A remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement

Abstract

A remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement is disclosed. The method includes the following steps: Step 1: The main branch gradually generates a set of representation points through initial and refinement stages to achieve location determination and type recognition; Step 2: A progressive class-related dual branch runs parallel to the main branch as an auxiliary branch, implicitly guiding the network to learn more robust appearance and semantic feature embeddings and adaptively generate high-quality prediction point sets; Step 3: The instance-guided enhancement module fully utilizes the abundant class-related semantic information in the auxiliary branches to explicitly enhance the discriminative features of the target in the main branch; Step 4: The optimization objectives during training mainly include classification, localization, and instance segmentation, improving the model's feature representation and generalization ability through multi-task learning. This method fully considers the target's special cross-shaped geometric structure and, combined with multi-task learning and interactive attention mechanisms, can be used to achieve accurate aircraft detection and recognition.

Description

Technical Field

[0001] This invention belongs to the field of target detection and recognition technology, and relates to an aircraft target detection and recognition method for optical remote sensing images, specifically a remote sensing image aircraft detection and recognition method that combines geometric prior progressive instance enhancement. Background Technology

[0002] Against the backdrop of increasing global emphasis on air superiority, the detection and identification of high-strategic-value targets, such as aircraft, is a current research hotspot in the field of remote sensing image processing, ensuring that air superiority is not compromised. Its development has significant implications for both military and civilian applications. In the military field, accurate detection of various military aircraft allows for the evaluation and analysis of enemy military dynamics, providing support for battlefield defense and early warning, as well as supporting military deployments in air superiority zones, thus providing strong guarantees for national security. In the civilian field, the detection and identification of aircraft targets based on remote sensing images can provide strong technical support for air traffic control. By analyzing the number and location of parked aircraft at airports and historical information, flight schedules, arrival and departure routes, and parking locations can be scientifically arranged.

[0003] Unlike natural images where targets are extracted using horizontal bounding boxes, remote sensing images, due to their unique overhead perspective, typically require rotated bounding boxes for precise target localization. Rotated target detection can determine the target's location and category simultaneously, as well as its orientation. It can be broadly categorized into two types: anchor-box-based detection and keypoint-based detection. Anchor-box-based methods inherit the processing flow from R-CNN and YOLO. These methods output orientation boxes by adding additional angle regression to the horizontal anchor boxes, while relying on special feature alignment operators to mitigate the mismatch between horizontal candidate boxes and rotated targets. However, due to the inherent periodicity of angle orientation, this modified angle regression often leads to discontinuous loss and inconsistent regression results.

[0004] In comparison, keypoint-based detection methods abandon manually set anchor boxes and use the prediction of key points such as diagonals and centers to achieve target detection. Among them, Oriented Repositories represent bounding boxes through point sets and has the potential to detect targets with different orientations, shapes, and postures. The steps of predicting point sets in this type of method are mainly divided into two parts: first, a preliminary point set is generated through ordinary convolution, and then deformable convolution is used to further mine the target's geometric structure information to generate a refined point set. After the point set is generated, a transformation function is used to generate a rotated box. However, this type of method lacks effective information guidance during the point set generation process, and only updates the network parameters by backpropagating the bounding boxes generated based on the point set envelope and the ground truth loss. Aircraft targets in remote sensing overhead scenes have a more special geometric structure than ships and vehicles, with an overall cross-shaped irregular structure. Without effective information supervision, the network has a greater difficulty in generating predicted point sets that represent the shape of aircraft targets. Moreover, the low duty cycle of aircraft introduces more background interference, affecting the representation of key target features and further increasing the probability of generating abnormal predicted points. In reality, aircraft, as a type of rigid body target, possess distinct geometric morphological characteristics. Their cross-shaped structure, formed by the fuselage and wings, allows them to be identified by humans. Therefore, introducing additional supervisory information containing prior knowledge about the aircraft can guide the point set to more fully represent key geometric and semantic features, enabling more accurate detection and identification of aircraft targets. Summary of the Invention

[0005] To address the challenges of precise location determination and category confusion caused by the unique structure and low duty cycle of aircraft in remote sensing scenarios, this invention proposes a remote sensing image aircraft detection and recognition method combining geometric prior and progressive instance enhancement. The proposed method fully considers the target's unique cross-shaped geometric structure and, by combining multi-task learning and interactive attention mechanisms, constructs a progressive class-related dual-branch and instance-guided enhancement module to fully mine the geometric semantic information of the aircraft target. It implicitly guides and explicitly enhances the representation of effective target features to achieve accurate aircraft detection and recognition.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement, the method being as follows:

[0008] Step 1: Based on the extraction of multi-scale features of the image by the backbone network and feature pyramid, the main branch gradually generates a set of representation points and extracts key features of the target through the initial stage and the refinement stage. The location determination and type recognition are realized in the localization branch and the classification branch, respectively.

[0009] Step 2: The progressive class-related dual branch runs in parallel with the main branch as an auxiliary branch. It provides additional supervision information, including the geometric prior of the aircraft target, through the coarse instance branch and the fine instance branch. This implicitly guides the network to learn more robust appearance and semantic feature embeddings and adaptively generate high-quality prediction point sets.

[0010] Step 3: The instance-guided enhancement module makes full use of the abundant class-related semantic information in the subsidiary branches and realizes the flow of instance-level information through the interactive attention mechanism, explicitly enhancing the identifiable features of the target in the initial and refinement stages of the main branch.

[0011] Step 4: The optimization objectives during training include three parts: classification, localization, and instance segmentation. Multi-task learning is used to simultaneously learn multiple related tasks within the same model to improve the model's feature representation and generalization ability. The three loss components originate from the classification branch and localization branch in Step 1, and the progressive class-related dual branch in Step 2, respectively.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] (1) This invention constructs an anchorless aircraft target detection and recognition network based on point set representation. It utilizes progressive class-related dual branches and instance-guided enhancement modules to fully mine the inherent geometric semantic information of rigid targets and adaptively integrates class-related semantic information into the localization and classification branches, thereby fully improving the ability to locate and identify aircraft targets.

[0014] (2) This invention designs a progressive class-related dual branch to fully utilize the prior knowledge of the special cross-shaped geometric structure of the aircraft target and implicitly guide the generation of a high-quality prediction point set. Based on the idea of ​​multi-task learning, this branch uses coarse foreground masks and fine cross-shaped masks to gradually guide the network to learn more robust appearance and semantic feature embeddings, and suppress the generation of abnormal prediction points in redundant background regions.

[0015] (3) This invention proposes an instance-guided enhancement module to fully utilize the efficient feature embedding in the auxiliary branches and explicitly enhance the distinguishable features of aircraft targets. This module adaptively fuses the rich class-related semantic information in the mask branch with the feature maps in the localization and classification branches through an interactive attention mechanism, thereby realizing the flow of instance-level information. Attached Figure Description

[0016] Figure 1 This is the overall architecture of the method proposed in this invention;

[0017] Figure 2 Example of mask truth used for progressive class-related bi-branch;

[0018] Figures 3-6 This is an example of aircraft target detection results. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a remote sensing image aircraft detection and recognition method that combines geometric prior and progressive instance enhancement. Based on the anchorless point set representation mode and the prior of aircraft geometry, it explicitly enhances the effective geometric and semantic feature representation of the target through progressive class-related bi-branch implicit guidance and instance-guided enhancement modules, thereby achieving accurate aircraft detection and recognition. Figure 1 As shown, the method includes the following steps:

[0021] Step 1: Based on the extraction of multi-scale features of the image by the backbone network and feature pyramid, the main branch gradually generates a set of representation points and extracts key features of the target through the initial stage and the refinement stage. The location determination and type recognition are realized in the localization branch and the classification branch, respectively.

[0022] Step 1-1: The backbone network extracts features at different scales from the input image, and then fuses high-level semantic features and low-level detail features via a feature pyramid. The feature pyramid outputs multi-level features with downsampling times of 3-7, denoted as F. n (n = 3, 4, 5, 6, 7).

[0023] Steps 1-2: In order to adapt to the special geometric structure and variable scale and orientation of aircraft targets in remote sensing scenes, the main branch of the prediction head uses a set of representation points for prediction, including regressing the target position through the localization branch and determining the target category through the classification branch.

[0024] Step 1-2-1: After the localization branch is attached to the feature pyramid, the generation of the representation point set offset is achieved step by step through two stages: initialization and refinement. Then, the target bounding box is generated through a transformation function. The pattern based on the representation point set prediction can adapt to target detection of different scales, orientations, and shapes. The target is represented by the point set R adaptively generated at the feature grid points.

[0025]

[0026] In the formula, n is the number of representation points, p i It is the i-th point in the set. In this representation model, the size, orientation, geometry, and morphological information of the target are implied through the relative positions of different points in the point set.

[0027] Step 1-2-1-1: In the initial stage, the features generated by the feature pyramid in Step 1-1 are first processed through l 3×3 cascaded convolutional layers (l is 3 by default) to obtain the initial localization features. Then, a learnable initial point set offset is generated at each feature grid point through a 1×1 convolution. If we let (x, y) represent the initial position of the grid point in the feature map, then the initialization of the representation point R1 is denoted as:

[0028]

[0029] Step 1-2-1-2: In the refinement stage, the refined localization features corresponding to the initial point set offset δ1 in the feature map are extracted through deformable convolution.

[0030]

[0031] In the formula, w i This represents the contribution of each point to the feature and is a learnable parameter in deformable convolution. Further adjustments to the point set offset are then generated. The final set of representation points output by the localization branch is denoted as:

[0032]

[0033] Steps 1-2-1-3: To facilitate supervised training of the network under existing ground truth and to achieve target rotation box prediction, it is necessary to transform the point set into a rotation box (OB) that encloses the target using a transformation function. The transformation process is denoted as:

[0034] OB = G(R²)

[0035] In the formula, OB is the rotated bounding box transformed by the point set, including the target's center position, size, and rotation angle. G is the transformation function, and during training, the differentiable function ConvexHull is selected to optimize the adaptive point learning process. During inference, MinAeraRect is used for post-processing to predict the target's rotated bounding box.

[0036] Step 1-2-2: In the classification branch, the features generated by the feature pyramid in Step 1-1 are first passed through l 3×3 cascaded convolutional layers (l defaults to 3) to obtain initial classification features. Then, refined classification features are extracted from the feature map based on the biases predicted in deformable convolution and localization branches. Finally, the target category is confirmed using Softmax.

[0037] Step 2: The progressive class-related dual branch runs in parallel with the main branch as an auxiliary branch. It provides additional supervision information, including the geometric prior of the aircraft target, through the coarse instance branch and the fine instance branch. This implicitly guides the network to learn more robust appearance and semantic feature embeddings and adaptively generate high-quality prediction point sets.

[0038] Step 2-1: The coarse instance branch and the fine instance branch have simple and similar structures to ensure that performance is improved without adding too much additional computational burden. The difference lies in the different supervision information during training. For the nth level feature output by the feature pyramid... H, W, and C represent the height, width, and number of channels of the feature map, respectively. First, a class-related reference feature map is obtained through l 3×3 convolutions (l defaults to 3).

[0039] Step 2-2: For the class-related reference feature map obtained in Step 2-1, directly generate the instance segmentation prediction result through a 1×1 convolution. N represents the number of target categories, and N+1 represents the background categories covered.

[0040] Steps 2-3: During training, the instance segmentation prediction result R in the two branches n The loss is calculated using the same mask ground truth method, and backpropagation optimizes parameter updates, implicitly guiding the representation point set to gradually move along the background-coarse foreground-fine ontology path. Simultaneously, under the supervision of the class-related mask ground truth method, R... n The R instance contains implicit instance-level information, which can be subsequently enhanced using the instance-guided enhancement module in step 3. n The feature maps in the main branch are fused together to achieve explicit enhancement of the target features.

[0041] Step 2-3-1: The truth value M of the coarse mask in the coarse instance branch corase The transformation is performed directly using the truth value of the rotated bounding box. The value is 0 in the background area, and the foreground area is filled with the corresponding target category. This branch is used to assist in focusing on the target foreground area, corresponding to the initial stage in the localization branch.

[0042] Step 2-3-2: Considering the cross-shaped structure of the aircraft target, it is difficult to extract its inherent geometric features by simply segmenting through rectangular instances. Therefore, the fine instance branching uses a cross-shaped fine mask ground truth M. refine Such ground truth values ​​more closely match the actual contours of aircraft targets, reducing the impact of redundant backgrounds on the representation of key target features and further guiding the adaptive point set to mine the inherent geometric structural characteristics of the target.

[0043] Step 2-3-2-1: The aircraft target presents a cross-shaped form composed of the fuselage and wings, which can be simulated by two mutually perpendicular ellipses. Let the true value of the target be the center position (x). c ,yc Given the target dimensions (w, h) and orientation angle θ, let the two ellipses representing the aircraft target cross mask be along the length and width of the rotating frame, respectively, to simulate the fuselage and wing. The longer sides of the two ellipses are the target ground truth values ​​h and w, and the shorter sides are w / τ and h / τ, respectively. The equations of the two ellipses are then written as:

[0044]

[0045]

[0046] In the formula, τ is an adjustment factor used to change the thickness of the longer side relative to the shorter side. When τ is 1, the true value degenerates into an elliptical mask.

[0047] Step 2-3-2-2: Fill the interior of the two elliptical regions with the target class cls, and fill the remaining regions with 0. This generates a cross-shaped mask ground truth for supervision of fine instance branches. Cross-shaped mask ground truth M refine Notation:

[0048]

[0049] Step 3: The instance-guided enhancement module makes full use of the abundant class-related semantic information in the subsidiary branches and realizes the flow of instance-level information through the interactive attention mechanism, explicitly enhancing the identifiable features of the target in the initial and refinement stages of the main branch.

[0050] Step 3-1: To achieve progressive enhancement of the high-quality point set generation, the feature maps of the classification branch and localization branch in the initial stage (corresponding to those in step 1-2-1-1) are... and in step 1-2-2 The feature maps of the classification and localization branches in the refinement stage are enhanced using reference features from the coarse instance branch; while the feature maps of the classification and localization branches in the refinement stage (corresponding to those in step 1-2-1-2) are enhanced. and in step 1-2-2 Enhancement is achieved through reference features in fine instance branches.

[0051] Step 3-2: Let the features to be enhanced in the main branch be... The reference feature used in the auxiliary branch for enhancement is The enhanced features are then obtained through the output of the feature-guided enhancement module.

[0052] Step 3-2-1: The feature-guided enhancement module fully utilizes contextual semantic information through an interactive attention mechanism, where query Q and value V originate from F. ori It is used to query key values ​​and identify aggregated and enhanced feature blocks, and is defined as Q = F. ori and V=Fori W v W v It is an embedding matrix implemented by 1×1 convolution. The key K comes from R and is defined as K = R, where sufficient instance information is used for query matching to determine F. ori Explicit enhancement.

[0053] Step 3-2-2: Process the key K through k×k grouped convolution to mine the static instance information K between local neighborhoods within the k×k grid. s K s It reflects the static context representation between local neighborhoods in R.

[0054] Step 3-2-3: Based on the concatenated static context representation K s And query Q, generate attention weights writing:

[0055] A = [K] s ,Q]W1W2

[0056] In the formula, C h To determine the number of heads in multi-head attention, W1 and W2 are alternating 1×1 convolutions, with a ReLU activation function added to W1. Features at all spatial locations in A are k×k×C. h A vector of dimension K. Unlike directly performing dot products on isolated query-key pairs, this attention is in the static context K. s Guided by this, we further explored the relationships between various queries and keys.

[0057] Step 3-2-4: Based on the obtained attention weights, aggregate and enhance the features of value V within the k×k grid to obtain the attention-enhanced feature K. enh ,writing:

[0058]

[0059] In the formula, K is the local matrix dot product operator within a k×k grid. enh The dynamic interaction between reference features and input features is explored, which is called dynamic instance augmentation of input features.

[0060] Step 3-2-5: Utilize the channel attention fusion structure to integrate static instance information K s and the feature K dynamically enhanced by instance information enh Adaptive fusion is performed to obtain the final enhanced feature F. fuse Channel attention can adaptively learn the importance of features from different channels and achieve feature aggregation through weight redistribution.

[0061] Step 3-2-5-1: First, for K s and K enh F is obtained by element-wise summation. sum Then, global pooling is performed along the spatial dimension to obtain channel-related global features. The method for calculating the k-th channel feature of G is as follows:

[0062]

[0063] In the formula, F sum (i,j,k) is F sum The value in the i-th row, j-th column, and k-th channel.

[0064] Step 3-2-5-2: Further explore the interdependencies between channels through fully connected layers to obtain more compact feature descriptors. writing:

[0065] G c =F fc (G) = ReLU(B(W3G))

[0066] In the formula, F fc It is a fully connected layer, consisting of an embedding matrix. It consists of batch normalization (B) and the nonlinear activation function ReLU. r is the channel scaling factor, used to reduce dimensionality and improve computational efficiency.

[0067] Step 3-2-5-3: Based on feature descriptor G c Using the Softmax operator, the channel soft attention weights w1 and w2 are calculated and written as:

[0068]

[0069] In the formula, A(k) and B(k) are respectively and The k-th row, w1(k) and w2(k) are the k-th elements of w1 and w2 respectively.

[0070] Step 3-2-5-4: Using the attention weights generated in step 3-2-5-3 to pair K s and K enh The weighted summation yields the final enhanced feature F. fuse ,writing:

[0071] F fuse =w1·K s +w2·K enh

[0072] Step 4: The optimization objectives during training include three parts: classification, localization, and instance segmentation. Multi-task learning is used to simultaneously learn multiple related tasks within the same model to improve the model's feature representation and generalization ability. The three loss components originate from the classification branch and localization branch in Step 1, and the progressive class-related dual branch in Step 2, respectively.

[0073] The overall loss function during training is written as follows:

[0074] L = L cls +L loc +λL mask

[0075] In the formula, λ is used to adjust the loss weight (default 0.05), and L... cls L loc and L mask These are classification loss, localization loss, and instance segmentation loss, respectively.

[0076] Step 4-1: Classification Loss L cls It originates from the classification loss and is directly calculated using the focus loss.

[0077] Step 4-2: Localization loss L loc Derived from the localization branch, including the initial stage loss L loc1 and refinement stage loss L loc2 Two parts. The localization loss in both stages continues to use the intersection-and-exchange loss L. IoU and spatial constraint loss L s.c Composition. L IoU Generated using the intersection-union-ratio (OCR) constraint bounding box, written as:

[0078]

[0079] In the formula, N p F is the number of positive samples. GIoU For the GIoU function, and These represent the bounding box and category to which the truth value is assigned, respectively. This is a directed polygon obtained by transforming a set of points using a transformation function.

[0080] L s.c This is used to constrain the generation of representation points within the bounding box. Let r... o and r c Let be points outside the truth box and the truth center point, respectively. Then the penalty function ρ ij and L s.c writing:

[0081]

[0082] In the formula, N o and N a These represent the number of points outside the truth box for each point and the number of positive sample points assigned to each target, respectively.

[0083] Step 4-3: Instance Segmentation Loss L mask Derived from the asymptotically correlated bibranch, by the coarse mask loss L mask1 And fine mask loss L mask2 The components correspond to the initial and fine-tuning stages of the localization process. The mask loss is calculated using a pixel-level Softmax cross-entropy function.

[0084] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement, characterized in that, The method includes the following steps: Step 1: Based on the extraction of multi-scale features of the image by the backbone network and feature pyramid, the main branch gradually generates a set of representation points and extracts key features of the target through the initial stage and the refinement stage. The location determination and type recognition are realized in the localization branch and the classification branch, respectively. Step 2: The progressive class-related dual branch runs in parallel with the main branch as an auxiliary branch. It provides additional supervision information, including the geometric prior of the aircraft target, through the coarse instance branch and the fine instance branch. It implicitly guides the network to learn more robust appearance and semantic feature embeddings and adaptively generate high-quality prediction point sets. Step 3: The instance-guided enhancement module makes full use of the abundant class-related semantic information in the subsidiary branches, realizes the flow of instance-level information through the interactive attention mechanism, and explicitly enhances the identifiable features of the target in the initial and refinement stages of the main branch; Step 4: The optimization objectives during training include three parts: classification, localization, and instance segmentation. By learning multiple related tasks simultaneously in the same model through multi-task learning, the model's feature representation and generalization ability can be improved. The three losses are respectively derived from the classification branch and localization branch in Step 1 and the progressive class-related dual branch in Step 2.

2. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement as described in claim 1, characterized in that, The specific steps of step 1 include: Step 1-1: The backbone network extracts features at different scales from the input image, and then fuses high-level semantic features and low-level detail features through a feature pyramid. The feature pyramid outputs multi-level features with downsampling times of 3-7, denoted as... ,in, ; Steps 1-2: In order to adapt to the special geometric structure and variable scale and orientation of aircraft targets in remote sensing scenes, the main branch of the prediction head uses a set of representation points for prediction, including regressing the target position through the localization branch and determining the target category through the classification branch.

3. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement according to claim 2, characterized in that, The specific steps of steps 1-2 are as follows: Step 1-2-1: After the localization branch is attached to the feature pyramid, the generation of the representation point set offset is achieved step by step through two stages: initialization and refinement. Then, the target bounding box is generated through a transformation function. The prediction mode based on the representation point set can adapt to target detection of different scales, orientations, and shapes. This is achieved by adaptively generating point sets at feature grid points. Characterization target: In the formula, The number of representation points, It is the first one In this representation model, the size, orientation, geometry, and morphological information of the target are implied through the relative positions of different points in the point set. Step 1-2-2: In the classification branch, the features generated by the feature pyramid in Step 1-1 are first passed through l 3×3 cascaded convolutional layers to obtain initial classification features. Then, based on the biases predicted in deformable convolution and localization branches, refined classification features are extracted from the feature map. Finally, the target category is confirmed using Softmax.

4. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement as described in claim 3, characterized in that... The specific steps of step 1-2-1 are as follows: Step 1-2-1-1: In the initial stage, the features generated by the feature pyramid in Step 1-1 are first processed through l 3×3 cascaded convolutional layers to obtain initial localization features. Then, a 1×1 convolution is used to generate a learnable initial point set offset for each feature grid point. If This indicates the initial position of the grid points in the feature map, at which point the representation points are initialized. Notation: Step 1-2-1-2: In the refinement stage, the initial point set offset in the feature map is extracted using deformable convolution. Corresponding refined positioning features : In the formula, This represents the contribution of each point to the feature, and is a learnable parameter in deformable convolution, further generating the adjusted point set offset. Then the set of representation points output by the locating branch is denoted as: Steps 1-2-1-3: To facilitate supervised training of the network under existing ground truth and to achieve target rotation box prediction, a transformation function is needed to convert the point set into a rotation box that surrounds the target. The conversion process is denoted as: In the formula, The rotating frame is transformed by a point set, including the target's center position, size, and rotation angle; For the transformation function, the differentiable function ConvexHull is selected during training to optimize the learning process of adaptive points, while MinAeraRect is used for post-processing during inference to predict the target's bounding box.

5. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement according to claim 2, characterized in that, The specific steps of step 2 are as follows: Step 2-1: For the nth level feature output by the feature pyramid , , , These represent the height, width, and number of channels of the feature map, respectively. First, a class-related reference feature map is obtained through l 3×3 convolutions. ; Step 2-2: For the class-related reference feature map obtained in Step 2-1, directly generate the instance segmentation prediction result through a 1×1 convolution. , For the number of target types, The representative covers the background class; Steps 2-3: Instance segmentation prediction results in the two branches during training. The loss is calculated using the same mask ground truth method, and backpropagation optimizes parameter updates. Implicitly, the representation point set is gradually guided to move along the background-coarse foreground-fine ontology path, while under the supervision of the class-related mask ground truth method. It contains instance-level information.

6. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement as described in claim 5, characterized in that... The specific steps of steps 2-3 are as follows: Step 2-3-1: Coarse Mask Truth Value in Coarse Instance Branch The transformation is performed directly through the truth value of the rotating box. The value is 0 in the background area and the foreground area is filled with the corresponding target category. This branch is used to assist in focusing on the target foreground area, corresponding to the initial stage in the positioning branch. Step 2-3-2: Fine instance branches use cross-shaped fine mask ground truth. .

7. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement as described in claim 6, characterized in that... The specific steps of step 2-3-2 are as follows: Step 2-3-2-1: The aircraft target presents a cross-shaped form composed of the fuselage and wings, which can be simulated by two mutually perpendicular ellipses. Let the true value of the target be the center position. ,size and orientation angle Let the two ellipses representing the cross mask of the aircraft target be along the length and width of the rotating frame, respectively, to simulate the fuselage and wing. The long sides of the two ellipses are the true values ​​of the target. and The shorter sides are respectively and Then the equations of the two ellipses can be written as: In the formula, This is an adjustment factor used to change the thickness of the longer side relative to the shorter side. When the value is 1, the truth value degenerates into an elliptical mask; Step 2-3-2-2: Fill the interior of the two elliptical regions with the target category. If the remaining areas are filled with 0, a cross-shaped fine mask ground truth can be generated for monitoring fine instance branches. Notation: 。 8. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement as described in claim 5, characterized in that... The specific steps of step 3 are as follows: Step 3-1: In order to achieve progressive enhancement of the high-quality point set generation, the feature maps of the classification branch and the localization branch in the initial stage are enhanced by the reference features in the coarse instance branch; The feature maps of the classification and localization branches in the refinement stage are enhanced by reference features in the fine instance branches; Step 3-2: Let the features to be enhanced in the main branch be... The reference feature used in the enhanced subsidiary branch is Then, the enhanced features are obtained through the output of the feature-guided enhancement module. .

9. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement according to claim 8, characterized in that, The specific steps of step 3-2 are as follows: Step 3-2-1: The feature-guided enhancement module fully utilizes contextual semantic information through an interactive attention mechanism, where the query... Sum Source It is used to query key values ​​and identify aggregated and enhanced feature blocks, and is defined as and , It is an embedding matrix implemented by 1×1 convolution, key Source , is defined as Sufficient instance information will be queried and matched for [the purpose of] [the following]. Explicit enhancement; Step 3-2-2: Process the key using k×k grouped convolution. To mine static instance information between local neighborhoods within a k×k grid. , Reflects Static context representation between local neighborhoods; Step 3-2-3: Based on the concatenated static context representation and query Generate attention weights ,writing: In the formula, The number of heads in the bullish attention. and For alternating 1×1 convolutions, Add the ReLU activation function to . The feature at all spatial locations is a Unlike directly performing dot products on isolated query-key pairs, this attention to dimensional vectors, within a static context, provides a different perspective. Guided by this, we further explored the relationships between various queries and keys; Step 3-2-4: Based on the obtained attention weights, pair values Features within a k×k grid are aggregated and enhanced to obtain attention-enhanced features. ,writing: In the formula, For local matrix dot product operators in a k×k grid, The dynamic interaction between reference features and input features was explored, which is called dynamic instance augmentation of input features; Step 3-2-5: Utilize the channel attention fusion structure to integrate static instance information. Features dynamically enhanced by instance information Adaptive fusion is performed to obtain the final enhanced features. Channel attention can adaptively learn the importance of features from different channels and achieve feature aggregation through weight redistribution.

10. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement according to claim 9, characterized in that, The specific steps of step 3-2-5 are as follows: Step 3-2-5-1: First, for and Element-by-element summation Then, global pooling is performed along the spatial dimension to obtain channel-related global features. , The method for calculating the k-th channel feature is as follows: In the formula, for The value in the i-th row, j-th column, and k-th channel; Step 3-2-5-2: Further explore the interdependencies between channels through fully connected layers to obtain more compact feature descriptors. ,writing: In the formula, It is a fully connected layer, consisting of an embedding matrix. Normalization and nonlinear activation functions composition, The channel scaling factor is used to reduce dimensionality and improve computational efficiency. Step 3-2-5-3: Based on feature descriptors The channel soft attention weights are calculated using the Softmax operator. and ,writing: In the formula, and They are respectively and The kth row, and They are respectively and The kth element; Step 3-2-5-4: Attention weights generated through 3-2-5-3 and The weighted summaries yield the final enhanced features. ,writing: 。 11. The remote sensing image aircraft detection and recognition method combining geometric prior progressive instance enhancement according to claim 1, characterized in that, The specific steps of step 4 are as follows: The overall loss function during training is written as follows: In the formula, Used to adjust loss weights , and These are classification loss, localization loss, and instance segmentation loss, respectively. Step 4-1: Classification Loss It originates from the classification loss and is directly calculated through the focus loss; Step 4-2: Locating the loss Derived from the localization branch, including the initial stage loss. and the loss in the refinement stage The positioning loss in both parts and stages continues to use the intersection-and-comparison loss. and spatial constraint loss composition, Generated using the intersection-union-ratio (OCR) constraint bounding box, written as: In the formula, The number of positive samples. For the GIoU function, and These represent the bounding box and category to which the truth value is assigned, respectively. To represent the directed polygon obtained by the transformation function of the point set; To constrain the generation of representation points within the bounding box, let and Let the points be outside the truth box and the truth center point, respectively, then the penalty function is... and writing: In the formula, and These represent the number of points outside the truth box for each point and the number of positive sample points assigned to each target, respectively. Step 4-3: Instance Segmentation Loss Derived from the progressively correlated bibranch, by coarse mask loss and fine mask loss The composition corresponds to the initial stage and fine-tuning stage in the localization process, and the mask loss is calculated using the pixel-level Softmax cross-entropy function.

Images