Low-altitude remote sensing image small target interpretation and iterative correction method, equipment and medium

Abstract

The invention discloses a low-altitude remote sensing image small target interpretation and iterative correction method and device and a medium, and the method comprises the steps: employing a pre-trained core algorithm model, employing a sliding window strategy to carry out the block reasoning prediction of an ultra-large-resolution low-altitude remote sensing image, and carrying out the vectorization of a result; with the help of GeoServer or GIS service, a prediction image combining image slices and vectorization prediction results is issued at an application front end; manually correcting a low-confidence or suspected error target in the predicted image, and writing a correction result into a vector database; and bringing the corrected prediction image into a training set, and carrying out incremental or fine training on the core algorithm model. According to the method, through fusing DINOv3 self-supervised visual basic model, improved ViT-Adapter and Cascade-RCNN multi-stage detection, high-precision automatic identification and continuous optimization of sparse small targets in an ultra-large remote sensing image are realized, and the method is especially suitable for low-altitude remote sensing image processing scenes with high resolution, large file volume, complex ground features and sparse targets.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent interpretation and target detection of remote sensing images, and in particular to a low-altitude remote sensing image small target interpretation and iterative correction method, device and medium. BACKGROUND

[0002] Low-altitude remote sensing technology (such as unmanned aerial vehicle remote sensing) has been widely used in ground object monitoring and target identification scenarios due to its high flexibility, short data acquisition period, and high resolution. However, there are still three major technical bottlenecks in the interpretation of small targets (pixel size usually < 50x50) in current low-altitude remote sensing images, and the three are intertwined, making it difficult for existing solutions to form an effective closed loop.

[0003] First, the data dilemma, the contradiction between massive data and fine annotation: reflected in the limitations of large image processing and annotation tools, the volume of a single low-altitude remote sensing image often exceeds 10GB, and it has the characteristics of "high resolution but local details are fuzzy, targets are sparse and small in size". Traditional general image annotation software cannot effectively load and process, and needs to rely on professional GIS software such as ArcGIS and QGIS, which has a high operation threshold and a complex process. The conflict between geographical continuity and slice annotation, sparse small targets such as power towers and tower cranes have the characteristics of "geographical continuity and structural coherence", and slice operation will destroy the integrity of the target and its relevance to the surrounding environment, making it difficult for annotators to use context to assist in identification, and the risk of missing annotation is significantly increased; while whole-image annotation requires higher tools and manpower. The cost of annotation is high, and without tool assistance, professionals need to fine-tune each point of the small and sparse target, which is time-consuming and labor-intensive, with very low annotation efficiency. Based on the above challenges in data dimension, it is difficult to establish an efficient annotation and processing flow relying only on traditional tools.

[0004] Second, model bottleneck, gap between general models and professional tasks, the current mainstream target detection technology has obvious limitations in low-altitude remote sensing small target interpretation: reflected in the convolutional neural network (CNN), which relies on local receptive field and inductive bias, and performs stably in traditional computer vision tasks, but when faced with complex structure and large target size remote sensing scenes, the detection accuracy and generalization ability are limited due to the limited receptive field. The Transformer model, represented by the visual Transformer (ViT) model of DINOv3, has strong general visual representation and remote sensing ground object semantic understanding ability through massive multi-modal data self-supervised learning (SSL). But it is restricted by the inherent defects of ViT architecture: lack of image-related inductive bias, high computational complexity of self-attention mechanism, and no multi-scale hierarchical structure of feature maps, which leads to poor performance (high miss rate, not sensitive to small targets) in small target detection and other dense prediction tasks, making it difficult to be directly applied to high-precision professional interpretation scenarios.

[0005] Third, the technical solution is fragmented and lacks systematic integration: it is reflected in the fact that existing technologies focus on single algorithm optimization, ignoring the most time-consuming core link of "data acquisition and labeling"; at the same time, data processing, pseudo-labeling, manual correction, and model iteration are scattered in different platforms / tools, and a complete closed loop of "data processing→model training→labeling iteration→performance improvement" has not been formed. The non-closed loop workflow leads to the fragmentation of data and algorithms, poor compatibility and low efficiency when the technology is landed.

[0006] For example, the invention application with application number 202510082463.5 discloses a real-time multi-source remote sensing image small target detection method assisted by super-resolution inference. In the detection process, the medium and large size anchor frame is removed, and a special anchor frame for small target detection is added, which improves the detection accuracy and reduces the inference time of the model. However, there are still problems such as lack of systematic integration, inefficient mass data processing, lack of systematic integration and utilization of data, and difficulty in ensuring stable detection performance when facing complex and variable low-altitude remote sensing image scenes. SUMMARY

[0007] In view of the above problems, the purpose of the present application is to provide a low-altitude remote sensing image small target interpretation and iterative correction method, device and medium, to solve the contradiction between efficient processing and fine labeling of high-resolution, super-large file remote sensing images, the insufficient precision and weak generalization ability of general Transformer / CNN models in small target detection, and the dispersion and fragmentation of data processing, algorithm training, and manual labeling, which makes it difficult to form a collaborative optimization closed loop.

[0008] The embodiment of the present application provides a low-altitude remote sensing image small target interpretation and iterative correction method, device and medium.

[0009] The first aspect is a low-altitude remote sensing image small target interpretation and iterative correction method, comprising:

[0010] S1, using a pre-trained core algorithm model, performing block inference prediction on a super-resolution low-altitude remote sensing image using a sliding window strategy, and performing vectorization processing on the inference prediction result;

[0011] S2, using GeoServer or GIS service, publishing the predicted image of the image slice joint vectorization prediction result in the application front end;

[0012] S3, manually adding, deleting or correcting low confidence or suspected error targets in the predicted image, and writing the predicted image correction result into a vector database;

[0013] S4, incorporating the corrected predicted image into the training set and performing incremental or fine training on the core algorithm model.

[0014] In an embodiment of the present application, a low-altitude remote sensing image small target interpretation and iterative correction system applying the method comprises:

[0015] A data processing and inference subsystem is configured to acquire a super-resolution low-altitude remote sensing image, perform block inference prediction based on a core algorithm model according to a sliding window strategy, and perform vectorization processing on the inference prediction result.

[0016] An interactive correction and sample generation subsystem is configured to provide a GeoServer or GIS service, publish a joint prediction image, generate a manual correction task list, manually correct a target of the prediction image, write a correction result into a vector database, and generate a sample based on the correction.

[0017] A core algorithm detection subsystem is configured to train and provide a core algorithm model.

[0018] In an embodiment of the present application, the data processing and inference subsystem comprises:

[0019] A spatial index module is configured to establish unified coordinate projection and metadata analysis for the super-resolution low-altitude remote sensing image, and provide spatial index.

[0020] A sliding window inference module is configured to perform window block on the image by using a parallel scheduling strategy.

[0021] A pseudo-labeled vectorization module is configured to convert a detection frame output by inference into a GeoJSON / Shapefile format.

[0022] In an embodiment of the present application, the interactive correction and sample generation subsystem comprises:

[0023] A GIS service publishing module is configured to publish pyramid image tiles, and realize superposition display of an image base map and a vector prediction result.

[0024] An intelligent screening module is configured to automatically screen a candidate target based on a business rule, and generate a manual correction task list.

[0025] A manual interactive editing module is configured to provide a manual correction target one-key jump, a batch labeling template, and a historical version comparison function by using a vector tile technology.

[0026] A sample generation module is configured to automatically crop an image according to a corrected vector frame, and generate a sample meeting a training requirement.

[0027] In an embodiment of the present application, the core algorithm detection subsystem takes the DINOv3 visual Transformer structure trained in a self-supervised manner as the backbone, adopts a multi-scale cascaded remote sensing target detection architecture based on the 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 frozen feature weight benchmark.

[0029] The DINOv3-Adapter multi-scale adaptation module is used for multi-type feature collaborative utilization, relative position information fidelity, lightweight multi-scale injection, and task-driven joint optimization.

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

[0031] In an embodiment of the present application, the Gram feature quality evaluation is represented by the following formula:

[0032]

[0033] Wherein, is the quality evaluation score, is the adapter output feature, is the frozen backbone feature, is the Gram matrix, is the Frobenius norm.

[0034] In an embodiment of the present application, the dynamic threshold control is based on The score is used to adjust the IoU threshold of the Cascade three-stage in real time, and the formula is represented as:

[0035]

[0036] Wherein, is the IoU threshold of the three-stage.

[0037] In an embodiment of the present application, the total loss function for realizing loss closed loop and bidirectional optimization is represented by the following formula:

[0038]

[0039]

[0040] Wherein, For Cascade-RCNN detection loss, For Gram loss, Take .

[0041] The second aspect: an electronic device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to realize the steps of the method provided in the first aspect.

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

[0043] The beneficial effects of the present application are:

[0044] 1. The present application realizes high-precision automatic identification and continuous optimization of sparse small targets in ultra-large remote sensing images by fusing DINOv3 self-supervised visual base model, improved ViT-Adapter and Cascade-RCNN multi-stage detection, especially suitable for high-resolution (centimeter to meter), large file size (single image ≥10GB), complex ground objects and sparse target low-altitude remote sensing image processing scene, which can be applied to power inspection, smart city construction, engineering supervision and other fields.

[0045] 2. The present application effectively solves the difficulty of high-resolution ultra-large file remote sensing image annotation and high cost by constructing an integrated data processing and interactive correction subsystem. The system uses spatial indexing and sliding window reasoning mechanism, can efficiently block process single 10GB level image, control the whole image reasoning time in hours, break through the limitation of traditional tools difficult to process large files, publish image and prediction vector results through GeoServer, provide Web efficient editing tool, free human from heavy full image annotation, only need to correct automatic reasoning result, make the overall annotation efficiency compared with traditional manual annotation tool increased by 3-5 times, reduce about 80% of invalid workload, seamlessly link the originally fragmented data processing and annotation link, form a high-efficiency closed loop.

[0046] 3. The application innovatively designs a core algorithm architecture of self-supervised backbone + adapter + Gram feedback, effectively bridging the gap between general basic models and professional remote sensing tasks. A frozen DINOv3 model trained by self-supervised training of hundreds of millions of multi-modal data is used as the backbone to provide strong and stable global feature representation capability. At the same time, a ViT-Adapter module is used to inject multi-scale spatial prior, reconstructing pyramid features suitable for detection, overcoming the lack of multi-scale features in pure ViT models. Gram feature quality evaluation and dynamic threshold control are introduced to adaptively adjust the sensitivity of the detector according to the feature clarity, ensuring the capture ability of small and sparse targets. Application shows that the method of the application can exhibit excellent generalization performance.

[0047] 4. The application integrates multiple originally dispersed links such as data processing, model inference, artificial correction and iterative training into an automatic closed-loop system, solves the problem of process fragmentation, and realizes the full-process automation from "automatic inference to generate pseudo-labeling" to "manual interactive correction" to "automatic sample generation and incremental training". High-quality data corrected by artificial correction will be automatically converted into training samples for incremental training of the model. This closed loop enables the model to continuously learn from newly labeled data, and the average mAP of small target detection can be improved by 3-5% after each iteration, forming a self-reinforcing cycle of "data accumulation-model optimization-performance improvement". At the same time, the system input and output adopts standard GIS format (such as GeoJSON / Shapefile), which has good compatibility with existing business systems (such as power inspection and smart city platform), and does not need complex conversion, shortening the technology deployment application period by 50%, greatly promoting the actual landing and long-term maintenance of technology. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is a flowchart of the method of the application;

[0049] Figure 2 is a principle flowchart of the method of the application;

[0050] Figure 3 is a principle flowchart of the data processing and inference subsystem of the application;

[0051] Figure 4 is a principle flowchart of the interactive correction and sample generation subsystem of the application;

[0052] Figure 5 is a principle flowchart of the core algorithm detection subsystem of the application;

[0053] Figure 6 is a structural schematic diagram of the electronic device of the application. DETAILED DESCRIPTION

[0054] Embodiments of the present application are described below in detail, examples of which are shown in the drawings, wherein the same or similar symbols represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are only used to explain the present application and cannot be understood as a limitation of the present application.

[0055] The existing low-altitude remote sensing image interpretation has problems of difficult annotation, high cost, poor adaptability of general models, and fragmented process,

[0056] To solve the above problems, the present application constructs a full-chain closed-loop system of automatic reasoning, interactive correction, and incremental / fine training in a large range, and realizes high-precision detection and continuous optimization of small targets in high-resolution remote sensing images of large files by combining the core algorithms of self-supervised feature extraction, multi-scale adaptation, and dynamic feedback detection.

[0057] The system on which the method of the present application relies is divided into three subsystems: a data processing and reasoning subsystem, an interactive correction and sample generation subsystem, and a core algorithm detection subsystem, which form a closed-loop process in cooperation.

[0058] Embodiment 1

[0059] The present embodiment discloses a low-altitude remote sensing image small target interpretation and iterative correction method. The method of the present embodiment is aimed at a single low-altitude remote sensing image of up to 10 GB, and realizes a complete closed loop from data import to model continuous optimization. As shown in Figure 1 and Figure 2 The method comprises the following steps:

[0060] S1, using a pre-trained core algorithm model, the super-resolution low-altitude remote sensing image is divided and reasoned by a sliding window strategy, and the reasoning prediction result is vectorized.

[0061] First, the data is preliminarily annotated and the model is small-scale pre-trained, then large-scale automatic reasoning is performed, the sliding window and parallel scheduling strategy are adopted to realize seamless block reasoning of large images; the boundary overlap is ensured to avoid target truncation, and the reasoning efficiency is maintained through GPU / multi-node parallelism, the high-resolution image and the prediction result vector are superimposed in the front end, and the target can be located through scaling and translation; During correction, batch addition, deletion, modification, version management and collaborative annotation are supported to realize rapid correction and quality control, at the same time, based on sample automatic generation and incremental training sample, the corrected vector data is automatically sliced into training samples according to the specified size and overlap rate, the new sample is included in the training set, and the core algorithm is incrementally or finely trained to realize continuous performance improvement.

[0062] The data processing and inference process includes, firstly, spatial indexing, responsible for reading and spatial indexing of large remote sensing images; by providing a unified coordinate projection and metadata analysis interface, the spatial reference basis is laid for subsequent slicing, inference and vectorization; then sliding window inference is carried out, sliding window and parallel scheduling strategy are adopted, seamless block inference of large images is realized; the boundary overlap is ensured to avoid target truncation, and the inference efficiency is maintained through GPU / multi-node parallel. Finally, the labeled vectorization is carried out, the detection box output is converted into GeoJSON / Shapefile and other vector standards, and the confidence and category attributes are recorded, and the spatial index is established to support fast query, spatial statistics and subsequent editing.

[0063] S2, using GeoServer or GIS service, publishing prediction images of image slice combined vectorization prediction results in application front end.

[0064] Based on the interactive correction and sample generation subsystem, GIS service data is published, pyramid image tiles are published through GeoServer, browser side non-delay scaling and translation are realized, vector prediction results are published in WFS / WMS service mode, and can be displayed with the base map,

[0065] The server side uses GeoServer or similar GIS service to jointly publish super high resolution image slices (tile pyramids) and prediction vector results, realizes online browsing, vector superposition and spatial query of 10GB level images. The system automatically filters low quality prediction results according to the rules of confidence, IoU and pseudo true value, and provides an efficient candidate list for manual correction.

[0066] S3, manually correcting low confidence or suspected error targets in prediction images by adding, deleting or changing, and writing prediction image correction results into vector database.

[0067] Firstly, intelligent screening is carried out, candidate targets are automatically screened according to confidence, IoU or business rules, and a task queue for manual correction is formed.

[0068] Then, manual interactive editing correction is carried out, the front end uses vector tile technology, which can carry out real-time rendering of millions of elements; correction supports one-key jump to prediction target, automatic adsorption, intelligent lasso and batch addition, deletion and modification.

[0069] Finally, sample generation is carried out, image blocks are automatically cropped according to the corrected vector frame, the size, overlap rate and boundary buffer are set, training samples in mainstream detection formats such as COCO / YOLO are generated, and the training pipeline is directly input into the downstream.

[0070] S4, the corrected prediction image is included in the training set, and the core algorithm model is incrementally or finely trained.

[0071] According to the corrected vector box, the image block is automatically cropped, the boundary buffer is supported, the standard sample conforming to the downstream training requirements is generated, and the model training pipeline is directly connected.

[0072] The artificially corrected vector target can be automatically sliced according to the set size and overlap rate, and a training sample set conforming to the detection task requirements is generated. The system incorporates the new sample into the training data, and performs incremental or fine training on the core algorithm model, realizes the continuous closed loop of “automatic reasoning-human correction-retraining”, and improves the robustness and generalization ability of the model to sparse small targets.

[0073] The embodiment method can realize wide-range automatic reasoning and result vectorization. A sliding window reasoning mechanism is adopted, and for a single low-altitude remote sensing image of up to 10 GB, a spatial index and a sliding window strategy are used for block reasoning, realizing seamless coverage of the whole image and keeping the computing resources controllable.

[0074] The prediction result vectorization is used, the candidate box and its confidence output by the model are automatically converted into a standard remote sensing vector format (such as GeoJSON or Shapefile), and a spatial index is established, which is convenient for subsequent retrieval, filtering and publishing.

[0075] Through vector data publishing and intelligent screening, the server side uses GeoServer or similar GIS services to jointly publish the super-high-resolution image slices (tile pyramid) and the prediction vector results, realizes online browsing, vector superposition and spatial query of 10 GB level images. The system automatically filters low-quality prediction results according to the rules of confidence, IoU with pseudo-true value, etc., and provides an efficient candidate list for manual correction.

[0076] Interactive manual correction and version management are carried out. The front end loads the prediction vector in real time on the high-resolution base map, supports fast positioning, zooming and panning. According to the confidence and spatial overlap degree strategy, low confidence or suspected error targets are preferentially presented, manual correction can directly add, delete and change all correction results, and real-time writing to the versioned vector database forms a traceable data asset.

[0077] At the same time, sample automatic generation and iterative training are carried out. The artificially corrected vector target can be automatically sliced according to the set size and overlap rate, and a training sample set conforming to the detection task requirements is generated.

[0078] The system incorporates the new sample into the training data, and performs incremental or fine training on the core algorithm model, realizes the continuous closed loop of “automatic reasoning-human correction-retraining”, and improves the robustness and generalization ability of the model to sparse small targets.

[0079] Embodiment 2:

[0080] The embodiment discloses a low-altitude remote sensing image small target interpretation and iterative correction system applying the method in embodiment 1, which comprises a data processing and reasoning subsystem, an interactive correction and sample generation subsystem and a core algorithm detection subsystem.

[0081] The data processing and reasoning subsystem is used for acquiring a super-resolution low-altitude remote sensing image, performing block reasoning prediction based on a core algorithm model according to a sliding window strategy, and performing vectorization processing on the reasoning prediction result. Figure 3 As shown in the figure, the data processing and reasoning subsystem specifically comprises:

[0082] A spatial index module is used for reading a large remote sensing image, establishing a unified coordinate projection and metadata analysis interface, providing a spatial reference basis for slicing and reasoning, and avoiding target positioning deviation caused by non-uniform coordinates.

[0083] A sliding window reasoning module is used for adopting a parallel scheduling strategy, slicing the image according to a 1024x1024 window, calling the core algorithm model for reasoning, ensuring target integrity through window overlap, and improving reasoning speed by using a GPU cluster (the reasoning time of a single 10GB image is controlled within hours).

[0084] A labeled vectorization module is used for converting the detection frame (including confidence and category) output by reasoning into a GeoJSON / Shapefile format, establishing a spatial index to support fast query (such as filtering targets according to a region range) and subsequent editing.

[0085] An interactive correction and sample generation subsystem is used for providing a GeoServer or GIS service, publishing the combined prediction image, generating a manual correction task list, manually correcting the target of the prediction image, writing the correction result into a vector database, and generating a sample based on the correction. Figure 4 As shown in the figure, the subsystem specifically comprises:

[0086] A GIS service publishing module is used for publishing pyramid image tiles through a GeoServer, publishing the vector prediction result in a WFS / WMS service mode, and realizing image base map and prediction result vector overlay display.

[0087] An intelligent screening module is used for automatically screening candidate targets based on business rules, generating a manual correction task list, and prioritizing the list from high to low as “missed targets > low confidence targets > suspected misdetected targets”.

[0088] An artificial interactive editing module is used for supporting real-time rendering of a million-level element on the front end by using a vector tile technology, providing target one-key jump, batch labeling templates and historical version comparison functions, and improving labeling efficiency (the labeling speed is 3-5 times higher than that of traditional tools.

[0089] A sample generation module automatically crops the image block according to the corrected vector frame, supports boundary buffering, generates standard samples meeting the requirements of downstream training, and directly connects the model training pipeline.

[0090] A core algorithm detection subsystem for training and providing a core algorithm model.

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

[0092] A self-supervised DINOv3 backbone module for efficient image feature mapping, rotary position encoding (RoPE) spatial prior, deep global semantic modeling, and frozen feature weight benchmark.

[0093] After the image is converted into a 1024-dimensional feature token through 16x16 patch embedding, a multi-layer self-attention (Transformer) structure is used to extract the global context across regions without human annotation.

[0094] Rotary relative position encoding (Rotary Position Embedding, RoPE) is used to directly introduce two-dimensional relative position information during attention matrix calculation. This mechanism maintains the geometric relationship between targets and is invariant to changes in resolution and viewing angle, making it particularly suitable for the geographical continuous distribution of small targets such as power towers and tower cranes.

[0095] In addition, the frozen weights maintain the generalization representation obtained from the self-supervised pre-training of the billion-level multi-modal remote sensing image as a stable feature benchmark. Furthermore, the global semantic vector, local patch token, and storage token provide a multi-scale, multi-semantic layer information source for downstream adaptation.

[0096] Taking the DINOv3 visual Transformer structure trained by the billion-level multi-modal remote sensing and natural image self-supervised training as the backbone, high-quality global semantics and spatial features are provided without the need for large-scale manual annotation, achieving:

[0097] Efficient image-feature mapping, which maps 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] Rotary position encoding (RoPE) spatial prior introduces relative position information in the row and column directions during attention calculation, avoiding the scale distortion of traditional absolute position encoding and ensuring the spatial geometric relationship of small targets.

[0099] Deep global semantic modeling, using multi-layer self-attention structure to extract cross-scale and cross-region context semantics, naturally adapts to the geographical continuity of sparse small targets.

[0100] Freeze the feature reference, keep the DINOv3 pre-training weight unchanged in the downstream task, only as a stable feature reference for subsequent module calling, to avoid overfitting and reduce the calculation cost.

[0101] DINOv3-Adapter multi-scale adaptation module is used for multi-type feature collaborative utilization, relative position information fidelity, lightweight multi-scale injection, and task-driven joint optimization.

[0102] The three types of features output by DINOv3 are sent to the lightweight adaptation channel for multi-path feature fusion, which is beneficial to maintain overall scene perception, strengthen edge and texture details, and provide cross-layer memory to supplement context.

[0103] The RoPE position relationship is retained in the adapter when spatial prior injection is performed, and the local geometric features and global representations are precisely aligned through the cross-attention mechanism. This design ensures that the relative position of small targets remains consistent when the resolution is scaled.

[0104] Moreover, the adapter projects the fused features into 1 / 32, 1 / 16, 1 / 8 multi-scale pyramid feature maps, providing resolution-progressive input for subsequent Cascade-RCNN.

[0105] In addition, the paradigm of freezing the backbone + fine-tuning the adapter + end-to-end training the detection head is adopted, and the detection loss directly drives the adapter to adjust the multi-scale feature distribution, ensuring that the local representation sensitive to small targets can still be learned in the small sample or pseudo-labeled scene.

[0106] To overcome the defect of ViT lacking multi-scale features and improve the performance of self-supervised visual base model in low-altitude remote sensing small target detection, the present application proposes to combine DINOv3 with ViT-Adapter organically, and on this basis, a task-driven multi-scale adaptation mechanism is constructed. The basic principle is: to retain the high-quality global and dense representations provided by DINOv3, while introducing lightweight image spatial prior and multi-scale fusion modules between the backbone and the detection head, so that global semantics and local details work together, thereby significantly improving the sensitivity and positioning accuracy of sparse small targets.

[0107] Multi-type feature collaborative utilization integrates the global semantic vector, local patch feature and storage token output by DINOv3, and sends the information of different semantic scales to different paths of the adapter, forming a global-local-context three-in-one feature expression.

[0108] Relative position information fidelity, DINOv3's rotation position encoding is preserved in the adapter and multi-scale fusion process, ensuring the consistency of relative position information during the adapter and multi-scale fusion process, and avoiding distortion of position information during scale transformation.

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

[0110] Task-driven joint optimization, using the strategy of freezing self-supervised backbone + fine-tuning adapter + detection head end-to-end training, local feature learning sensitive to small targets is realized on a small amount of labeled or pseudo-labeled data, while keeping the backbone representation stable.

[0111] Gram feedback Cascade-RCNN detection component module, for multi-scale feature reception, Gram feature quality evaluation, dynamic threshold control, loss closed loop and bidirectional optimization.

[0112] Multi-scale feature and dynamic detection requirements for low-altitude remote sensing small targets, multi-stage detection system with Gram feedback and dynamic threshold control is constructed.

[0113] Multi-scale feature reception, multi-layer features output by the previous adapter are sent to the Cascade-RCNN three-stage detection head according to the 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 class confirmation, realizing progressive detection from coarse screening to refinement.

[0114] Gram feature quality evaluation, calculate the Gram similarity between the adapter output and the initial global feature, quantify the consistency of feature texture and spatial structure, as a small target feature clarity index.

[0115] Dynamic threshold control: adjust the IoU and confidence threshold of each detection stage in real time according to the Gram quality score, the clearer the feature, the stricter the threshold, and when the feature is blurred, the threshold is relaxed and the adapter is triggered to enhance the spatial prior.

[0116] Loss closed loop and bidirectional optimization, introduce Gram constraint term in the overall loss, jointly optimize with detection loss, freeze self-supervised backbone, only update adapter and detection head, realize "feature enhancement-detection optimization" synchronous promotion.

[0117] This module introduces feature quality measurement and dynamic threshold control based on traditional Cascade-RCNN, and strengthens the adaptability to sparse small targets.

[0118] Gram feature quality evaluation on adapter output features and frozen backbone features Compute Gram similarity:

[0119]

[0120] where is the Gram matrix, is the Frobenius norm.

[0121] The higher the value, the more consistent the small target features after adaptation with the original global representation, the better the quality.

[0122] Dynamic threshold control, according to Real-time adjustment of the IoU threshold of the Cascade three-stage:

[0123]

[0124] When the feature is clear, the threshold is strict to suppress redundancy; when the feature is blurred, the threshold is relaxed and the adapter is triggered to enhance spatial prior injection, improving small target recall rate.

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

[0126]

[0127]

[0128] where, is the Cascade-RCNN detection loss, take In addition, only the adapter and detection platform parameters are updated, and the self-supervised backbone is always frozen; Gram loss, used to measure feature quality, as part of the optimization goal to improve small target detection accuracy, through Gram constraint, maintain the stability of the feature space, realize the "feature enhancement-detection optimization" bidirectional promotion.

[0129] Application example:

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

[0131] Data acquisition and preprocessing include: data source is commercial high-resolution satellite (typical resolution 0.5m or 1m) to obtain multi-spectral or panchromatic images of target line area regularly. Low-altitude application, local re-shooting and registration are carried out by using unmanned aerial vehicle or ground station, forming "low-altitude task" high-resolution remote sensing data set. Image volume, after splicing, a single orthographic mosaic map can reach 8-12GB, with the characteristics of super large file of typical low-altitude inspection.

[0132] Initial model training, select a small number of vector labeled samples (about 300, resolution 0.5-1m) of existing power towers and surrounding environment for training. DINOv3 self-supervised backbone remains frozen, only fine-tune ViT-Adapter and Cascade-RCNN detection head.

[0133] Large-scale automatic reasoning and vectorization, use sliding window (1024x1024, overlap 20% pixels) strategy to perform parallel reasoning on GPU cluster, generate candidate boxes and confidence, and automatically convert the results into GeoJSON / Shapefile vector.

[0134] Interactive manual correction, publish pyramid tiles and vector services through GeoServer. The inspection personnel browse the 0.5-1m resolution base map on the Web front end, quickly check the low confidence or suspected missed targets and modify them online.

[0135] Incremental training and closed-loop iteration, according to the correction results, crop 1024x1024 training blocks (overlap rate 0.25) to generate standard COCO format samples. Start incremental training to update ViT-Adapter and detection head parameters, continuously improve the recall rate and positioning accuracy of sparse small targets.

[0136] Through the application example method, the annotation efficiency can be improved by 3-5 times, and through "automatic screening of low-quality results + convenient editing tool", 80% of invalid annotation workload is reduced. The inference time of a single 10GB image is controlled within hours, supporting online browsing and editing without relying on professional GIS tools for offline processing. The average precision (mAP) of small targets (<50x50 pixels) such as power towers and tower cranes is improved by 5-10% compared to traditional ViT models, and the missed detection rate is reduced by 25%; in different resolutions (0.5-1m) and different scenes of remote sensing images, the mAP fluctuation is ≤5%, suitable for multiple business scenarios. After each iteration, the small target detection mAP is improved by an average of 3-5%, realizing the positive cycle of "data accumulation-model optimization"; the system output conforms to the GIS annotation standard, and can be directly connected to power inspection, smart city and other business systems, improving compatibility and shortening deployment cycle by 50%.

[0137] The application also provides an electronic device, Figure 6 The structure schematic diagram of the electronic device provided by the embodiment of the application is shown as Figure 6 As shown in the figure, the electronic device can include a processor, a communications interface, a memory and a communications bus, wherein the processor, the communications interface and the memory complete mutual communication through the communications bus. The processor can call logical instructions in the memory, for example, to execute the following method:

[0138] S1, using a pre-trained core algorithm model, using a sliding window strategy to block inference prediction on a super high-resolution low-altitude remote sensing image, and performing vectorization processing on the inference prediction result;

[0139] S2, using a GeoServer or GIS service, publishing a predicted image of the image slice combined with the vectorization prediction result in an application front end;

[0140] S3, manually adding, deleting or correcting low-confidence or suspected error targets in the predicted image, and writing the predicted image correction result into a vector database;

[0141] S4, incorporating the corrected predicted image into the training set, and performing incremental or fine training on the core algorithm model.

[0142] In addition, the logical instructions in the above-mentioned memory can be implemented in the form of a software functional unit and sold or used as a stand-alone product, and can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0143] The embodiments of the present application also provide a non-transitory computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the method provided by the above-mentioned embodiments, for example, comprising:

[0144] S1, using a pre-trained core algorithm model, using a sliding window strategy to block inference prediction on a super high-resolution low-altitude remote sensing image, and performing vectorization processing on the inference prediction result;

[0145] S2, using a GeoServer or GIS service, publishing a predicted image of the image slice combined with the vectorization prediction result in an application front end;

[0146] S3, manually adding, deleting or correcting low-confidence or suspected error targets in the predicted image, and writing the predicted image correction result into a 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.

2. The method according to claim 1, characterized in that, 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.

3. The method according to claim 2, 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.

4. The method according to claim 2, 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.

5. The method according to claim 2, characterized in that, 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.

6. The method according to claim 5, characterized in that, The Gram feature quality assessment formula is expressed as follows: in, For quality assessment scores, For adapter output characteristics, Characteristics of frozen skeletal structure For Gram matrices, It is the Frobenius norm.

7. The method according to claim 5, 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.

8. The method according to claim 5, 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 .

9. 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 8.

10. 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 8.

Images