Directed ship fine granularity identification method based on remote sensing image, electronic equipment and program product

Abstract

The invention discloses a directed ship fine-grained identification method based on a remote sensing image, a ship is identified through a ship fine-grained detection model, and the ship fine-grained detection model is constructed by the following steps: obtaining a ship identification data set; preprocessing the ship identification data set; taking ResNet-50 as a backbone network of the ship fine-grained detection model to extract ship multi-scale visual features, and outputting a multi-layer feature map; and inputting the preprocessed ship identification data set into the ship fine-grained detection model for forward propagation, back propagation and loss function calculation, completing model training and storing parameters. The ship fine-grained detection model comprises a double-flow feature fusion module, and the module fuses feature map information from each layer in two fusion directions from top to bottom and from bottom to top. The ship fine-grained detection model includes a polarized feature focusing module that captures key features related to a particular task by applying polarized attention.

Description

Technical Field

[0001] This disclosure belongs to the field of ship management technology, and specifically relates to a method, electronic equipment and program products for directional fine-grained ship identification based on remote sensing images. Background Technology

[0002] Ship identification is an image processing technique used to detect and classify ship instances in digital images. With advancements in remote sensing technology, ship identification has been widely applied in the maritime field. Optical remote sensing images, as an important data source, offer advantages such as intuitive understanding, rich semantic information, and high resolution, making them highly helpful for accurately identifying ship targets. In recent years, the improvement in the quality of optical remote sensing images, particularly in ship identification against complex backgrounds, has attracted widespread attention.

[0003] Most current research focuses solely on ship detection tasks, distinguishing only between ship targets and non-ship targets. Some coarse-grained ship identification methods roughly categorize ships into 2 to 5 classes, such as military and civilian vessels. Fine-grained ship identification requires classifying ships into more detailed types, such as tugboats, cruise ships, fishing boats, and speedboats. However, research on fine-grained ship identification remains relatively limited, despite its significant applications and value in areas such as ship safety, marine environmental protection, maritime traffic management, and navigation. For example, identifying the types and behaviors of surrounding vessels can improve a ship's autonomous navigation capabilities, reduce accident risks, and enhance traffic coordination between vessels. Furthermore, tracking fishing vessels and monitoring fisheries activities helps control illegal fishing, maintain sustainable fisheries development, and protect the health of marine ecosystems.

[0004] Compared with general target detection tasks, fine-grained ship identification in remote sensing images mainly presents the following three challenges:

[0005] 1) Arbitrary Direction: Ships often appear in remote sensing images in arbitrary directions. Traditional horizontal bounding boxes are no longer suitable for locating rotating targets because they may contain a lot of background information, leading to misleading detection results. While rotating anchors attempt to address this issue, they struggle to achieve good spatial alignment with ground truth boxes and fail to provide sufficiently good semantic information and regression accuracy.

[0006] 2) High aspect ratio: When a ship has a high aspect ratio, a small angular deviation between the predicted bounding box and the ground reality bounding box will cause the IoU to drop rapidly. Due to the distortion of the ship's shape and the inaccurate overlap between the predicted and ground reality bounding boxes, using IoU for label assignment may lead to incorrect classification results.

[0007] 3) Significant Variations in Appearance and Scale: In fine-grained ship identification, differences in regulations across countries and regions lead to substantial intra-class appearance variations for certain ship categories. Furthermore, ship categories with similar uses also exhibit inter-class similarities, potentially confusing the classification branches of the network. In addition, the scale of ships varies considerably, especially for small vessels, increasing the risk of loss or misidentification. Summary of the Invention

[0008] One embodiment of this disclosure provides a directed fine-grained ship identification method based on key feature extraction from remote sensing images. This method identifies ships using a fine-grained ship detection model, the construction steps of which include:

[0009] S1, Obtain the ship identification dataset;

[0010] S2, preprocessing the ship identification dataset;

[0011] S3, ResNet-50 is used as the backbone network of the ship fine-grained detection model to extract multi-scale visual features of the ship and output multi-layer feature maps;

[0012] S4. Input the preprocessed ship identification dataset into the ship fine-grained detection model for forward propagation, back propagation, and loss function calculation to complete model training and save parameters. Attached Figure Description

[0013] The above and other objects, features, and advantages of this disclosure will become readily apparent from the following detailed description of exemplary embodiments, taken in conjunction with the accompanying drawings. Several embodiments of this disclosure are illustrated in the drawings by way of example and not limitation, in which:

[0014] Figure 1 A model architecture diagram of a directed fine-grained ship identification method based on key feature extraction from remote sensing images, according to one embodiment of the present disclosure.

[0015] Figure 2 An improved polarization feature focusing model architecture diagram according to one embodiment of the present disclosure.

[0016] Figure 3 Example diagram of the HRSC2016 dataset according to one embodiment of this disclosure.

[0017] Figure 4 Example image of the visualization results of the HRSC2016 dataset of the present disclosure.

[0018] Figure 5 Example image of port scene visualization results based on one embodiment of the HRSC2016 dataset according to this disclosure.

[0019] Figure 6 Example diagram of the dense distribution of ships in the HRSC2016 dataset according to one of the embodiments of this disclosure. Detailed Implementation

[0020] To address the challenges posed by the aforementioned difficulties in high-performance remote sensing ship detection, this disclosure introduces the concept of key features and provides a directed fine-grained ship identification method based on key feature extraction from remote sensing images. This method aims to achieve fine-grained ship identification in optical remote sensing images. This disclosure employs a novel feature fusion network, powerful feature representation, and an efficient label allocation strategy.

[0021] According to one or more embodiments, a directed fine-grained ship identification method based on key feature extraction from remote sensing images includes the following steps:

[0022] S1. Prepare the ship dataset: Select the HRSC2016 ship dataset and complete the conversion of the annotation file format;

[0023] S2. Perform data preprocessing: Perform data augmentation operations such as normalization, random flipping, and random cropping on the images in the dataset;

[0024] S3. Constructing a Fine-Grained Ship Detection Model: The backbone network uses a pre-trained ResNet-50 for initial feature extraction and size reduction, with the output being a feature map. To fully utilize multi-level information, this disclosure proposes a dual-stream feature fusion module. This module employs a top-down and bottom-up fusion strategy to effectively fuse information from various levels, thereby improving the ability to identify ships at multiple scales. Then, a polarization feature focusing module is introduced, applying polarization attention to capture key features relevant to specific tasks. This feature decoupling provides more useful semantic information for each task, thus contributing to accurate classification and regression. Furthermore, an adaptive anchor label allocation strategy is adopted, selecting anchor points with high-quality regression features to compensate for inconsistencies between classification and regression.

[0025] S4. Training the model: The preprocessed image is fed into the model described in S3 for forward propagation, back propagation and loss function calculation, to complete model training and save parameters.

[0026] S5. Ship Inspection: Load the model parameters obtained from training and directly inspect ships at sea.

[0027] Furthermore, the model design specifically includes: First, this disclosure constructs a multi-scale feature pyramid through a dual-stream feature fusion module, fusing the feature maps in both bottom-up and top-down directions. Unlike traditional feature pyramids, the top-down fusion part of this disclosure considers all layers above the current layer. Next, this disclosure proposes a polarization feature focusing module to generate decoupled features sensitive to classification and regression, effectively extracting task-specific key features. Finally, this disclosure introduces an adaptive anchor label allocation strategy to dynamically select and capture high-quality anchors, mitigating inconsistencies between classification and regression, and introduces a harmony compensation factor to construct a harmony-sensitive loss function, enabling the model to focus more on anchors with high potential.

[0028] Furthermore, firstly, the feature maps are uniformly sized using 1x1 convolutions. During the fusion process, this disclosure considers two fusion directions: top-down and bottom-up. For the bottom-up part, the lower-layer feature maps are input into the channel attention module to adjust the weights of each channel, and then fused element-wise with the higher-layer feature maps. Due to the different resolutions of the feature maps, resizing is required during the fusion process; this disclosure uses bilinear interpolation to achieve this step. For the top-down part, the two-stream feature fusion module fuses all layers above the current layer. To better adapt to a wider range of anchor scales, this disclosure performs max pooling on P6 to obtain P7.

[0029] Furthermore, this disclosure constructs an independent feature pyramid network, called a dual FPN, for each task. Then, this disclosure uses a carefully designed polarization attention mechanism to obtain enhanced feature representations. Utilizing the polarization function, different branches generate the discriminative features required for their respective tasks. Specifically, for classification tasks, this disclosure prefers to select global features with high response to reduce the impact of feature interference. For regression tasks, this disclosure focuses more on the features of the target boundary and suppresses the influence of irrelevant highly activated regions.

[0030] Given intermediate features P∈RC*H*W as input, this disclosure uses convolution operations to sequentially infer a 1D channel attention map Mc∈RC*1*1 and a 2D spatial attention map Ms∈R1*H*W from the input features. Channel attention is used to model the relationships between different channels in the input feature map, while spatial attention is used to model the dependencies between pixels in the input image. The formula is as follows:

[0031] (1)

[0032] (2)

[0033] in The sigmoid function is obtained from the input features P through global average pooling, where W0∈R. C / r*C and W1∈R C*C / r is the weight of the fully connected layer, and r is the reduction ratio. 3*3 This represents a 3x3 filter convolution. d 1*3 c d 3*1 c d 3*3 ...

[0034] (3)

[0035] (4)

[0036] in ⊙ represents tensor product, and ⊙ represents element-wise multiplication. Based on this, this disclosure constructs a robust task-sensitive key feature representation by designing specific polarization functions to meet the needs of classification and regression tasks. For classification tasks, this disclosure enhances the high-response portions of the feature map, giving them higher weight in the feature representation, while ignoring less important cue portions that may only be used for localization or introduce interfering noise. Simultaneously, for regression tasks, key features are often scattered around the object edges; this disclosure reduces the response of specific regions, aiming to concentrate the feature map on as many visual cues as possible for object localization, such as object contours and contextual information.

[0037] (5)

[0038] (6)

[0039] Here, the tuning factor controls the intensity of feature activation (default is 15). For key classification features, high-response regions are sufficient for accurate classification, without needing to pursue excessive information. Therefore, this disclosure incentivizes high-response key classification features and suppresses irrelevant features with attention weights less than 0.5. This allows the classifier to reduce its focus on difficult-to-classify regions, enhance sensitivity to key features, and reduce the risk of overfitting and misclassification. Unlike classification tasks, a strong response to a small region at the edge of a target is detrimental to locating the entire target. The suppression function suppresses high-response regions in the regression features, forcing the model to seek potential visual cues for accurate localization.

[0040] (7)

[0041] Finally, a task-sensitive feature representation is constructed according to the above formula. By summing and merging the attention-weighted features, input features P, and attention response map M', this disclosure yields a powerful feature representation F for accurate ship identification. The polarization feature focusing module ensures that the extracted regression key features are evenly distributed on the target, which helps to identify the target boundary and perform accurate localization. The classification key features are concentrated on the most easily identifiable parts of the target, avoiding interference from other parts on the classification results, thus making the classification results more accurate.

[0042] Furthermore, this disclosure introduces the concept of Harmony Score (HS), which utilizes prior information about spatial matching, feature alignment capability, and regression uncertainty to measure the anchor point's localization capability, defined as follows:

[0043] (8)

[0044] (9)

[0045] This disclosure uses and to weight different tasks, and is a penalty term obtained by the change in IoU before and after regression, used to quantify the regression uncertainty of the model during training. IoU_i represents the IoU between the anchor point and the ground truth box, used to represent prior information about spatial alignment. IoU_o, on the other hand, represents the IoU between the predicted box and the ground truth box, used to measure the model's ability to align key features. However, in reality, due to the existence of regression uncertainty, relying solely on IoU_o as a criterion is unreliable. For example, a high-quality anchor with low IoU_i but high IoU_o might be incorrectly classified as a negative sample in the early stages of training. Therefore, to mitigate the impact of regression uncertainty, this disclosure introduces a penalty term to penalize samples with excessively large changes in IoU after regression, ensuring the reasonable selection of training samples. This disclosure adopts a simple and intuitive form for constructing the penalty term, using the change in IoU before and after regression to represent the probability of incorrect anchor evaluation.

[0046] By utilizing the newly defined harmony score, this disclosure enables a more optimized label assignment strategy. During the training phase, this disclosure first calculates the harmony score between the ground truth bounding box and the anchor, then selects anchors with harmony scores higher than a specific threshold (default 0.6) as positive samples, while labeling the remaining anchors as negative samples. To achieve a more stable training process, this disclosure gradually adjusts the influence of IoU_i. The specific adjustment schedule is as follows:

[0047] (20)

[0048] Where t = number of iterations / maximum number of iterations It is the final weighting factor that appears in (13).

[0049] The further technical solution is as follows: In order to enable the detector to treat anchor points with different localization capabilities differently, and to give more attention to positive samples with higher localization potential, this disclosure proposes a harmony compensation factor. For each GT g First, calculate its harmony score with all anchors as hs. Then, select positive samples based on a threshold. The harmony score of the positive samples is represented as hs. pos , of which hs pos ∈hs. Assume the maximum harmonic fraction of g is hs. max The compensation value is denoted as hs, then The harmony compensation factor c is formed by summing the harmony scores of all positive samples:

[0050] (10)

[0051] (11)

[0052] This disclosure embeds a harmony compensation factor into the training process to further enhance the correlation between regression and classification for high-quality ship identification, and proposes a harmony-sensitive loss. The classification-sensitive loss and regression-sensitive loss are defined as follows:

[0053] (12)

[0054] (13)

[0055] in, and N and N0 represent the positive samples for all anchor and harmony score threshold selections, respectively. p These represent the total number of all anchors and the total number of positive anchors, respectively. FL(·) represents the focal loss. L smoothL1 This represents the smoothed L1 loss of the regression. j The harmonic compensation factor c is used to distinguish positive samples with different localization potentials. Introducing the harmonic compensation factor c prevents the loss of high-quality positive samples from being drowned out by the main loss of samples with poor alignment with the ground truth bounding box, allowing the model to adaptively focus on samples with higher localization potential, rather than just good initial spatial alignment. The multi-task loss of the model is defined as follows:

[0056] (14)

[0057] Where the value p and the vector t represent the predicted classification score and the offset of the predicted bounding box, respectively. * The class label representing the anchor (for positive samples, p)* =1; for negative samples p * =0). t and t * These are the offsets of the predicted bounding box and the target bounding box, respectively, t * =(t x * ,t y * ,t w * ,t h * ,t θ * ).

[0058] According to one or more embodiments, such as Figure 1 As shown, a directed fine-grained ship identification method based on key feature extraction from remote sensing images specifically includes the following steps:

[0059] Step S1: Obtain the HRSC2016 remote sensing image dataset for fine-grained ship classification. This includes three tasks: ship detection, coarse-grained ship recognition, and fine-grained ship recognition.

[0060] Step S2: Perform data preprocessing on the input fine-grained remote sensing ship data to be identified:

[0061] Step S21: Convert the dataset annotation grid to DOTA format.

[0062] Step S22: Since the remote sensing images are very large and of inconsistent size, they need to be divided into 1024*1024 images.

[0063] Step S3: The image data preprocessed in Step 2 is layered and fused to generate a preliminary fused multi-scale feature map.

[0064] Step S31: The feature map is first unified in channel dimension through 1*1 convolution.

[0065] Step S32: For the bottom-up part, input the lower-level feature map into the channel attention module to adjust the weight of each channel, and then merge it upwards, adding it element by element to the higher-level feature map (addition operation is always accompanied by resizing operation, which is implemented by bilinear interpolation algorithm).

[0066] Step S33: For the top-down part, unlike the general FPN which only merges two adjacent layers, the bidirectional feature extraction network merges all layers above the current layer.

[0067] Step S4: Send the fused intermediate features generated in step S33 into the polarization feature focusing module, such as... Figure 2For each task, an independent feature pyramid network (dual FPN) is constructed, and the discriminative features required for each task are generated using the polarization function.

[0068] Step S41: Given intermediate features P∈R C*H*W As input, a 1D channel attention map Mc∈R is sequentially inferred from the input features through convolution operations. C*1*1 And 2D spatial attention graph Ms∈R 1*H*W Channel attention is used to model the relationships between different channels in the input feature map, while spatial attention is used to model the dependencies between pixels in the input image.

[0069] (1)

[0070] (2)

[0071] in This represents the sigmoid function. It is obtained from the input feature P through global average pooling, where W0∈R C / r*C and W1∈R C*C / r is the weight of the fully connected layer, and r is the reduction ratio. This represents a 3x3 filter convolution. , , These represent dilated convolutions with different kernel sizes, and the dilation rate is 2. The concatenation of features is represented here. Dilated convolution is used to expand the receptive field of the convolutional kernel. Simultaneously, convolutional kernels with different aspect ratios are used to better detect slender objects like ships. The classification cost is calculated. The ground truth bounding box is obtained from the prediction results.

[0072] Step S42: This disclosure obtains the attention response map M' for a specific task by concatenating two attention maps, as shown in the following formula:

[0073] (3)

[0074] (4)

[0075] in Represents tensor product, This represents element-wise multiplication. Building upon this, a robust task-sensitive key feature representation is constructed by designing specific polarization functions to meet the needs of classification and regression tasks. For classification tasks, high-response portions of the feature map are enhanced, giving them higher weight in the feature representation, while less important cue portions that might only be used for localization or introduce interfering noise are ignored. Simultaneously, for regression tasks, key features are typically scattered around the object edges; reducing the response in specific regions aims to concentrate the feature map on as many visual cues as possible for object localization, such as object contours and contextual information.

[0076] Step S43: This disclosure uses the following activation and suppression functions to process the input features:

[0077] (5)

[0078] (6)

[0079] Here, the tuning factor controls the intensity of feature activation (default is 15). For key classification features, high-response regions are sufficient for accurate classification, without needing to pursue excessive information. Therefore, this disclosure incentivizes high-response key classification features and suppresses irrelevant features with attention weights less than 0.5. This allows the classifier to reduce its focus on difficult-to-classify regions, enhance its sensitivity to key features, and reduce the risk of overfitting and misclassification. Unlike classification tasks, a strong response to a small region at the edge of a target is detrimental to locating the entire target.

[0080] Step S44: A task-sensitive feature representation is constructed according to the following expression. By summing and merging the attention-weighted features, input features P, and attention response map M', this disclosure obtains a powerful feature representation F for accurate ship identification. Planned feature focusing ensures that the extracted regression key features are evenly distributed across the target, aiding in target boundary identification and accurate localization. Classification key features are concentrated on the most easily identifiable parts of the target, avoiding interference from other parts, thus resulting in more accurate classification results.

[0081] (7)

[0082] Step S5: This disclosure introduces the concept of harmony score, which uses prior information on spatial matching, feature alignment capability, and regression uncertainty to measure the anchor point's localization capability, defined as follows:

[0083] (8)

[0084] (9)

[0085] Among them, the use of and To weight different tasks. This is a penalty term, obtained by measuring the change in IoU before and after regression, used to quantify the regression uncertainty of the model during training. IoU_i represents the IoU between the anchor point and the ground truth box, used to represent prior information about spatial alignment. IoU_o, on the other hand, represents the IoU between the predicted box and the ground truth box, used to measure the model's ability to align key features. However, in reality, due to the existence of regression uncertainty, relying solely on IoU_o as a criterion is unreliable. For example, a high-quality anchor with low IoU_i but high IoU_o might be incorrectly classified as a negative sample in the early stages of training. Therefore, to mitigate the impact of regression uncertainty, this disclosure introduces a penalty term. This disclosure penalizes samples with excessively large changes in IoU after regression to ensure reasonable selection of training samples. The penalty term is constructed in a simple and intuitive form, using the change in IoU before and after regression to represent the probability of incorrect anchor evaluation. Utilizing a newly defined harmony score, this disclosure enables a more optimized label assignment strategy. During the training phase, this disclosure first calculates the harmony score between the ground truth bounding box and the anchor, then selects anchors with harmony scores higher than a specific threshold (default 0.6) as positive samples, while labeling the remaining anchors as negative samples.

[0086] Step S6: To enable the detector to differentiate between anchor points with different localization capabilities, and to give more attention to positive samples with higher localization potential, this disclosure proposes a harmony compensation factor. For each GT... g First, calculate its harmony score with all anchors as hs. Then, select positive samples based on a threshold. The harmony score of the positive samples is represented as hs. pos , of which hs pos ∈hs. Assume the maximum harmonic fraction of g is hs. max The compensation value is denoted as hs, then The harmony compensation factor c is formed by summing the harmony scores of all positive samples:

[0087] (10)

[0088] (11)

[0089] By embedding a harmony compensation factor into the training process, the correlation between regression and classification is further enhanced to achieve high-quality ship identification, and a harmony-sensitive loss is proposed. The classification-sensitive loss and regression-sensitive loss are defined as follows:

[0090] (12)

[0091] (13)

[0092] in, and N and N0 represent the positive samples for all anchor and harmony score threshold selections, respectively. p These represent the total number of all anchors and the total number of positive anchors, respectively. FL(·) represents the focal loss. L smoothL1 This represents the smoothed L1 loss of the regression. j The harmonic compensation factor c is used to distinguish positive samples with different localization potentials. Introducing the harmonic compensation factor c prevents the loss of high-quality positive samples from being drowned out by the main loss of samples with poor alignment with the ground truth bounding box, allowing the model to adaptively focus on samples with higher localization potential, rather than just good initial spatial alignment. The multi-task loss of the model is defined as follows:

[0093] (14)

[0094] Where the value p and the vector t represent the predicted classification score and the offset of the predicted bounding box, respectively. * The class label representing the anchor (for positive samples, p) * =1; for negative samples p * =0). t and t * These are the offsets of the predicted bounding box and the target bounding box, respectively, t * =(t x * ,t y * ,t w * ,t h * ,t θ * ).

[0095] In this embodiment of the disclosure, such as Figure 1 The model architecture diagram shown illustrates the following steps: input remote sensing image → ResNet-50 (C2, C3, C4, C5, C6) extracts multi-layer features → dual-stream feature fusion module (F2~F6) → polarization feature focusing module → detector head → output directed ship bounding boxes + fine-grained categories. Among these,

[0096] ResNet-50 serves as the backbone network, outputting five feature maps {C2, C3, C4, C5, C6}. A dual-stream feature fusion module propagates low-level details upwards and high-level semantics downwards, merging all layers above the current one. In the diagram, F2~F6 represent the multi-scale feature pyramid obtained after fusion. Then, a polarization feature focusing module further decouples each level of F2~F6, generating "classification features" and "regression features." By judging whether "hs>0.5?", only anchor points with a harmony score (HS) higher than the threshold are retained as positive samples during training. The detector head receives both classification and regression features, outputting a class score and a 5-dimensional rotated bounding box (x, y, w, h, θ). Therefore, Figure 1 The structure connects "dual-stream fusion, polar decoupling, and adaptive labeling," visually demonstrating the complete forward path from raw remote sensing images to fine-grained directed detection results.

[0097] like Figure 2 The diagram shows the model architecture of the polarization feature focusing module. The model input is the intermediate feature P ∈ RC×H×W after two-stream fusion, and the model outputs two independent branches: a classification branch and a regression branch. The classification branch outputs the polarized key classification feature Fcls (high-response regions are "activated"), and the regression branch outputs the polarized key regression feature Freg (high-response regions are "suppressed," and edge information is enhanced). The module includes a channel attention module, a spatial attention module, and activation and suppression functions to obtain classification and regression features, respectively. Figure 2 The abstract concept of "polarized attention" is unfolded into a pipeline diagram of "dual flow-dual attention-dual polarization," which shows how to "split" the same feature map into two sets of key features that are suitable for classification and localization respectively, thereby solving the problem of "inter-class similarity and intra-class difference" in fine-grained recognition.

[0098] The beneficial effects of the embodiments disclosed herein include:

[0099] (1) This disclosure adopts a top-down and bottom-up fusion strategy to effectively integrate information from all levels, thereby improving the ability to identify ships at multiple scales.

[0100] (2) This disclosure captures key features relevant to a specific task by applying polarized attention. This feature decoupling provides more useful semantic information for each task, thereby helping to make accurate classification and regression.

[0101] (3) This disclosure adopts an adaptive anchor label allocation strategy, selecting anchor points with high-quality regression features to compensate for the inconsistency between classification and regression. In addition, a harmony compensation factor is introduced, and a harmony-sensitive loss is proposed, enabling the detector to pay more attention to anchor points with higher potential.

[0102] It should be understood that in the embodiments of this disclosure, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0103] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this disclosure, and these modifications or substitutions should all be covered within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for directional fine-grained ship identification based on remote sensing images, characterized in that, Ships are identified using a fine-grained ship detection model. The steps for constructing the ship fine-grained detection model include: S1, Obtain the ship identification dataset; S2, preprocessing the ship identification dataset; S3, ResNet-50 is used as the backbone network of the ship fine-grained detection model to extract multi-scale visual features of the ship and output multi-layer feature maps; S4. Input the preprocessed ship identification dataset into the ship fine-grained detection model for forward propagation, back propagation, and loss function calculation to complete model training and save parameters.

2. The method according to claim 1, characterized in that, The ship fine-grained detection model includes a dual-flow feature fusion module, which uses two fusion directions, top-down and bottom-up, to fuse information from feature maps at each layer.

3. The method according to claim 2, characterized in that, The ship fine-grained detection model includes a polarization feature focusing module, which captures key features relevant to a specific task by applying polarization attention.

4. The method according to claim 1, characterized in that, The ship fine-grained detection model adopts an adaptive anchor label allocation strategy to select anchor points with high-quality regression characteristics based on harmony scores.

5. The method according to claim 2, characterized in that, For bottom-up fusion, the dual-stream feature fusion module inputs the lower-level feature map into the channel attention module to adjust the weight of each channel, and then fuses it with the higher-level feature map element by element.

6. The method according to claim 5, characterized in that, For top-down fusion, the dual-stream feature fusion module fuses all layers above the current layer.

7. The method according to claim 1, characterized in that, The preprocessing of the ship identification dataset includes converting the dataset annotation grid to the DOTA format and segmenting it into images of the same size.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method as described in any one of claims 1 to 7.

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

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed by a processor to implement the method according to any one of claims 1 to 7.

Images