A budget-aware visual feature compression method for remote sensing image understanding

Abstract

The application discloses a budget-aware visual feature compression method for remote sensing image understanding, comprising: acquiring a remote sensing image and a text query corresponding to the remote sensing image, and converting the text query into a text feature; constructing the remote sensing image into multiple scale image views, and encoding each scale image view; obtaining visual feature importance by calculating attention weights from the text feature to the visual feature; determining an anchor scale, mapping the visual feature importance of the anchor scale to other scales to obtain aligned visual feature importance of each scale; allocating scale budgets for each scale according to a global visual feature budget, and dividing the visual feature into multiple regions within each scale; calculating regional average importance in each region to form a candidate visual feature sequence; sorting the candidate visual feature sequence, and selecting visual features according to the corresponding scale budget to obtain a compressed visual feature sequence satisfying the global visual feature budget.

Description

Technical Field

[0001] This invention discloses a budget-aware visual feature compression method for remote sensing image understanding, belonging to the fields of computer vision and multimodal artificial intelligence technology. Background Technology

[0002] Multimodal large language models typically consist of a visual encoder, a visual-language connection module, and a language model. They can receive image and text queries and output natural language responses. In remote sensing scenarios, satellite images, aerial images, and similar images are characterized by large coverage areas, high resolution, and a wide range of target sizes. The model needs to understand both kilometer-level scene layouts and identify local details such as vehicles, ships, islands, and road boundaries, which may occupy only a few pixels. Ultra-high resolution remote sensing images refer to remote sensing images with large spatial coverage and / or high spatial resolution, such that the number of visual features generated after encoding by the visual encoder exceeds the preset visual feature budget or the effective context processing capability of the visual language model. Ultra-high resolution remote sensing images can include satellite images, aerial images, UAV images, etc., and they typically contain both large-scale scene structures and fine-grained targets or local evidence occupying only a few pixels.

[0003] Existing visual language models typically employ downsampling, dense tiling, or global Top-K pruning when processing ultra-high-resolution remote sensing images. Direct downsampling loses small objects and boundary details; while dense tiling preserves local information, it drastically increases the number of visual features, leading to significant attention computation and memory overhead, and potentially disrupting the global context. Global Top-K pruning tends to concentrate the budget in salient regions, ignoring sparse but query-relevant key evidence.

[0004] Therefore, a feature compression method that can combine text queries, cross-scale visual information, and regional structure is needed to enable the model to retain the global context and local fine-grained evidence relevant to the problem under strict visual feature budget, thereby improving the accuracy and inference efficiency of understanding ultra-high resolution remote sensing images. Summary of the Invention

[0005] Purpose of the invention: To address the problem that existing ultra-high resolution remote sensing visual language understanding methods struggle to balance global context, local details, and computational efficiency under a fixed visual feature budget, this invention proposes a budget-aware visual feature compression method for remote sensing image understanding. This method utilizes text queries to guide visual feature importance estimation and employs a region-fidelity retention and merging mechanism to retain query-related evidence and compress redundant background information under a fixed budget, making it suitable for processing ultra-high resolution remote sensing images.

[0006] Technical solution: In a first aspect, this invention proposes a budget-aware visual feature compression method for remote sensing image understanding, comprising:

[0007] Acquire remote sensing images and corresponding text queries, and convert the text queries into text features;

[0008] The remote sensing image is constructed into image views at multiple scales, and the visual encoder in the visual language model is used to encode the image views at each scale to obtain the visual features at each scale.

[0009] A cross-modal attention mechanism based on a visual language model is used to calculate the attention weights from text features to visual features, and to obtain the importance of visual features based on the attention weights. The importance of visual features is used to characterize the degree of correlation between the corresponding visual features and the text query. Anchor scales are determined from multiple scales, and the importance of visual features at the anchor scales is used as the reference importance for cross-scale alignment, thereby obtaining the aligned visual feature importance at each scale.

[0010] Scale budgets are allocated to each scale based on the global visual feature budget, and visual features are divided into multiple non-overlapping regions within each scale. Based on the importance of aligned visual features, the average importance of each region is calculated. Visual features whose importance after alignment is not lower than the average importance of the region are retained. All visual features whose importance is lower than the average importance of the region are merged as representative features of the region. The retained visual features are concatenated with the representative features of the region to form a candidate visual feature sequence.

[0011] The candidate visual feature sequences are sorted according to the average importance of the region, and visual features are selected from the sorting results according to the corresponding scale budget to obtain a compressed visual feature sequence that satisfies the global visual feature budget.

[0012] The compressed visual feature sequence and the text features are input into the language model in the visual language model, and the language model generates an answer to the text query.

[0013] Furthermore, the visual feature importance obtained based on the attention weights is expressed as follows:

[0014]

[0015] Where M represents the number of text features. This represents the attention weight of the j-th text feature to the i-th visual feature. This indicates the importance of the i-th visual feature.

[0016] Furthermore, the visual encoder in the visual language model encodes the image views at each scale to obtain visual features at each scale, including:

[0017] For the first Encode the image view at each scale to obtain the first scale. Visual features at various scales are obtained, and two-dimensional positional embedding and scale embedding are added to these visual features to obtain visual features enhanced at scale-specific locations, as shown below:

[0018]

[0019] in, Let i be the i-th visual feature at the s-th scale. The position embedding is obtained by two-dimensional position embedding interpolation. For the scale embedding corresponding to the s-th scale, Indicates the number after enhancement at a specific location based on scale. The first scale A visual feature.

[0020] Furthermore, the step of determining the anchor point scale from multiple scales, and using the visual feature importance of the anchor point scale as a reference importance for cross-scale alignment, thereby obtaining the aligned visual feature importance for each scale, includes:

[0021] The visual feature importance at the anchor point scale is rearranged into a two-dimensional importance map. For any grid position at the s-th scale... , will the At any grid location on any scale Mapped to anchor point scale coordinates, and based on the two-dimensional importance map, bilinear interpolation is used to obtain the visual feature importance after alignment at any grid position on the s-th scale. ;

[0022] Let the set of grids covered by the i-th visual feature at the s-th scale be . The importance of visual features after alignment The following formula is used to calculate:

[0023]

[0024] in, Indicates the first Network location at each scale The importance of aligned visual features Indicates the first The first of the scales The importance of visual features after alignment.

[0025] Furthermore, the step of allocating scale budgets to each scale based on the global visual feature budget, and dividing the visual features into multiple non-overlapping regions within each scale, includes:

[0026] Set the global visual feature budget to Scale budgets are assigned to each scale based on the global visual feature budget, and the scale budget for the s-th scale is... satisfy ;

[0027] Within each scale, visual features are divided into multiple non-overlapping regions, satisfying... , Let be the number of regions within the s-th scale.

[0028] Furthermore, based on the importance of aligned visual features, the average importance of the region is calculated within each region, expressed as:

[0029]

[0030] In the formula, This represents the m-th region at the s-th scale. This represents the average regional importance of the m-th region at the s-th scale. This indicates the importance of visual features after alignment.

[0031] Furthermore, the step of sorting the candidate visual feature sequences according to the average importance of the regions, and selecting visual features from the sorting results according to the corresponding scale budget, includes:

[0032] First, sort the regions in descending order according to their average importance.

[0033] Secondly, within each region, the retained visual features are sorted in descending order according to the importance of the aligned visual features, and the representative features of the region are added to the candidate visual feature sequence corresponding to that region to obtain the sorting result.

[0034] Finally, visual features are selected from the sorting results according to the scale budget.

[0035] Secondly, the present invention provides an electronic device, the electronic device comprising:

[0036] At least one processor;

[0037] and a memory communicatively connected to the at least one processor;

[0038] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform a budget-aware visual feature compression method for remote sensing image understanding.

[0039] Thirdly, the present invention proposes a computer-readable storage medium storing computer instructions for causing a processor to execute a budget-aware visual feature compression method for remote sensing image understanding.

[0040] Fourthly, the present invention proposes a computer program product, which includes a computer program that, when executed by a processor, implements a budget-aware visual feature compression method for remote sensing image understanding.

[0041] Beneficial Effects: This invention acquires ultra-high resolution remote sensing images and corresponding text queries; constructs image views at multiple scales and obtains visual features at each scale through a visual encoder; calculates the importance of visual features perceived in the query based on text-to-visual attention in a visual language model; obtains the importance of features at each scale through cross-scale interpolation alignment, using an importance map at a low-resolution anchor scale as a reference; allocates visual feature budgets to each scale and divides the features at each scale into multiple semantically consistent regions; retains features with importance no less than the average importance of the region within each region, and merges the remaining redundant features into representative features of the region; sorts candidate features according to regional importance and feature importance, and selects a compressed visual feature sequence under budget constraints; inputs the compressed visual feature sequence and text features into a language model to generate a response. Based on the above technical solution, this invention has the following advantages compared with the prior art:

[0042] (1) The present invention introduces text query information into the visual feature compression process, which can prioritize the retention of visual evidence related to the current problem and reduce the budget occupied by background features unrelated to the query.

[0043] (2) The present invention constructs a multi-scale image view and enhances the scale-specific location and cross-scale importance alignment, while taking into account both the global scene structure and local fine-grained information.

[0044] (3) The present invention proposes a region retention and merging mechanism, which retains relatively important visual features in each region and merges redundant features into compact representations, thereby reducing the consumption of visual feature budget by repetitive background information.

[0045] (4) This invention maintains global context and local fine-grained features while satisfying a fixed visual feature budget, and outputs a compressed visual feature sequence, which helps to reduce the attention computation overhead and memory usage at the language model end.

[0046] (5) This invention is applicable to tasks such as ultra-high resolution remote sensing visual question answering, target position relationship judgment, counting, ground object recognition, and scene understanding, and can also be extended to other high resolution image understanding scenarios. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall process of the budget-aware visual feature compression method for remote sensing image understanding proposed in this embodiment of the invention. Detailed Implementation

[0048] The visual language model in this embodiment of the invention includes a visual encoder, a visual language connection module, and a language model. The visual encoder extracts visual features from an image view; the visual language connection module maps these visual features to an embedding space that the language model can process; and the language model generates natural language responses based on textual and visual features. Therefore, the language model used in this embodiment is the language model within the visual language model, and not a separate model independent of the visual language model.

[0049] The technical solution of the present invention will now be further described in conjunction with the accompanying drawings and embodiments.

[0050] Example 1:

[0051] like Figure 1 As shown in the figure, this invention proposes a budget-aware visual feature compression method for understanding ultra-high resolution remote sensing images, including the following steps:

[0052] Step 1: Obtain the ultra-high resolution remote sensing image I and the text query x corresponding to the ultra-high resolution remote sensing image I, and input the text query x into the text embedding layer or text segmenter to obtain the text feature T.

[0053] Step 2: Construct the ultra-high resolution remote sensing image I into S scale image views { The lowest resolution scale is used as the anchor scale to preserve the global structure, while the high resolution scale is used to provide fine-grained information. A visual encoder is used to encode the image view at each scale to obtain the visual features at each scale, and then map them to the language model embedding space.

[0054] To avoid confusion in spatial location and scale semantics between visual features from different scales, the i-th visual feature at the s-th scale is obtained through scale-specific location enhancement:

[0055]

[0056] in, Let i be the i-th visual feature at the s-th scale. The position embedding is obtained by interpolation of the two-dimensional basic position embedding. Let be the learnable scale embedding corresponding to the s-th scale. Indicates the number after enhancement at a specific location based on scale. The first scale There are several visual features. Location embedding is used to represent the spatial position of the visual feature within the corresponding image view, while scale embedding is used to represent the image scale to which the visual feature belongs. The visual features, after incorporating location and scale embeddings, are then used for subsequent attention weight calculation, importance estimation, region segmentation, and feature compression.

[0057] Step 3: Obtain the attention matrix A from text features T to visual features from the cross-modal attention calculation process of the visual language model, and aggregate it according to the text feature dimension to obtain the visual feature importance score.

[0058] As one implementation method, the visual feature importance score can be calculated using the following formula:

[0059]

[0060] Where M represents the number of text features. This represents the attention weight of the j-th text feature to the i-th visual feature. This represents the importance of the i-th visual feature. When multiple scales are involved, the i-th... The visual feature refers to the first visual feature at the current scale. The first visual feature, that is, the second... The first scale A visual feature.

[0061] In this embodiment of the invention, the attention matrix is ​​used to characterize the degree of attention paid by text features to corresponding visual features, and the visual feature importance is used to characterize the degree of correlation between the corresponding visual features and the text query.

[0062] Step 4: Using the anchor scale among multiple scales as the reference scale, map the visual feature importance of the anchor scale to other scales to obtain the cross-scale aligned visual feature importance.

[0063] As one implementation method, the importance of visual features at the anchor point scale is rearranged into a two-dimensional importance graph. For the grid location (u,v) at the s-th scale, a deterministic scaling mapping is used. Mapping to anchor point scale coordinates and using bilinear interpolation, the importance of aligned visual features at the grid position (u,v) on the s-th scale is obtained. , is represented as:

[0064]

[0065] in, Indicates the first Grid positions at each scale The importance of visual features after alignment A two-dimensional importance plot representing the anchor point scale. For the deterministic scaling mapping from the target mesh at the s-th scale to the anchor-scale mesh, This indicates bilinear interpolation.

[0066] The set of grids covered by the i-th visual feature at the s-th scale is The importance of visual features after alignment The following formula is used to calculate:

[0067]

[0068] In the formula, Represented as the first The first of the scales A set of grids covered by visual features. Indicates the first Network location at each scale The importance of aligned visual features Indicates the first The first of the scales The importance of visual features after alignment. These correspond to grid position level importance and visual feature level importance, respectively.

[0069] Step 5: Assign scale budgets to each scale based on the global visual feature budget B. , satisfy Within each scale, visual features are divided into regions, forming several non-overlapping regions. The s-th scale is set... ,in Let be the number of regions within the s-th scale, such that each region can output at least one representative feature.

[0070] As one implementation method, region partitioning can be achieved through clustering in a joint space composed of visual features and two-dimensional coordinates, or by generating region masks using a segmentation model and mapping them to a visual feature grid. Region partitioning is used to aggregate semantically consistent and spatially adjacent visual features, enabling the compression process to not only rely on global importance ranking but also retain representative information about the regions where sparse targets are located.

[0071] Step 6: Calculate the average importance of each region, retain visual features with importance no less than the average importance of the region, and merge the remaining redundant features into representative features of the region through mean pooling to obtain the candidate visual feature sequence.

[0072] In the process of region preservation and merging, for the m-th region at the s-th scale The regional average importance is:

[0073]

[0074] Retain satisfaction The visual features constitute the set of retained features. Redundant sets These are the visual features that were not retained. The set of redundant features that were not retained. The features are merged into representative regional features through mean pooling:

[0075]

[0076] in .

[0077] Step 7: Sort the candidate visual feature sequences according to regional importance, and select visual features according to the corresponding scale budget to obtain a compressed visual feature sequence that satisfies the global visual feature budget.

[0078] As one implementation method, the regions are first sorted in descending order according to their average importance. Then, within each region, the retained features are sorted in descending order according to their importance after alignment, and the retained features are concatenated with the region's representative features to form a candidate visual feature subsequence for that region. Next, the candidate visual feature subsequences of each region are concatenated according to the region sorting results to form a candidate visual feature sequence. Finally, the sequence is calculated according to scale estimation. Before choosing The visual feature is obtained as the first... A compressed visual feature sequence at each scale.

[0079] Step 8: Concatenate the compressed visual feature sequence with the text features and input the concatenation into the language model in the visual language model. The language model will then output the answer to the text query.

[0080] Example 2:

[0081] This embodiment uses an ultra-high resolution remote sensing visual question answering task as an example. The input includes an ultra-high resolution remote sensing image and a natural language query, such as asking about the positional relationship of a small island relative to an aquaculture raft. Since the original image size is large, directly inputting all visual features into the language model would result in excessively high context length and computational cost. Therefore, this embodiment uses the budget-aware compression method proposed in Embodiment 1 before the visual features are input into the language model.

[0082] First, the original image is scaled down into multiple image views. Low-resolution views are used to capture large-scale structures such as coastlines, islands, roads, and farmland, while high-resolution views are used to preserve fine-grained evidence such as boats, vehicles, islets, and building boundaries. Visual features are obtained from each scale image view through a frozen or trainable visual encoder, and then mapped to the language model embedding space through a visual language projection layer.

[0083] Secondly, the attention relationship between text queries and visual features is calculated at the anchor point scale. By aggregating the attention weights from text features to visual features, an importance map of visual features relevant to the current query can be obtained. This importance map reflects the degree of attention the text query pays to different image regions, for example, giving higher weights to regions related to entities such as "small islands" and "fishing rafts".

[0084] Furthermore, the importance map at the anchor scale is propagated to other scales via bilinear interpolation, ensuring that features at high-resolution scales also receive comparable importance scores. This avoids redundant, costly attention calculations at each scale while guaranteeing that features at different scales participate in budget allocation according to a unified standard.

[0085] Then, region segmentation is performed at each scale. Region segmentation can employ clustering methods such as K-means to divide semantically similar and spatially continuous regions in a joint space composed of visual features and normalized two-dimensional coordinates; alternatively, an image segmentation model can be used to obtain a mask, which is then mapped onto a visual feature grid. The average importance of each region is calculated, and features with importance at least equal to this average are retained. Redundant features not retained are not directly discarded but are merged into a single region representative feature through mean pooling, thereby reducing the number of features while preserving coarse-grained context.

[0086] Finally, the preserved features and representative features of each region are sequenced according to priority. Specifically, regions are first arranged from highest to lowest average importance, then the preserved features are arranged from highest to lowest feature importance within each region, and the merged representative features are appended to the end of the corresponding region. This selection is performed under scale budget constraints. The system extracts 10 features, ultimately resulting in a compressed visual feature sequence that satisfies the global budget B. This sequence, along with the text features, is input into the language model to generate the final answer.

[0087] In this embodiment of the invention, the number of multi-scale views S, the global visual feature budget B, and the scale budget for each scale are... and number of regions The settings can be adjusted based on device computing power, language model context length, image resolution, and task accuracy requirements. Higher resolution scales can allocate more budget to preserve local details, while also setting... This ensures that each region has at least one representative feature that can be used in subsequent reasoning.

[0088] To verify the effectiveness of this embodiment, it was tested on three ultra-high resolution remote sensing image understanding benchmarks: XLRS-Bench, RSHR-Bench, and MME-RealWorld-RS. The experiments used a unified prompt template inference setting, primarily comparing the model's accuracy on remote sensing visual question answering, fine-grained perception, and spatial reasoning tasks. Tables 1 to 3 present the experimental results on the three datasets, with all values ​​representing percentages of accuracy.

[0089] Table 1 Performance of the ultra-high resolution remote sensing visual question answering task on the XLRS-Bench dataset

[0090]

[0091] As shown in Table 1, the method in this embodiment achieves an overall weighted accuracy of 44.0% at the 7B model scale, which is higher than representative baselines such as GeoLLaVA-8K and Qwen2.5-VL-7B. This indicates that it can more effectively preserve global context and local evidence related to text queries under a fixed visual feature budget.

[0092] Table 2 Performance of high-resolution remote sensing and inference tasks on the RSHR-Bench dataset.

[0093]

[0094] As shown in Table 2, the method in this embodiment achieves a perception mean of 29.2% and an inference mean of 45.0% on RSHR-Bench, with a combined mean of 34.8%. Compared to remote sensing-specific models such as EarthDial, GeoChat, GeoLLaVA-8K, and VHM, this embodiment demonstrates better stability in both perception and inference tasks.

[0095] Table 3 Performance of location, color, and counting tasks on the MME-RealWorld-RS dataset.

[0096]

[0097] As shown in Table 3, the method in this embodiment achieves an average accuracy of 33.33% on MME-RealWorld-RS, and 44.00% and 42.00% on location understanding and color recognition tasks, respectively. This indicates that the query-aware importance estimation and region preservation merging mechanism can preserve fine-grained evidence such as small targets, location relationships, and ground feature colors in ultra-high resolution remote sensing images.

[0098] In summary, the experimental results in Tables 1 to 3 show that the method of this embodiment has good accuracy performance on multiple ultra-high resolution remote sensing image understanding benchmarks. In particular, it can take into account global scene structure, local fine-grained evidence, and task-related region coverage under limited visual feature budget, thus verifying the effectiveness of the present invention.

Claims

1. A budget-aware visual feature compression method for remote sensing image understanding, characterized in that: include: Acquire remote sensing images and corresponding text queries, and convert the text queries into text features; The remote sensing image is constructed into image views at multiple scales, and the visual encoder in the visual language model is used to encode the image views at each scale to obtain the visual features at each scale. A cross-modal attention mechanism based on a visual language model is used to calculate the attention weights from text features to visual features, and to obtain the importance of visual features based on the attention weights. The importance of visual features is used to characterize the degree of correlation between the corresponding visual features and the text query. Anchor scales are determined from multiple scales, and the importance of visual features at the anchor scales is used as the reference importance for cross-scale alignment, thereby obtaining the aligned visual feature importance at each scale. Scale budgets are allocated to each scale based on the global visual feature budget, and visual features are divided into multiple non-overlapping regions within each scale. Based on the importance of aligned visual features, the average importance of each region is calculated. Visual features whose importance after alignment is not lower than the average importance of the region are retained. All visual features whose importance is lower than the average importance of the region are merged as representative features of the region. The retained visual features are concatenated with the representative features of the region to form a candidate visual feature sequence. The candidate visual feature sequences are sorted according to the average importance of the region, and visual features are selected from the sorting results according to the corresponding scale budget to obtain a compressed visual feature sequence that satisfies the global visual feature budget. The compressed visual feature sequence and the text features are input into the language model in the visual language model, and the language model generates an answer to the text query.

2. The budget-aware visual feature compression method for remote sensing image understanding according to claim 1, characterized in that: The importance of visual features obtained based on the attention weights is expressed as follows: ； Where M represents the number of text features. This represents the attention weight of the j-th text feature to the i-th visual feature. This indicates the importance of the i-th visual feature.

3. The budget-aware visual feature compression method for remote sensing image understanding according to claim 1, characterized in that: The method employs a visual encoder in a visual language model to encode image views at each scale, obtaining visual features at each scale, including: For the Encode the image view at each scale to obtain the first scale. Visual features at various scales are obtained, and two-dimensional positional embedding and scale embedding are added to these visual features to obtain visual features enhanced at scale-specific locations, as shown below: ； in, Let i be the i-th visual feature at the s-th scale. The position embedding obtained by two-dimensional position embedding interpolation, For the scale embedding corresponding to the s-th scale, Indicates the first [unit] after enhancement at a specific scale location. The first scale A visual feature.

4. The budget-aware visual feature compression method for remote sensing image understanding according to claim 1, characterized in that: The process of determining anchor scales from multiple scales, using the visual feature importance of the anchor scales as a reference importance for cross-scale alignment, and thereby obtaining the aligned visual feature importance for each scale, includes: The visual feature importance at the anchor point scale is rearranged into a two-dimensional importance map. For any grid position at the s-th scale... , will the At any grid location on any scale Mapped to anchor point scale coordinates, and based on the two-dimensional importance map, bilinear interpolation is used to obtain the visual feature importance after alignment at any grid position on the s-th scale. ; Let the set of grids covered by the i-th visual feature at the s-th scale be . The importance of visual features after alignment The following formula is used to calculate: ； in, Indicates the first Network location at each scale The importance of aligned visual features Indicates the first The first of the scales The importance of visual features after alignment.

5. The budget-aware visual feature compression method for remote sensing image understanding according to claim 1, characterized in that: The process of allocating scale budgets to each scale based on global visual feature budgets, and dividing visual features into multiple non-overlapping regions within each scale, includes: Set the global visual feature budget to Scale budgets are assigned to each scale based on the global visual feature budget, and the scale budget for the s-th scale is... satisfy ; Within each scale, visual features are divided into multiple non-overlapping regions, satisfying... , Let be the number of regions within the s-th scale.

6. The budget-aware visual feature compression method for remote sensing image understanding according to claim 1, characterized in that: The average importance of a region within each region is calculated based on the importance of aligned visual features, and is expressed as follows: ； In the formula, This represents the m-th region at the s-th scale. This represents the average regional importance of the m-th region at the s-th scale. This indicates the importance of visual features after alignment.

7. The budget-aware visual feature compression method for remote sensing image understanding according to claim 1, characterized in that: The step of sorting the candidate visual feature sequences according to the average importance of the regions and selecting visual features from the sorting results according to the corresponding scale budget includes: First, sort the regions in descending order according to their average importance. Secondly, within each region, the retained visual features are sorted in descending order according to the importance of the aligned visual features, and the representative features of the region are added to the candidate visual feature sequence corresponding to that region to obtain the sorting result. Finally, visual features are selected from the sorting results according to the scale budget.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform a budget-aware visual feature compression method for remote sensing image understanding according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute and implement the budget-aware visual feature compression method for remote sensing image understanding according to any one of claims 1-7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements a budget-aware visual feature compression method for remote sensing image understanding as described in any one of claims 1-7.

Images