A low-altitude remote sensing image small target interpretation and iterative correction method, device and medium

Abstract

This invention discloses a method, device, and medium for interpreting and iteratively correcting small targets in low-altitude remote sensing images. The method includes: using a pre-trained core algorithm model, performing block-by-block inference prediction on ultra-large resolution low-altitude remote sensing images using a sliding window strategy, and vectorizing the results; publishing a predicted image combining image tiles and vectorized prediction results on the application front-end using GeoServer or GIS services; manually correcting low-confidence or suspected erroneous targets in the predicted image, and writing the correction results into a vector database; and incorporating the corrected predicted image into a training set for incremental or fine-tuning training of the core algorithm model. This invention achieves high-precision automatic identification and continuous optimization of sparse small targets in ultra-large remote sensing images by integrating the DINOv3 self-supervised visual base model, the improved ViT-Adapter, and Cascade-RCNN multi-stage detection. It is particularly suitable for processing low-altitude remote sensing images with high resolution, large file size, complex terrain features, and sparse targets.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation and target detection technology of remote sensing images, and in particular to a method, device and medium for interpreting and iteratively correcting small targets in low-altitude remote sensing images. Background Technology

[0002] Low-altitude remote sensing technologies (such as UAV remote sensing) have been widely used in ground object monitoring and target identification scenarios due to their advantages of high flexibility, short data acquisition cycle, and high resolution. However, the interpretation of small targets (pixel size usually <50×50) in low-altitude remote sensing images still faces three major technical bottlenecks, and these three bottlenecks are intertwined, making it difficult for existing solutions to form an effective closed loop.

[0003] First, there is the data dilemma: the contradiction between massive amounts of data and fine-grained annotation. This is reflected in the limitations of large-scale image processing and annotation tools. Low-altitude remote sensing images often have a single image size exceeding 10GB, and are characterized by "high resolution but blurred local details, sparse targets, and small scale." Traditional general-purpose image annotation software cannot effectively load and process this data, requiring reliance on specialized GIS software such as ArcGIS and QGIS, which has a high operational threshold and cumbersome processes. Second, there is a conflict between geographic continuity and tile annotation. Sparse small targets such as power towers and cranes have the characteristics of "geographical continuity and structural coherence." Tile operations can destroy the integrity of the targets and their correlation with the surrounding environment, making it difficult for annotators to use contextual aids for identification, significantly increasing the risk of missed annotations. On the other hand, whole-map annotation places higher demands on both tools and manual work. Third, annotation costs are high. Without tool assistance, professionals must meticulously annotate every single point of tiny, sparse targets, resulting in a large workload, long processing time, and extremely low annotation efficiency. Based on these challenges across the data dimensions, relying solely on traditional tools is insufficient to establish an efficient annotation and processing workflow.

[0004] Secondly, there is a bottleneck in modeling. The gap between general-purpose models and specialized tasks is significant. Current mainstream target detection technologies have obvious limitations in interpreting small targets in low-altitude remote sensing. Convolutional neural networks (CNNs), relying on local receptive fields and inductive biases, perform stably in traditional computer vision tasks, but their detection accuracy and generalization ability are limited when facing remote sensing scenes with complex structures and large target scales due to their limited receptive fields. Transformer-type models, such as the Visual Transformer (ViT) model represented by DINOv3, possess powerful general visual representation and semantic understanding capabilities for remote sensing objects through self-supervised learning (SSL) on massive multimodal data. However, they are constrained by inherent defects in the ViT architecture: lack of image-related inductive biases, high computational complexity of self-attention mechanisms, and the absence of multi-scale hierarchical structures in feature maps. This results in poor performance in dense prediction tasks such as small target detection (high false negative rate, insensitive to small targets), making it difficult to directly apply to high-precision specialized interpretation scenarios.

[0005] Third, the technical solutions are fragmented and lack systematic integration: existing technologies often focus on optimizing a single algorithm, neglecting the most time-consuming core step of "data acquisition and annotation"; at the same time, data processing, pseudo-annotation, manual correction, and model iteration are scattered across different platforms / tools, failing to form a complete closed loop of "data processing → model training → annotation iteration → performance improvement". This non-closed-loop workflow leads to a separation between data and algorithms, resulting in poor compatibility and low efficiency when the technology is implemented.

[0006] For example, invention application No. 202510082463.5 discloses a real-time multi-source remote sensing image small target detection method using super-resolution assisted inference. This application removes medium and large-sized anchor frames during the detection process and adds dedicated anchor frames for small target detection, improving detection accuracy while reducing model inference time. However, it also suffers from a lack of systematic integration, inefficient processing of massive amounts of data, and a lack of systematic integration and utilization of data, making it difficult to guarantee stable detection performance when facing complex and ever-changing low-altitude remote sensing image scenarios. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a method, device, and medium for interpreting and iteratively correcting small targets in low-altitude remote sensing images. This addresses the contradiction between efficient processing and precise annotation of high-resolution, ultra-large file remote sensing images, the insufficient accuracy and weak generalization ability of general Transformer / CNN models in small target detection, and the fragmented and disjointed processes of data processing, algorithm training, and manual annotation, which make it difficult to form a collaborative optimization loop.

[0008] This invention provides a method, device, and medium for interpreting and iteratively correcting small targets in low-altitude remote sensing images.

[0009] First aspect: A method for interpreting and iteratively correcting small targets in low-altitude remote sensing images, including:

[0010] S1. Using the pre-trained core algorithm model, a sliding window strategy is adopted to perform block inference prediction on ultra-high resolution low-altitude remote sensing images, and the inference prediction results are vectorized.

[0011] S2. Utilize GeoServer or GIS services to publish the predicted image of the image tile joint vectorization prediction results on the application front end.

[0012] S3. Manually add, delete, or modify low-confidence or suspected erroneous targets in the predicted image and write the correction results of the predicted image into the vector database.

[0013] S4. Incorporate the corrected predicted images into the training set and perform incremental or fine-tuning training on the core algorithm model.

[0014] In one embodiment of the present invention, a low-altitude remote sensing image small target interpretation and iterative correction system applying the method described above includes:

[0015] The data processing and inference subsystem is used to acquire ultra-high resolution low-altitude remote sensing images. Based on the core algorithm model, it performs block-based inference and prediction using a sliding window strategy, and then vectorizes the inference and prediction results.

[0016] The interactive calibration and sample generation subsystem is used to provide GeoServer or GIS services, publish joint prediction images, generate a list of manual calibration tasks, manually calibrate the targets in the prediction images, write the calibration results into the vector database, and generate samples based on the calibration.

[0017] The core algorithm detection subsystem is used to train and provide the core algorithm model.

[0018] In one embodiment of the present invention, the data processing and reasoning subsystem includes:

[0019] The spatial indexing module is used to establish a unified coordinate projection and metadata parsing for ultra-high resolution low-altitude remote sensing images, and to provide a spatial index.

[0020] The sliding window inference module is used to divide the image into blocks by window using a parallel scheduling strategy;

[0021] The pseudo-annotation vectorization module is used to convert the detection boxes output by inference into GeoJSON / Shapefile format.

[0022] In one embodiment of the present invention, the interactive correction and sample generation subsystem includes:

[0023] The GIS service publishing module is used to publish pyramid-shaped image tiles, enabling the overlay display of image base maps and vector prediction results.

[0024] The intelligent filtering module is used to automatically filter candidate targets based on business rules and generate a list of manual correction tasks;

[0025] The manual interactive editing module uses vector tile technology and provides one-click jump to manually corrected targets, batch annotation templates, and historical version comparison functions;

[0026] The sample generation module automatically cropps the image based on the corrected vector box to generate samples that meet the training requirements.

[0027] In one embodiment of the present invention, the core algorithm detection subsystem, with a self-supervised trained DINOv3 visual Transformer structure as its backbone, adopts a multi-scale cascaded remote sensing target detection architecture based on DINOv3-Adapter multi-scale adaptation and Gram feedback Cascade-RCNN detection components, including:

[0028] The self-supervised DINOv3 backbone module is used for efficient image feature mapping, rotation position encoding (RoPE) spatial prior, deep global semantic modeling, and freezing feature weight benchmarks.

[0029] The DINOv3-Adapter multi-scale adaptation module is used for collaborative utilization of multiple types of features, ensuring the fidelity of relative position information, lightweight multi-scale injection, and achieving task-driven joint optimization.

[0030] The Gram-feedback Cascade-RCNN detection module is used for multi-scale feature reception, Gram feature quality evaluation, dynamic threshold control, loss closure and bidirectional optimization.

[0031] In one embodiment of the present invention, the Gram feature quality assessment is expressed by the formula:

[0032]

[0033] in, For quality assessment scores, For adapter output characteristics, Characteristics of frozen skeletal structure For Gram matrices, It is the Frobenius norm.

[0034] In one embodiment of the present invention, dynamic threshold control is performed based on... The score is used to adjust the IoU threshold of the three stages of Cascade in real time, as expressed by the formula:

[0035]

[0036] in, The IoU threshold is for the three stages.

[0037] In one embodiment of the present invention, the total loss function for achieving loss closure and bidirectional optimization is expressed as follows:

[0038]

[0039]

[0040] in, For Cascade-RCNN detection loss, For Gram loss, Pick .

[0041] Second aspect: An electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method provided in the first aspect.

[0042] Third aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0043] The beneficial effects of this invention are:

[0044] 1. This invention achieves high-precision automatic identification and continuous optimization of sparse small targets in ultra-large remote sensing images by integrating the DINOv3 self-supervised vision basic model, the improved ViT-Adapter, and the Cascade-RCNN multi-stage detection. It is especially suitable for low-altitude remote sensing image processing scenarios with high resolution (centimeter to meter level), large file size (single image ≥ 10GB), complex ground features, and sparse targets. It can be applied to fields such as power line inspection, smart city construction, and engineering supervision.

[0045] 2. This invention effectively solves the dilemma of difficult and costly annotation of high-resolution, ultra-large file remote sensing images by constructing an integrated data processing and interactive correction subsystem. The system adopts a spatial indexing and sliding window inference mechanism, which can efficiently process single images of 10GB level in blocks, and control the inference time of the whole map to the hour level. It breaks through the limitation of traditional tools that are difficult to handle large files. By publishing images and predicted vector results through GeoServer, it provides efficient editing tools on the web, freeing manual labor from the heavy work of full map annotation. Manual labor only needs to correct the automatic inference results, which improves the overall annotation efficiency by 3-5 times compared with traditional manual annotation tools, reduces about 80% of the invalid workload, and seamlessly connects the originally fragmented data processing and annotation links to form an efficient closed loop.

[0046] 3. This invention innovatively designs a core algorithm architecture of self-supervised backbone + adapter + Gram feedback, effectively bridging the gap between general-purpose basic models and specialized remote sensing tasks. It uses a frozen DINOv3 model, self-supervised and trained with hundreds of millions of multimodal data points, as the backbone, providing powerful and stable global feature representation capabilities. Simultaneously, by injecting multi-scale spatial priors through the ViT-Adapter module, it reconstructs pyramid features suitable for detection, overcoming the deficiency of pure ViT models in lacking multi-scale features. Furthermore, it introduces Gram feature quality assessment and dynamic threshold control, adaptively adjusting the detector's sensitivity based on feature clarity, ensuring the ability to capture small and sparse targets. Applications demonstrate that the method of this invention exhibits excellent generalization performance.

[0047] 4. This invention organically integrates multiple previously fragmented processes, such as data processing, model inference, manual correction, and iterative training, into an automated closed-loop system. This solves the problem of fragmented processes, achieving full automation from "automatic inference to generate pseudo-labels" to "interactive manual correction" and then to "automatic sample generation and incremental training." High-quality data after manual correction is automatically converted into training samples for incremental model training. This closed loop allows the model to continuously learn from newly labeled data. With each iteration, the mAP for small target detection can be improved by an average of 3-5%, forming a self-reinforcing cycle of "data accumulation - model optimization - performance improvement." Simultaneously, the system's input and output both use standard GIS formats (such as GeoJSON / Shapefile), ensuring good compatibility with existing business systems (such as power line inspection and smart city platforms) without complex conversions. This shortens the technology deployment and application cycle by 50%, greatly promoting the practical implementation and long-term maintenance of the technology. Attached Figure Description

[0048] Figure 1 This is a schematic flowchart of the method of the present invention;

[0049] Figure 2 This is a flowchart illustrating the principle of the method of the present invention;

[0050] Figure 3 This is a flowchart illustrating the principle of the data processing and reasoning subsystem of this invention.

[0051] Figure 4 This is a flowchart illustrating the principle of the interactive calibration and sample generation subsystem of this invention.

[0052] Figure 5 This is a flowchart illustrating the principle of the core algorithm detection subsystem of this invention.

[0053] Figure 6 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0054] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0055] Current low-altitude remote sensing image interpretation suffers from problems such as difficult annotation, high cost, poor adaptability of general models, and fragmented workflow.

[0056] To address the aforementioned issues, this invention constructs a closed-loop system encompassing large-scale automatic inference, interactive correction, and incremental / refined training. Combining core algorithms of self-supervised feature extraction, multi-scale adaptation, and dynamic feedback detection, it achieves high-precision detection and continuous optimization of small targets in high-resolution, ultra-large file remote sensing images.

[0057] The system upon which the method of this invention is based is divided into three major subsystems: a data processing and reasoning subsystem, an interactive correction and sample generation subsystem, and a core algorithm detection subsystem. These three subsystems work together to form a closed-loop process.

[0058] Example 1:

[0059] This embodiment discloses a method for interpreting and iteratively correcting small targets in low-altitude remote sensing images. This method is designed for low-altitude remote sensing images up to 10GB in size, achieving a complete closed loop from data import to continuous model optimization. Figure 1 and Figure 2 As shown, the steps include:

[0060] S1. Using the pre-trained core algorithm model, a sliding window strategy is adopted to perform block inference prediction on ultra-high resolution low-altitude remote sensing images, and the inference prediction results are vectorized.

[0061] First, preliminary data annotation and small-scale model pre-training are performed. Then, large-scale automatic inference is carried out, employing a sliding window and parallel scheduling strategy to achieve seamless block-based inference of large images. Boundary overlap is ensured to avoid target truncation, while inference efficiency is maintained through GPU / multi-node parallelism. The front end provides high-resolution images and prediction results vector overlay, and targets can be located one by one through scaling and translation. During correction, batch addition, deletion, modification, version management, and collaborative annotation are supported to achieve rapid correction and quality control. At the same time, training samples are automatically generated based on samples and incremental training samples. The corrected vector data is automatically sliced ​​according to the specified size and overlap rate to generate training samples. The new samples are included in the training set to perform incremental or fine training on the core algorithm to achieve continuous performance improvement.

[0062] The data processing and inference process includes: first, spatial indexing, which is responsible for reading large-scale remote sensing imagery and establishing the spatial index; by providing a unified coordinate projection and metadata parsing interface, it lays the spatial reference foundation for subsequent tiling, inference, and vectorization; then, sliding window inference is performed, using sliding window and parallel scheduling strategies to achieve seamless block-based inference of large images; boundary overlap is ensured to avoid target truncation, while inference efficiency is maintained through GPU / multi-node parallelism. Finally, annotation vectorization is performed, converting the detection box output into vector standards such as GeoJSON / Shapefile, recording confidence and category attributes, and establishing a spatial index to support fast querying, spatial statistics, and subsequent editing.

[0063] S2. Utilize GeoServer or GIS services to publish the predicted image of the image tile joint vectorization prediction results on the application front end.

[0064] Based on the interactive calibration and sample generation subsystem, GIS service data is published. Pyramid-style image tiles are published via GeoServer, enabling browser-side zooming and panning without delay. Vector prediction results are published as WFS / WMS services, which can be overlaid with base maps for display.

[0065] The server-side utilizes GeoServer or similar GIS services to jointly publish ultra-high resolution image tiles (tile pyramids) and predicted vector results, enabling online browsing, vector overlay, and spatial querying of 10GB-level images. The system automatically filters low-quality prediction results based on rules such as confidence level and IoU with pseudo-true values, providing an efficient candidate list for manual correction.

[0066] S3. Manually add, delete, or modify low-confidence or suspected erroneous targets in the predicted image and write the corrected predicted image results into the vector database.

[0067] First, intelligent filtering is performed, automatically filtering candidate targets based on confidence level, IoU, or business rules to form a task queue for manual correction.

[0068] Then, manual interactive editing and correction are performed. The front end uses vector tile technology, which can handle real-time rendering of millions of elements. The correction supports one-click jump to the predicted target, automatic snapping, intelligent lasso, and batch addition, deletion and modification.

[0069] Finally, sample generation is performed. Based on the corrected vector bounding box, image blocks are automatically cropped. The size, overlap rate and boundary buffer can be set to generate training samples that conform to mainstream detection formats such as COCO / YOLO, which are then directly input into the downstream training pipeline.

[0070] S4. Incorporate the corrected predicted images into the training set and perform incremental or fine-tuning training on the core algorithm model.

[0071] Automatically crop image blocks based on the corrected vector bounding boxes, support boundary buffering, generate standard samples that meet the requirements of downstream training, and directly connect to the model training pipeline.

[0072] The manually corrected vector targets can be automatically sliced ​​according to the set size and overlap rate to generate a training sample set that meets the requirements of the detection task. The system incorporates the new samples into the training data to perform incremental or fine training on the core algorithm model, realizing a continuous closed loop of "automatic reasoning - manual correction - retraining" and improving the model's robustness and generalization ability to sparse small targets.

[0073] The method in this embodiment can achieve large-scale automatic reasoning and result vectorization. It adopts a sliding window reasoning mechanism and uses spatial indexing and sliding window strategy to perform block reasoning for low-altitude remote sensing images that can reach 10GB per image, achieving seamless full map coverage while keeping computing resources controllable.

[0074] By vectorizing the prediction results, the candidate boxes and their confidence scores output by the model are automatically converted into standard remote sensing vector formats (such as GeoJSON or Shapefile) and a spatial index is created to facilitate subsequent retrieval, filtering and publishing.

[0075] Through vector data publishing and intelligent filtering, the server utilizes GeoServer or similar GIS services to jointly publish ultra-high resolution image tiles (tile pyramids) and predicted vector results, enabling online browsing, vector overlay, and spatial querying of 10GB-level images. The system automatically filters low-quality prediction results based on rules such as confidence level and IoU with pseudo-true values, providing an efficient candidate list for manual correction.

[0076] Interactive manual correction and version management are implemented. The front end loads predicted vectors in real time on a high-resolution base map, supporting rapid positioning, zooming, and translation. Low-confidence or suspected erroneous targets are prioritized based on strategies such as confidence level and spatial overlap. Manual intervention allows for the direct addition, deletion, and modification of all correction results, which are written to a versioned vector database in real time, forming traceable data assets.

[0077] Simultaneously, automatic sample generation and iterative training are performed. The manually corrected vector targets can be automatically sliced ​​according to the set size and overlap rate to generate a training sample set that meets the requirements of the detection task.

[0078] The system incorporates new samples into the training data and performs incremental or fine-tuning training on the core algorithm model to achieve a continuous closed loop of "automatic reasoning-human correction-retraining" and improve the model's robustness and generalization ability to sparse small targets.

[0079] Example 2:

[0080] This embodiment discloses a low-altitude remote sensing image small target interpretation and iterative correction system that applies the method described in Embodiment 1, including a data processing and inference subsystem, an interactive correction and sample generation subsystem, and a core algorithm detection subsystem.

[0081] The data processing and inference subsystem is used to acquire ultra-high resolution low-altitude remote sensing images. Based on the core algorithm model, it performs block-based inference and prediction using a sliding window strategy, and then vectorizes the inference and prediction results. Figure 3 As shown, the specific items include:

[0082] The spatial indexing module is used to read large-scale remote sensing images, establish a unified coordinate projection and metadata parsing interface, provide a spatial reference basis for tiling and inference, and avoid target positioning deviations caused by inconsistent coordinates.

[0083] The sliding window inference module is used to divide the image into 1024×1024 window blocks using a parallel scheduling strategy, call the core algorithm model for inference, ensure the integrity of the target through window overlap, and use GPU clusters to improve inference speed (the inference time for a single 10GB image is controlled within hours).

[0084] The annotation vectorization module is used to convert the detection boxes (including confidence and category) output by inference into GeoJSON / Shapefile format, and to build a spatial index to support fast query (such as filtering targets by "region range") and subsequent editing.

[0085] The interactive calibration and sample generation subsystem provides GeoServer or GIS services, publishes the combined predicted imagery, generates a list of manual calibration tasks, allows manual calibration of targets in the predicted imagery, writes the calibration results into a vector database, and generates samples based on the calibration, such as... Figure 4 As shown, where:

[0086] The GIS service publishing module publishes pyramid-shaped image tiles through GeoServer, and publishes vector prediction results in the form of WFS / WMS services, realizing the overlay display of image base map and prediction results vector overlay.

[0087] The intelligent filtering module automatically filters candidate targets based on business rules and generates a list of manual correction tasks, with priority from high to low as "missed targets > low confidence targets > suspected false detection targets".

[0088] The manual interactive editing module uses vector tile technology on the front end, supports real-time rendering of millions of features, and provides functions such as one-click target jump, batch annotation templates, and historical version comparison, improving annotation efficiency (annotation speed is 3-5 times faster than traditional tools).

[0089] The sample generation module automatically cropes image blocks based on the corrected vector boxes, supports boundary buffering, and generates standard samples that meet the requirements of downstream training, directly connecting to the model training pipeline.

[0090] The core algorithm detection subsystem is used to train and provide the core algorithm model.

[0091] like Figure 5 As shown, the core algorithm detection subsystem of this embodiment adopts a multi-scale cascaded remote sensing target detection structure based on self-supervised DINOv3-Adapter and Gram feedback, including:

[0092] The self-supervised DINOv3 backbone module is used for efficient image feature mapping, rotation position encoding (RoPE) spatial prior, deep global semantic modeling, and freezing feature weight benchmarks.

[0093] The images are transformed into 1024-dimensional feature tokens after 16×16 patch embedding. A multi-layer self-attention (Transformer) structure is adopted to extract global context across regions without manual annotation.

[0094] Rotary Position Embedding (RoPE) is employed to directly incorporate two-dimensional relative position information during attention matrix calculation. This mechanism preserves the geometric relationships between targets and is invariant to changes in resolution and viewpoint, making it particularly suitable for the geographically continuous distribution of small targets such as power towers and cranes.

[0095] Furthermore, the frozen weights are maintained on the generalized representations obtained from self-supervised pre-training on hundreds of millions of multimodal remote sensing images, serving as a stable feature benchmark. Additionally, global semantic vectors, local patch tokens, and storage tokens provide multi-scale, multi-semantic layer information sources for downstream adaptation.

[0096] Using the DINOv3 visual Transformer structure, which has been self-supervised and trained on hundreds of millions of multimodal remote sensing and natural images, as the backbone, it provides high-quality global semantic and spatial features without the need for large-scale manual annotation, achieving:

[0097] Efficient image-feature mapping achieves the mapping from pixels to high-dimensional tokens through image slicing and linear projection, preserving fine-grained texture information and providing dense input for subsequent global attention.

[0098] Rotation Position Encoding (RoPE) spatial prior introduces relative position information of row and column directions into attention calculation, avoiding scale distortion of traditional absolute position encoding and ensuring the spatial geometric relationship of small targets.

[0099] Deep global semantic modeling utilizes a multi-layer self-attention structure to extract cross-scale and cross-regional contextual semantics, naturally adapting to the geographical continuity of sparse small targets.

[0100] The frozen feature benchmark keeps the DINOv3 pre-trained weights unchanged in downstream tasks, serving only as a stable feature benchmark for subsequent modules to use, thus avoiding overfitting and reducing computational costs.

[0101] The DINOv3-Adapter multi-scale adaptation module is used for collaborative utilization of multiple types of features, ensuring the fidelity of relative position information, lightweight multi-scale injection, and achieving task-driven joint optimization.

[0102] By sending the three types of features output by DINOv3 into the lightweight adaptation channel for multi-path feature fusion, it is beneficial to maintain overall scene perception, enhance edge and texture details, and provide cross-layer memory to supplement the context.

[0103] During spatial prior injection, the RoPE positional relationships are preserved in the adapter, and a cross-attention mechanism is used to achieve fine alignment between local geometric features and global representations. This design ensures that the relative positions of small targets remain consistent when the resolution is scaled.

[0104] Furthermore, the adapter projects the fused features into multi-scale pyramid feature maps of 1 / 32, 1 / 16, and 1 / 8 scales, providing progressively higher resolution inputs for subsequent Cascade-RCNN.

[0105] Furthermore, by adopting a paradigm of frozen backbone + fine-tuned adapter + end-to-end training of the detection head, the detection loss directly drives the adapter to adjust the multi-scale feature distribution, ensuring that it can still learn local representations sensitive to small targets in small sample or pseudo-labeled scenarios.

[0106] To overcome the lack of multi-scale features in ViT and improve the performance of self-supervised visual models in low-altitude remote sensing small target detection, this invention proposes to organically combine DINOv3 with ViT-Adapter and build a task-driven multi-scale adaptation mechanism on this basis. The basic principle is to retain the high-quality global and dense representations provided by DINOv3, while introducing lightweight image spatial priors and multi-scale fusion modules between the backbone and the detection head, enabling global semantics and local details to work together, thereby significantly improving the sensitivity and localization accuracy for sparse small targets.

[0107] By leveraging multiple feature types, the global semantic vector, local patch features, and stored tokens output by DINOv3 are integrated, and information at different semantic scales is sent to different pathways of the adapter to form a three-in-one feature representation of global, local, and contextual information.

[0108] Relative position information is preserved. The rotational position encoding of DINOv3 is retained in the adapter and multi-scale fusion process. The consistency of relative position information is ensured during the adapter and multi-scale fusion process, and the position information is avoided from being distorted during scale transformation.

[0109] Lightweight multi-scale injection, through cross-attention and spatial prior injection, fuses local geometric features with the dense representation of DINOv3 to reconstruct pyramid features suitable for the detector.

[0110] Task-driven joint optimization employs a strategy of freezing the self-supervised backbone, fine-tuning the adapter, and end-to-end training of the detection head. This approach enables the learning of local features sensitive to small targets on a small amount of labeled or pseudo-labeled data, while maintaining the stability of the backbone representation.

[0111] The Gram-feedback Cascade-RCNN detection module is used for multi-scale feature reception, Gram feature quality evaluation, dynamic threshold control, loss closure and bidirectional optimization.

[0112] To address the multi-scale characteristics and dynamic detection requirements of small targets in low-altitude remote sensing, a multi-stage detection system with Gram feedback and dynamic threshold control is constructed.

[0113] Multi-scale feature reception feeds the multi-layer features output from the preceding adapter into the Cascade-RCNN three-stage detection head according to resolution. Low-resolution features are used for global coarse screening, medium-resolution features are used for boundary refinement, and high-resolution features are used for fine-grained category confirmation, realizing progressive detection from coarse screening to refinement.

[0114] Gram feature quality assessment calculates the Gram similarity between the adapter output and the initial global features, quantifies the consistency of feature texture and spatial structure, and serves as an indicator of the sharpness of small target features.

[0115] Dynamic threshold control: The IoU and confidence thresholds for each detection stage are adjusted in real time based on the Gram quality score. The clearer the features, the stricter the threshold. When the features are blurry, the threshold is appropriately relaxed and the adapter is used to enhance the spatial prior.

[0116] The loss loop closure and bidirectional optimization introduce Gram constraint terms into the overall loss and jointly optimize them with the detection loss. The self-supervised backbone is frozen, and only the adapter and detection head are updated, so as to achieve the synchronous improvement of "feature enhancement and detection optimization".

[0117] This module introduces feature quality measurement and dynamic threshold control on the basis of traditional Cascade-RCNN, which enhances its adaptability to sparse small targets.

[0118] Gram feature quality assessment of adapter output features Characteristics of frozen skeletal structure Calculate Gram similarity:

[0119]

[0120] in For Gram matrices, It is the Frobenius norm.

[0121] A higher value indicates that the adapted small target features are more consistent with the original global representation, and the better the quality.

[0122] Dynamic threshold control, based on Real-time adjustment of IoU thresholds for the three phases of Cascade:

[0123]

[0124] When features are clear, the threshold is strictly set to suppress redundancy; when features are ambiguous, the threshold is relaxed and the adapter is triggered to strengthen spatial prior injection, thereby improving the recall rate of small targets.

[0125] Loss loop closure and bidirectional optimization, the total loss function is expressed as:

[0126]

[0127]

[0128] in, For Cascade-RCNN detection loss, Pick In addition, only the adapter and testing station parameters are updated; the self-monitoring backbone remains frozen. Gram loss is used to measure feature quality and is part of the optimization objective to improve the accuracy of small target detection. Through Gram constraints, the stability of the feature space is maintained, achieving a two-way improvement of "feature enhancement - detection optimization".

[0129] Application examples:

[0130] Application of 0.5-1m resolution satellite imagery in low-altitude power line inspection.

[0131] Data acquisition and preprocessing include: Data sources are commercial high-resolution satellites (typically 0.5m or 1m resolution) that periodically acquire multispectral or panchromatic images of the target route area. Low-altitude applications utilize UAVs or ground stations for local re-enhancing and registration, forming a high-resolution remote sensing dataset for "low-altitude mission-oriented" applications. Image size is significant; a single orthophoto mosaic, after stitching, can reach 8-12GB, exhibiting the characteristics of a typical large file for low-altitude inspections.

[0132] For initial model training, a small number of vector-annotated samples (approximately 300 images, resolution 0.5-1m) of existing power towers and their surrounding environment were selected for training. The DINOv3 self-supervised backbone was kept frozen, and only the ViT-Adapter and Cascade-RCNN detection heads were fine-tuned.

[0133] Large-scale automatic inference and vectorization: A sliding window (1024×1024, 20% overlap) strategy is used for parallel inference on a GPU cluster to generate candidate boxes and confidence scores. The results are automatically converted into GeoJSON / Shapefile vectors.

[0134] Interactive manual correction is implemented, publishing pyramid-shaped tiles and vector services via GeoServer. Inspectors can browse 0.5-1m resolution base maps on a web frontend to quickly check for low-confidence or suspected missed targets and make online corrections.

[0135] Incremental training and closed-loop iteration are employed. Based on the correction results, 1024×1024 training blocks (overlap rate 0.25) are automatically pruned to generate standard COCO format samples. Incremental training is then initiated to update the ViT-Adapter and detector head parameters, continuously improving the recall rate and localization accuracy for sparse small targets.

[0136] This application example demonstrates a 3-5 times improvement in annotation efficiency, reducing invalid annotation workload by 80% through automatic filtering of low-quality results and convenient editing tools. Inference time for a single 10GB image is controlled within hours, supporting online viewing and editing without relying on offline processing using professional GIS tools. The average accuracy (mAP) for small targets (<50×50 pixels) such as power towers and tower cranes is improved by 5-10% compared to traditional ViT models, with a 25% reduction in false negatives. In remote sensing images of different resolutions (0.5-1m) and scenes, mAP fluctuation is ≤5%, adapting to multiple business scenarios. After each iteration, the average mAP for small target detection improves by 3-5%, achieving a positive cycle of "data accumulation - model optimization." The system output conforms to GIS annotation standards and can be directly integrated with business systems such as power line inspection and smart cities, improving compatibility and shortening the deployment cycle by 50%.

[0137] The present invention also provides an electronic device, Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 6 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method:

[0138] S1. Using the pre-trained core algorithm model, a sliding window strategy is adopted to perform block inference prediction on ultra-high resolution low-altitude remote sensing images, and the inference prediction results are vectorized.

[0139] S2. Utilize GeoServer or GIS services to publish the predicted image of the image tile joint vectorization prediction results on the application front end.

[0140] S3. Manually add, delete, or modify low-confidence or suspected erroneous targets in the predicted image and write the correction results of the predicted image into the vector database.

[0141] S4. Incorporate the corrected predicted images into the training set and perform incremental or fine-tuning training on the core algorithm model.

[0142] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0144] S1. Using the pre-trained core algorithm model, a sliding window strategy is adopted to perform block inference prediction on ultra-high resolution low-altitude remote sensing images, and the inference prediction results are vectorized.

[0145] S2. Utilize GeoServer or GIS services to publish the predicted image of the image tile joint vectorization prediction results on the application front end.

[0146] S3. Manually add, delete, or modify low-confidence or suspected erroneous targets in the predicted image and write the correction results of the predicted image into the vector database.

[0147] S4. Incorporate the corrected predicted images into the training set and perform incremental or fine-tuning training on the core algorithm model.

[0148] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0149] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A low-altitude remote sensing image small target interpretation and iterative correction method, characterized in that, include: S1. Using the pre-trained core algorithm model, a sliding window strategy is adopted to perform block inference prediction on ultra-high resolution low-altitude remote sensing images, and the inference prediction results are vectorized. S2. Utilize GeoServer or GIS services to publish the predicted image of the image tile joint vectorization prediction results on the application front end. S3. Manually add, delete, or modify low-confidence or suspected erroneous targets in the predicted image and write the correction results of the predicted image into the vector database. S4. Incorporate the corrected predicted images into the training set and perform incremental or fine-tuning training on the core algorithm model. The low-altitude remote sensing image small target interpretation and iterative correction system applying the method described above includes: The data processing and inference subsystem is used to acquire ultra-high resolution low-altitude remote sensing images. Based on the core algorithm model, it performs block-based inference and prediction using a sliding window strategy, and then vectorizes the inference and prediction results. The interactive calibration and sample generation subsystem is used to provide GeoServer or GIS services, publish joint prediction images, generate a list of manual calibration tasks, manually calibrate the targets in the prediction images, write the calibration results into the vector database, and generate samples based on the calibration. The core algorithm detection subsystem is used to train and provide the core algorithm model; The core algorithm detection subsystem, with the self-supervised trained DINOv3 visual Transformer structure as its backbone, adopts a multi-scale cascaded remote sensing target detection architecture based on DINOv3-Adapter multi-scale adaptation and Gram feedback Cascade-RCNN detection components, including: The self-supervised DINOv3 backbone module is used for efficient image feature mapping, rotation position encoding (RoPE) spatial prior, deep global semantic modeling, and freezing feature weight benchmarks. The DINOv3-Adapter multi-scale adaptation module is used for collaborative utilization of multiple types of features, ensuring the fidelity of relative position information, lightweight multi-scale injection, and realizing task-driven joint optimization. The Gram feedback Cascade-RCNN detection module is used for multi-scale feature reception, Gram feature quality evaluation, dynamic threshold control, loss closure and bidirectional optimization. The Gram feature quality assessment formula is expressed as follows: wherein, is a quality assessment score, is an adapter output feature, is a frozen backbone feature, is a Gram matrix, is a Frobenius norm.

2. The method according to claim 1, characterized in that, The data processing and reasoning subsystem includes: The spatial indexing module is used to establish a unified coordinate projection and metadata parsing for ultra-high resolution low-altitude remote sensing images, and to provide a spatial index. The sliding window inference module is used to divide the image into blocks by window using a parallel scheduling strategy; The pseudo-annotation vectorization module is used to convert the detection boxes output by inference into GeoJSON / Shapefile format.

3. The method according to claim 1, characterized in that, The interactive correction and sample generation subsystem includes: The GIS service publishing module is used to publish pyramid-shaped image tiles, enabling the overlay display of image base maps and vector prediction results. The intelligent filtering module is used to automatically filter candidate targets based on business rules and generate a list of manual correction tasks; The manual interactive editing module uses vector tile technology and provides one-click jump to manually corrected targets, batch annotation templates, and historical version comparison functions; The sample generation module automatically cropps the image based on the corrected vector box to generate samples that meet the training requirements.

4. The method according to claim 1, characterized in that, Dynamic threshold control is performed based on... The score is used to adjust the IoU threshold of the three stages of Cascade in real time, as expressed by the formula: in, The IoU threshold is for the three stages.

5. The method according to claim 1, characterized in that, The total loss function for achieving loss loop closure and bidirectional optimization is expressed as follows: in, For Cascade-RCNN detection loss, For Gram loss, Pick .

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images