evaluating the saliency of each scale saliency map in the multi-scale saliency map pyramid, and performing weighted processing on the saliency map of each scale based on the saliency to generate a saliency map of the original image; After edges and fine structures in the original image are identified using the Canny edge detection algorithm, a mixed saliency map is generated based on the edge detection result and the saliency map of the original image.

5. The method according to claim 2, characterized in that: Screening and removing redundancy of regions of interest in the mixed saliency map include: Binarizing the mixed saliency map to obtain a region of interest in the mixed saliency map; Based on the YOLOv5 model, a semantic segmentation result of the original image is obtained, and based on the image entropy and inverse moment of the semantic segmentation result, the region of interest in the mixed saliency map that does not conform to the eigenvalue distribution is removed to obtain a first screening result of the region of interest; Using the vegetation index based on the red-green-blue model, the first screening result of the area of ​​interest is screened to obtain a second screening result of the area of ​​interest; According to the size of the region of interest and the position between the regions of interest, the regions of interest in the redundancy of the second screening result of the regions of interest are deleted.

6. The method according to claim 1, characterized in that The resolution of the non-interest region image is generated based on the downlink transmission bandwidth, including: Define the transmission resolution level of the image of non-interest areas; Determine the time step for predicting downlink transmission bandwidth based on the satellite's image acquisition frequency and orbital dynamics; Obtaining the downlink transmission bandwidth at several historical moments, and using a regression fitting model to predict the downlink transmission bandwidth at the next moment; wherein the regression fitting model includes: a LightGBM model; Determine whether the downlink transmission bandwidth at the next moment is a sudden increase in bandwidth; In the case that the downlink transmission bandwidth at the next moment is not a sudden increase in bandwidth, based on the downlink transmission bandwidth at the next moment, the time step of predicting the downlink transmission bandwidth and the data amount of the original image, calculating the maximum available resolution, and determining the resolution of the non-interest region image in the transmission resolution level according to the maximum available resolution; In the case that the downlink transmission bandwidth at the next moment is a sudden increase in bandwidth, the resolution of the non-interesting image transmitted last time is used as the resolution of the non-interesting image.

7. The method according to claim 6, characterized in that The step of obtaining the downlink transmission bandwidth at a plurality of historical moments and predicting the downlink transmission bandwidth at the next moment by using a regression fitting model includes: Calculate the rolling mean and rolling standard deviation of the downlink transmission bandwidth at several historical moments; Perform Z-score normalization on the downlink transmission bandwidth at several historical moments and the corresponding rolling mean and rolling standard deviation; The standardized data is input into a trained regression fitting model to obtain the downlink transmission bandwidth at the next moment; wherein the parameters in the trained regression fitting model are obtained by grid search optimization.

8. The method according to claim 6, characterized in that The determining whether the downlink transmission bandwidth at the next moment is a sudden increase in bandwidth includes: Generate a time series according to the downlink transmission bandwidth at a historical moment and the downlink transmission bandwidth at a next moment; The time series data is divided into a plurality of windows of fixed size by a sliding window, and the standard deviation in each window is calculated; If the standard deviation of the window where the downlink transmission bandwidth at the next moment is located and the standard deviation of other windows is greater than the set threshold, and the bandwidth increase and decrease trend is reversed, it is determined that the downlink transmission bandwidth at the next moment is a bandwidth sudden increase.

9. The method according to claim 6, characterized in that The step of determining the resolution of the non-interest region image in the transmission resolution level according to the available maximum resolution comprises: Among the transmission resolution levels, obtaining a level that is not greater than the available maximum resolution; Calculating an excitation factor for each level not greater than the maximum resolution available; According to the maximum value of the excitation factor, a resolution level is determined in a level greater than the maximum available resolution to obtain the resolution of the image of the non-interest region.

10. An adaptive image transmission planning system based on image content and downlink transmission bandwidth, characterized in that: The system comprises: The first device is used to obtain the size and resolution of the original image; generate an image of the region of interest and an image of the region of non-interest by extracting the region of interest in the original image, and obtain the position of the region of interest in the original image; wherein the resolution of the image of the region of interest is the resolution of the original image, and the resolution of the image of the region of non-interest is generated based on the downlink transmission bandwidth; and transmit the size of the image of the region of interest, the image of the region of non-interest, the original image, and the position of the region of interest in the original image to the second device via the downlink; The second device is used to restore the original image based on the image of the region of interest, the image of the non-region of interest, the size of the original image, and the region of interest in the original image.

Images