Aerial small target detection and segmentation method and system based on reliability guidance
By generating modal reliability maps and cross-modal local migration fields in aerial images, and combining them with a reliability-migration joint gating fusion network, the stability and consistency problems of small target detection in aerial images are solved, and efficient detection and segmentation of sparse small targets are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG AIRCRAFT DESIGN INST AVIATION IND CORP OF CHINA
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies struggle to address dual-modal quality differences, local spatial shifts, and sparse candidate regions in aerial images, resulting in insufficient stability in small target detection and a lack of closed-loop correction in the detection results.
By acquiring RGB and infrared aerial images, performing time synchronization and radiometric normalization, a modal reliability map is generated, the cross-modal local migration field is estimated, feature fusion is performed using a reliability-migration joint gating fusion network, and closed-loop correction is performed for target detection and segmentation.
It improves the stability and consistency of aerial target recognition in complex environments, reduces the missed detections and false detections of sparse small targets, and enhances the stability of target bounding boxes and contours.
Smart Images
Figure CN122435482A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of computer vision and aerial intelligent perception technology, and specifically relates to a reliability-guided aerial small target detection and segmentation method and system. Background Technology
[0002] Target detection and pixel-level contour extraction in aerial images are crucial foundational capabilities for applications such as UAV inspection, urban management, emergency search and rescue, traffic monitoring, and security reconnaissance. Compared to ground-view images, targets in aerial images are generally smaller in scale, sparsely distributed, have complex backgrounds, and span a wide scale range. Furthermore, they are more susceptible to spatial misalignment between different modalities due to changes in flight attitude and sensor installation deviations. In harsh environments, RGB images are easily affected by low light, rain, fog, shadows, and partial occlusion, leading to a loss of texture details. While infrared images can highlight thermal targets, they can also be affected by thermal noise, thermal drift, and boundary blurring, causing distortion of target shapes.
[0003] Existing bimodal methods often employ a registration-then-fusion or direct feature stitching approach, assuming that both modalities have the same level of reliability in all regions. This makes it difficult to adapt to complex aerial photography scenarios where the degree of degradation varies with spatial location. On the other hand, existing small target detection techniques typically focus on the multi-scale structure of the detection network itself, but rarely establish explicit compensation mechanisms for the weak registration problem of bimodal methods. Furthermore, existing segmentation methods are often independent of the detection module, resulting in a lack of feedback correction between the detection bounding box and the segmentation contour. Ultimately, this easily leads to problems such as missed detections, false detections, and unstable boundaries for sparse small targets.
[0004] Therefore, there is an urgent need for a technical solution that can take into account the differences in quality between the two modes, local spatial offset, sparse candidate region screening and detection segmentation closed-loop correction, so as to improve the recognition stability and consistency of sparse small targets in aerial photography in complex environments. Summary of the Invention
[0005] The purpose of this application is to provide a reliability-guided aerial small target detection and segmentation method and system to solve the problems in the prior art, such as the susceptibility of direct fusion of dual-modal images to degradation and weak registration, insufficient stability of sparse small target detection, and lack of closed-loop correction between detection and segmentation results.
[0006] The technical solution of this application is:
[0007] The first aspect of this application provides a reliability-guided aerial small target detection and segmentation method, including:
[0008] Step S1: Acquire RGB aerial images and infrared aerial images of the same scene, and perform time synchronization, radiometric normalization, and coarse registration.
[0009] Step S2: Perform a quality degradation assessment based on the brightness, contrast, noise, fogging, occlusion, and missing features of the RGB aerial image and the infrared aerial image, and generate the corresponding modal reliability map;
[0010] Step S3: Estimate the cross-modal local migration field based on the local correlation between the RGB aerial image and the infrared aerial image, and generate sparse candidate regions by combining the modal reliability map;
[0011] Step S4: Input the RGB aerial image, the infrared aerial image, the modal reliability map, and the cross-modal local offset field into the reliability-offset joint gating fusion network to obtain the fused feature map;
[0012] Step S5: Perform small target detection based on the fused feature map to obtain the target bounding box and target category;
[0013] Step S6: Perform bounding box guided contour segmentation based on the target bounding box to obtain the target contour mask;
[0014] Step S7: Perform closed-loop correction on the target bounding box and target confidence based on the target contour mask, and output the target category, target location and target contour information.
[0015] In at least one embodiment of this application, step S1 includes: performing timestamp matching on RGB aerial images and infrared aerial images; normalizing the size and radiometric response of images of different modalities; and using a coarse registration method based on global feature points and platform attitude information to ensure that the dual-modal images meet the initial alignment conditions for subsequent local offset estimation.
[0016] In at least one embodiment of this application, step S2 includes: extracting brightness stability, texture clarity, and occlusion ratio features from RGB aerial images; extracting thermal saliency, thermal noise intensity, and edge continuity features from infrared aerial images; performing region-level quality scoring on the dual-modal images based on the features; and generating pixel-level or block-level modal reliability maps; wherein the modal reliability map is used to characterize the reliability of the corresponding region in participating in fusion calculation under the current environment.
[0017] In at least one embodiment of this application, step S3 includes: extracting local corresponding blocks on the multi-scale feature maps of the RGB aerial image and the infrared aerial image; estimating the offset vectors of each candidate region based on cross-modal correlation and geometric constraints; combining the offset vectors into a cross-modal local offset field; and screening sparse candidate regions based on thermal saliency, structural complexity, and offset consistency.
[0018] In at least one embodiment of this application, in step S4, the reliability-offset joint gating fusion network includes a reliability branch, an offset compensation branch, and a joint gating branch; the reliability branch outputs bimodal region validity weights based on the modal reliability map; the offset compensation branch performs flexible alignment of region features based on the cross-modal local offset field; the joint gating branch combines the bimodal region validity weights, offset consistency, and sparse priors to generate region-level fusion weights to control the fusion strength and feature flow direction of RGB features and infrared features;
[0019] The joint gating branch adopts a progressive fusion strategy, which preserves target edge and thermally significant difference information in the shallow feature stage, and enhances semantic consistency and class separability in the deep feature stage.
[0020] In at least one embodiment of this application, step S5 includes: constructing a small target detection network based on the fused feature map; the small target detection network includes at least a candidate screening detection head and a refinement detection head, the candidate screening detection head is used to quickly screen out low-confidence background regions, and the refinement detection head is used to refine the bounding box position, scale and category of the candidate targets.
[0021] In at least one embodiment of this application, step S6 includes: generating a segmentation seed region based on the target bounding box; feeding the segmentation seed region and the corresponding fusion feature into a contour segmentation network to obtain a target contour mask; the contour segmentation network refines the target region at different scales and suppresses background leakage based on edge continuity constraints.
[0022] In at least one embodiment of this application, step S7 includes: correcting the position and size of the target bounding box based on the outer region of the target contour mask; recalibrating the target detection confidence based on contour integrity, region consistency and boundary overlap; and suppressing false detections for detection results with unstable contours or inconsistent with candidate regions.
[0023] The second aspect of this application provides a reliability-guided aerial small target detection and segmentation system, comprising:
[0024] The image acquisition module is used to acquire RGB aerial images and infrared aerial images of the same scene;
[0025] The coarse registration module is used for time synchronization, radiometric normalization, and coarse registration of dual-modal images;
[0026] The modal reliability assessment module is used to generate modal reliability maps based on image degradation features.
[0027] The offset compensation module is used to estimate the cross-modal local offset field and generate sparse candidate regions;
[0028] The joint gating fusion module is used to output a fused feature map through a reliability-offset joint gating fusion network;
[0029] The small target detection module is used to output the target bounding box and target category;
[0030] The contour segmentation module is used to generate a target contour mask based on the target bounding box;
[0031] The closed-loop refinement module is used to correct the detection results based on the target contour mask and output the final detection and segmentation results.
[0032] In at least one embodiment of this application, the modal reliability assessment module includes a quality degradation feature extraction unit and a reliability map generation unit. The quality degradation feature extraction unit is used to extract at least one feature from the following: brightness stability, texture clarity, thermal saliency, noise intensity, occlusion ratio, and missing degree. The offset compensation module includes a local correlation estimation unit and a candidate region screening unit. The closed-loop refinement module includes a bounding box correction unit, a confidence recalibration unit, and a false detection suppression unit. The bounding box correction unit corrects the target bounding box based on the outer region of the target contour mask. The confidence recalibration unit updates the target confidence based on contour integrity and boundary overlap. The false detection suppression unit removes abnormal detection results based on contour stability and region consistency.
[0033] The invention has at least the following beneficial technical effects:
[0034] The reliability-guided aerial small target detection and segmentation method proposed in this application has the following advantages: First, it explicitly models the reliability of different modalities in different regions to avoid excessive interference from degraded regions on the fusion results. Second, it reduces feature misalignment under weak registration conditions through cross-modal local offset compensation. Third, it improves the target saliency and class separability of sparse small targets by dynamically allocating fusion weights at the region level through a reliability-offset joint gating fusion network. Fourth, it improves the consistency and stability of target bounding boxes and contour results through a detection and segmentation closed-loop refinement mechanism. Attached Figure Description
[0035] Figure 1 This is a flowchart of a reliability-guided aerial small target detection and segmentation method according to one embodiment of this application;
[0036] Figure 2 This is a schematic diagram of dual-modal reliability assessment and offset field estimation according to one embodiment of this application;
[0037] Figure 3 This is a schematic diagram of a reliability-offset joint gating fusion network according to one embodiment of this application;
[0038] Figure 4 This is a flowchart of the detection, segmentation, closed-loop refinement, and output process according to one embodiment of this application;
[0039] Figure 5 This is a schematic diagram of a reliability-guided aerial small target detection and segmentation system according to one embodiment of this application. Detailed Implementation
[0040] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.
[0041] The following is in conjunction with the appendix Figures 1 to 5 This application will be described in further detail.
[0042] This application provides a reliability-guided aerial small target detection and segmentation method, such as... Figure 1 As shown, it includes the following steps:
[0043] Step S1: Acquire RGB aerial images and infrared aerial images of the same scene, and perform time synchronization, radiometric normalization and coarse registration.
[0044] In one implementation, the acquisition platform can be a UAV, an airborne pod, or a fixed-wing inspection platform. RGB aerial images are used to provide texture, color, and edge structure information, while infrared aerial images are used to provide thermally salient target information. Considering the attitude changes of the aerial platform, lens installation errors, and frame rate differences between different sensors, timestamp matching is first performed on the dual-modal images. Then, the images are normalized in size and radiometric response according to the platform attitude, focal length, and imaging size. Finally, coarse registration is completed using global feature points or attitude priors to ensure that the same target is in adjacent regions in both modalities, so that the dual-modal images meet the initial alignment conditions for subsequent local offset estimation.
[0045] Step S2: Based on the brightness, contrast, noise, fogging, occlusion, and missing features of the RGB aerial images and infrared aerial images, perform a quality degradation assessment and generate the corresponding modal reliability map.
[0046] Unlike traditional bimodal fusion methods that assume equal weighting of both modalities in the calculation, this application first determines the reliability of each modality within the current region. For RGB aerial images, features such as brightness stability, local contrast, texture clarity, occlusion ratio, and edge interruption degree can be extracted; for infrared aerial images, features such as thermal saliency, thermal boundary continuity, thermal noise intensity, edge continuity, and thermal drift degree can be extracted. These features are then mapped to pixel-level or block-level quality scores to obtain a modal reliability map. The modal reliability map characterizes the reliability of a corresponding region in the current environment for participating in fusion calculations; a higher modal reliability map value indicates that the corresponding modality is more suitable for participating in subsequent fusion and detection in that region.
[0047] Step S3: Estimate the cross-modal local migration field based on the local correlation between RGB aerial images and infrared aerial images, and generate sparse candidate regions by combining the modal reliability map.
[0048] Because RGB and infrared images often exhibit local misalignment, parallax differences, and heat source diffusion in aerial photography, direct pixel-level fusion can easily lead to target blurring and feature contamination. To address this, this application extracts local corresponding blocks within a multi-scale feature space and estimates the offset vector of each candidate region through cross-modal correlation, geometric constraints, and adjacent region smoothing constraints, thus constructing a cross-modal local offset field. Subsequently, sparse candidate regions are selected by comprehensively considering thermal saliency, structural complexity, modal reliability, and offset consistency to reduce the interference of large areas of blank background on subsequent detection. Regions more likely to contain real targets are prioritized, reducing the consumption of network inference resources by large areas of blank background.
[0049] like Figure 2 As shown, the modal reliability map and the cross-modal local migration field work together to generate the sparse candidate region map. Regions with low reliability and unstable migration can be processed later or have their fusion weight reduced, while regions with significant thermal response and high structural continuity are preferentially included in subsequent detection branches.
[0050] Step S4: Input the RGB aerial image, infrared aerial image, modal reliability map, and cross-modal local offset field into the reliability-offset joint gating fusion network to obtain the fused feature map.
[0051] The innovative model structure of this application is reflected in the reliability-misalignment joint gating fusion network. This network does not directly stitch together bimodal features, nor does it simply employ a single attention module; instead, it constructs three collaborative branches. First, the reliability branch outputs modal validity weights for each region based on the modal reliability map, used to suppress modal regions severely affected by rain, fog, low light, occlusion, and missing features. Second, the misalignment compensation branch performs flexible alignment of regional features based on cross-modal local misalignment fields, ensuring that local target responses are corrected before fusion. Third, the joint gating branch generates region-level fusion weights based on modal validity weights, misalignment consistency, and sparse priors, determining the dominant relationship between RGB and infrared features in different regions at different levels.
[0052] like Figure 3 As shown, in the shallow feature stage, the joint gating branch prioritizes preserving edge contours, hotspot ranges, and fine-grained spatial distribution information to enhance the visibility of small targets. In the deep feature stage, the joint gating branch further enhances category semantics and cross-modal consistency to improve the classification stability of sparse small targets. The resulting fused feature map balances detailed representation and discriminative ability.
[0053] Step S5: Perform small target detection based on the fused feature map to obtain the target bounding box and target category.
[0054] In one implementation, the small target detection module may employ a dual-head structure consisting of a candidate screening detection head and a refinement detection head. The candidate screening detection head is used to quickly screen potential targets in sparse candidate regions, filtering out regions that clearly do not meet the target priors; the refinement detection head further regresses the bounding box position, scale, and category based on high-resolution features, especially refining the localization of small targets in aerial photography, which are characterized by small size, blurred edges, and strong background interference. As an optional implementation, the detection sub-network may adopt an improved single-stage detection architecture or a lightweight multi-scale detection architecture, but its specific network name is not a core limitation of this application.
[0055] Step S6: Perform bounding box guided contour segmentation based on the target bounding box to obtain the target contour mask.
[0056] To avoid the high-cost pixel-level segmentation of the entire aerial image, this application employs a bounding box-guided contour segmentation strategy. Specifically, the target bounding box output by the detection module is used as the seed region, and the corresponding local fusion features, boundary priors, and thermally salient priors are input into the contour segmentation network to obtain the target contour mask. The contour segmentation network can refine the target region at multiple scales and improve the contour closure degree through edge continuity constraints, region consistency constraints, and background suppression constraints, thereby reducing the adhesion between the target and the background.
[0057] Step S7: Perform closed-loop correction on the target bounding box and target confidence based on the target contour mask, and output the target category, target location and target contour information.
[0058] like Figure 4 As shown, this application does not treat detection and segmentation as two independent endpoints, but instead constructs a closed-loop refinement mechanism for detection and segmentation. First, the target bounding box is recalculated based on the circumscribed region of the target contour mask to correct for issues such as the initial detection box being too large, too small, or locally offset. Second, the detection confidence is recalibrated based on contour integrity, boundary overlap, and region consistency, enhancing the confidence of high-contour, reliable targets and reducing the weight of unstable targets. Third, false detections are suppressed for results with abnormally fragmented contours, significant inconsistencies with candidate regions, or overlaps and conflicts with adjacent targets, thereby improving the stability of the final output.
[0059] In one training implementation, the reliability-offset joint gating fusion network can be optimized through multi-task joint training. During training, reliability supervision terms, offset consistency supervision terms, detection loss terms, and contour segmentation loss terms can be introduced simultaneously, enabling reliability modeling, local alignment, object detection, and contour segmentation to converge collaboratively within the same optimization framework. For degraded samples in the training data, the model's adaptability to complex environments can be enhanced by simulating rain, fog, low light, local missing data, and occlusion perturbations.
[0060] In one inference implementation, the bimodal image is first synchronized, coarsely registered, and its quality degradation is assessed. Then, sparse candidate regions are generated based on the reliability map and offset field, and enhancement fusion and fine detection are performed only on priority regions. Subsequently, bounding box-guided contour segmentation is performed, and the detection bounding box and target confidence are corrected using the segmentation results to close loops. Finally, the target category, center position, bounding box, target contour, and their visualization results are output. This inference process can balance the detection accuracy and boundary stability of sparse small targets in complex environments.
[0061] The second aspect of this application provides a reliability-guided aerial small target detection and segmentation system, the structure of which is as follows: Figure 5 As shown, it includes an image acquisition module, a coarse registration module, a modal reliability assessment module, an offset compensation module, a joint gating fusion module, a small target detection module, a contour segmentation module, and a closed-loop refinement module.
[0062] The image acquisition module is used to acquire RGB aerial images and infrared aerial images of the same scene;
[0063] The coarse registration module is used for time synchronization, radiometric normalization, and coarse registration of dual-modal images;
[0064] The modal reliability assessment module is used to generate modal reliability maps based on image degradation features.
[0065] The offset compensation module is used to estimate the cross-modal local offset field and generate sparse candidate regions;
[0066] The joint gating fusion module is used to output a fused feature map through a reliability-offset joint gating fusion network;
[0067] The small target detection module is used to output the target bounding box and target category;
[0068] The contour segmentation module is used to generate a target contour mask based on the target bounding box;
[0069] The closed-loop refinement module is used to correct the detection results based on the target contour mask and output the final detection and segmentation results.
[0070] In a preferred embodiment of this application, the modal reliability assessment module includes a quality degradation feature extraction unit and a reliability map generation unit. The quality degradation feature extraction unit is used to extract at least one feature from the following: brightness stability, texture clarity, thermal saliency, noise intensity, occlusion ratio, and missing degree. The offset compensation module includes a local correlation estimation unit and a candidate region screening unit. The closed-loop refinement module includes a bounding box correction unit, a confidence recalibration unit, and a false detection suppression unit. The bounding box correction unit corrects the target bounding box based on the outer region of the target contour mask. The confidence recalibration unit updates the target confidence based on contour integrity and boundary overlap. The false detection suppression unit removes abnormal detection results based on contour stability and region consistency.
[0071] Each module can be deployed in the processor as software, or in an integrated hardware and software configuration in airborne terminals, edge computing devices, or ground workstations.
[0072] Compared with existing technologies, this application has at least the following advantages: First, by modal reliability modeling and cross-modal local offset compensation, it solves the problem of distortion caused by direct fusion of dual-modal aerial images in complex environments; Second, by using a reliability-offset joint gating fusion network, it dynamically determines the fusion strength and feature flow direction at the regional level, improving the saliency and class separability of sparse small targets; Third, by using sparse candidate region guidance and bounding box-guided contour segmentation, it reduces the interference of large-area background on detection and segmentation; Fourth, by using a closed-loop refinement mechanism for detection and segmentation, it further improves the consistency and stability of target bounding boxes and contour results.
[0073] It should be noted that the network backbone, detection head form, segmentation head form, loss function form, and specific solution methods for the local offset field mentioned in the above embodiments can all be replaced or adjusted without departing from the core idea of this application. All equivalent technical solutions based on modal reliability modeling, cross-modal local offset compensation, regional joint gating fusion, and closed-loop refinement of detection and segmentation should fall within the protection scope of this application.
[0074] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A reliability-guided aerial small target detection and segmentation method, characterized in that, include: Step S1: Acquire RGB aerial images and infrared aerial images of the same scene, and perform time synchronization, radiometric normalization, and coarse registration. Step S2: Perform a quality degradation assessment based on the brightness, contrast, noise, fogging, occlusion, and missing features of the RGB aerial image and the infrared aerial image, and generate the corresponding modal reliability map; Step S3: Estimate the cross-modal local migration field based on the local correlation between the RGB aerial image and the infrared aerial image, and generate sparse candidate regions by combining the modal reliability map; Step S4: Input the RGB aerial image, the infrared aerial image, the modal reliability map, and the cross-modal local offset field into the reliability-offset joint gating fusion network to obtain the fused feature map; Step S5: Perform small target detection based on the fused feature map to obtain the target bounding box and target category; Step S6: Perform bounding box guided contour segmentation based on the target bounding box to obtain the target contour mask; Step S7: Perform closed-loop correction on the target bounding box and target confidence based on the target contour mask, and output the target category, target location and target contour information.
2. The reliability-guided aerial small target detection and segmentation method according to claim 1, characterized in that, Step S1 includes: performing timestamp matching on RGB aerial images and infrared aerial images; normalizing the size and radiometric response of images of different modalities; and using a coarse registration method based on global feature points and platform attitude information to ensure that the dual-modal images meet the initial alignment conditions for subsequent local offset estimation.
3. The aerial small target detection and segmentation method based on reliability guidance according to claim 1, characterized in that, Step S2 includes: extracting brightness stability, texture clarity, and occlusion ratio features from RGB aerial images; extracting thermal saliency, thermal noise intensity, and edge continuity features from infrared aerial images; performing region-level quality scoring on the dual-modal images based on the features, and generating pixel-level or block-level modal reliability maps; wherein, the modal reliability map is used to characterize the credibility of the corresponding region in participating in fusion calculation under the current environment.
4. The aerial small target detection and segmentation method based on reliability guidance according to claim 1, characterized in that, Step S3 includes: extracting local corresponding blocks from the multi-scale feature maps of the RGB aerial image and the infrared aerial image; estimating the offset vectors of each candidate region based on cross-modal correlation and geometric constraints; combining the offset vectors into a cross-modal local offset field; and filtering sparse candidate regions based on thermal saliency, structural complexity, and offset consistency.
5. The reliability-guided aerial small target detection and segmentation method according to claim 1, characterized in that, In step S4, the reliability-offset joint gating fusion network includes a reliability branch, an offset compensation branch, and a joint gating branch; the reliability branch outputs bimodal region validity weights based on the modal reliability map; the offset compensation branch performs flexible alignment of region features based on the cross-modal local offset field; the joint gating branch combines the bimodal region validity weights, offset consistency, and sparse priors to generate region-level fusion weights to control the fusion strength and feature flow direction of RGB features and infrared features. The joint gating branch adopts a progressive fusion strategy, which preserves target edge and thermally significant difference information in the shallow feature stage, and enhances semantic consistency and class separability in the deep feature stage.
6. The aerial small target detection and segmentation method based on reliability guidance according to claim 1, characterized in that, Step S5 includes: constructing a small target detection network based on the fused feature map; the small target detection network includes at least a candidate screening detection head and a refinement detection head, the candidate screening detection head is used to quickly screen out low-confidence background regions, and the refinement detection head is used to refine the bounding box position, scale and category of candidate targets.
7. The reliability-guided aerial small target detection and segmentation method according to claim 1, characterized in that, Step S6 includes: generating a segmentation seed region based on the target bounding box; feeding the segmentation seed region and the corresponding fusion feature into a contour segmentation network to obtain a target contour mask; the contour segmentation network refines the target region at different scales and suppresses background leakage based on edge continuity constraints.
8. The aerial small target detection and segmentation method based on reliability guidance according to claim 1, characterized in that, Step S7 includes: correcting the position and size of the target bounding box based on the outer region of the target contour mask; recalibrating the target detection confidence based on contour integrity, region consistency and boundary overlap; and suppressing false detections for detection results with unstable contours or inconsistent with candidate regions.
9. A reliability-guided aerial small target detection and segmentation system, characterized in that, include: The image acquisition module is used to acquire RGB aerial images and infrared aerial images of the same scene; The coarse registration module is used for time synchronization, radiometric normalization, and coarse registration of dual-modal images; The modal reliability assessment module is used to generate modal reliability maps based on image degradation features. The offset compensation module is used to estimate the cross-modal local offset field and generate sparse candidate regions; The joint gating fusion module is used to output a fused feature map through a reliability-offset joint gating fusion network; The small target detection module is used to output the target bounding box and target category; The contour segmentation module is used to generate a target contour mask based on the target bounding box; The closed-loop refinement module is used to correct the detection results based on the target contour mask and output the final detection and segmentation results.
10. The aerial small target detection and segmentation system based on reliability guidance according to claim 9, characterized in that, The modal reliability assessment module includes a quality degradation feature extraction unit and a reliability map generation unit. The quality degradation feature extraction unit is used to extract at least one feature from brightness stability, texture clarity, thermal saliency, noise intensity, occlusion ratio, and missing degree. The offset compensation module includes a local correlation estimation unit and a candidate region screening unit. The closed-loop refinement module includes a bounding box correction unit, a confidence recalibration unit, and a false detection suppression unit. The bounding box correction unit corrects the target bounding box based on the outer region of the target contour mask. The confidence recalibration unit updates the target confidence based on contour integrity and boundary overlap. The false detection suppression unit removes abnormal detection results based on contour stability and region consistency.