Low resolution optical ship identification method and apparatus

Abstract

The application provides a low-resolution optical ship identification method and device, and relates to the technical field of ship identification. The method comprises the following steps: acquiring geographical position information of a to-be-identified ship, a low-resolution ship image containing the to-be-identified ship and imaging angle and imaging time of the ship image, and the ship image has a ship feature; enhancing the ship feature based on a feature enhancement module, performing a deformable convolution operation on the enhanced ship image based on the imaging angle of the ship image, obtaining an angle-calibrated ship feature, and performing pooling processing on the angle-calibrated ship feature to obtain a feature descriptor; generating an environment compensation feature based on the imaging time of the ship image and the geographical position information of the to-be-identified ship; fusing the feature descriptor and the environment compensation feature to obtain a fused feature; and comparing the fused feature with a plurality of reference ship features in a ship feature library to determine a target ship feature meeting a preset reference condition from the plurality of reference ship features.

Description

Technical Field

[0001] This application relates to the field of ship identification technology, and in particular to a low-resolution optical ship identification method and apparatus. Background Technology

[0002] Traditional ship identification technologies often rely on deep learning models (such as residual network classification models) for ship classification and identification. However, in low-resolution optical remote sensing image scenarios, existing methods have the following three problems:

[0003] First, low resolution leads to severe loss of image details, significantly reducing recognition accuracy. Furthermore, super-resolution reconstruction methods are computationally too complex to be deployed on mobile terminals.

[0004] Secondly, the appearance differences caused by changes in ship attitude are huge. Existing methods require storing a massive number of multi-angle templates. When the storage volume is large and the angle deviation is too large, the recognition accuracy drops sharply.

[0005] Third, environmental changes such as lighting conditions, sea surface reflection, and severe weather pose significant challenges, and existing methods lack effective utilization of environmental information, resulting in poor robustness. Summary of the Invention

[0006] In view of the above problems, this application provides a low-resolution optical ship identification method and apparatus.

[0007] This application provides a low-resolution optical ship identification method, comprising: acquiring the geographical location information of the ship to be identified, a low-resolution ship image containing the ship to be identified, and the imaging angle and imaging time of the ship image, wherein the ship image contains ship features associated with the ship to be identified; enhancing the ship features in the ship image based on a preset feature enhancement module, performing a deformable convolution operation on the enhanced ship image based on the imaging angle of the ship image to obtain angle-calibrated ship features, and performing pooling processing on the angle-calibrated ship features to obtain feature descriptors; generating environmental compensation features of the ship image based on the imaging time of the ship image and the geographical location information of the ship to be identified, wherein the environmental compensation features include an offset caused by solar information and seasonal information corresponding to the imaging time; fusing the feature descriptors and the environmental compensation features to obtain fused features; comparing the fused features with multiple reference ship features pre-stored in a ship feature library to determine target ship features that meet preset reference conditions from the multiple reference ship features, wherein the target ship features are used to identify the identification information of the ship to be identified.

[0008] According to an embodiment of this application, ship features in a ship image are enhanced based on a preset feature enhancement module, including: extracting ship features in the ship image that are easily submerged by background noise; and enhancing ship features that are easily submerged by background noise using an attention mechanism; wherein the attention mechanism includes a channel attention mechanism, a spatial attention mechanism, and a combination of a channel attention mechanism and a spatial attention mechanism.

[0009] According to an embodiment of this application, based on the imaging angle of a ship image, a deformable convolution operation is performed on the enhanced ship image to obtain angle-calibrated ship features. This includes: generating a spatial offset corresponding to the imaging angle of the ship image, where the imaging angle includes the roll angle, pitch angle, and yaw angle of the ship to be identified when the ship image is captured; adjusting the sampling position of the convolution kernel in the deformable convolution operation based on the spatial offset to match the spatial orientation of the ship to be identified; and extracting features from the enhanced ship image through the adjusted deformable convolution operation to obtain angle-calibrated ship features.

[0010] According to an embodiment of this application, the ship features after angle calibration are pooled to obtain feature descriptors, including: extracting a basic feature map from the ship features after angle calibration through a preset backbone network; performing generalized average pooling on the basic feature map along the spatial dimension to obtain global features; uniformly dividing the basic feature map into multiple horizontal strips along the height direction; performing generalized average pooling on each horizontal strip to obtain multiple local features; sequentially concatenating the global features and multiple local features to obtain multi-granularity descriptors; and sequentially performing weight standardization and batch normalization on the multi-granularity descriptors to obtain feature descriptors.

[0011] According to an embodiment of this application, the solar information includes illumination information and solar altitude angle information; based on the imaging time of the ship image and the geographical location information of the ship to be identified, environmental compensation features of the ship image are generated, including: determining the solar altitude angle information based on the imaging time and the geographical location information of the ship to be identified; determining illumination information based on the solar altitude angle information; determining seasonal information based on the imaging time; and generating environmental compensation features based on the solar altitude angle information, illumination information, and seasonal information.

[0012] According to an embodiment of this application, the solar altitude angle is determined based on the imaging time and the geographical location information of the ship to be identified, including: calculating the solar declination angle and hour angle using an astronomical algorithm based on the imaging time; obtaining the longitude and latitude of the ship to be identified based on the geographical location information of the ship to be identified; and determining the solar altitude angle based on the solar declination angle, hour angle, longitude, and latitude.

[0013] According to an embodiment of this application, determining seasonal information based on imaging time includes: determining the season type of the imaging time based on the month information in the imaging time; encoding the season type into a ternary season vector, and using the ternary season vector as seasonal information.

[0014] According to an embodiment of this application, the feature descriptor and the environmental compensation feature are fused to obtain the fused feature, including: generating a feature gating vector based on the environmental compensation feature through a preset gating network; and multiplying the feature gating vector with the feature descriptor to obtain the fused feature.

[0015] According to an embodiment of this application, the fused feature is compared with multiple reference ship features pre-stored in a ship feature library to determine the target ship feature that meets the preset reference conditions from the multiple reference ship features, including: calculating the similarity between the fused feature and any one of the multiple reference ship features; when the similarity is greater than a preset similarity threshold, determining any one of the reference ship features as the target ship feature.

[0016] This application, in another aspect, provides a low-resolution optical ship identification device, comprising: an acquisition module for acquiring the geographical location information of a ship to be identified, a low-resolution ship image containing the ship to be identified, and the imaging angle and imaging time of the ship image, wherein the ship image contains ship features associated with the ship to be identified; and an adjustment module for enhancing the ship features in the ship image based on a preset feature enhancement module, performing a deformable convolution operation on the enhanced ship image based on the imaging angle of the ship image to obtain angle-calibrated ship features, and performing pooling processing on the angle-calibrated ship features to obtain feature... The system comprises: a descriptor; a compensation module, used to generate environmental compensation features for the ship image based on the imaging time and the geographical location information of the ship to be identified, the environmental compensation features including the offset caused by solar information and seasonal information corresponding to the imaging time; a fusion module, used to fuse the feature descriptor with the environmental compensation features to obtain fused features; and an identification module, used to compare the fused features with multiple reference ship features pre-stored in a ship feature library to determine the target ship features that meet preset reference conditions from the multiple reference ship features, the target ship features being used to identify the identification information of the ship to be identified.

[0017] The low-resolution optical ship identification method and apparatus provided in this application have at least the following technical advantages:

[0018] 1. This application uses a preset feature enhancement module and channel attention and spatial attention mechanisms to adaptively enhance local structural features (such as island, mast, and deck edges) in ship images that are easily submerged by background noise. Without increasing the image resolution, it effectively restores the response intensity of key discriminative features and overcomes the problem of detail loss caused by low resolution.

[0019] 2. This application adopts a lightweight re-identification feature extraction network, which replaces standard convolution by combining depthwise convolution and pointwise convolution, and adopts generalized average pooling and horizontal strip pooling strategies. While ensuring feature discrimination ability, it significantly reduces the amount of computation and parameters. Compared with traditional super-resolution reconstruction methods, it has significant computational efficiency advantages, can be deployed on mobile terminals and meet real-time processing requirements.

[0020] 3. This application dynamically generates spatial offset based on the ship's imaging angle, and adjusts the sampling position of the convolution kernel based on the spatial offset to match the deformed receptive field with the ship's spatial orientation, thereby adapting to geometric deformation under different viewpoints. Through angle adaptive calibration, the same ship can generate consistent feature representations at different angles, eliminating the need to store massive multi-angle templates, significantly reducing the feature library size, and avoiding a decrease in recognition accuracy due to excessive angle deviation.

[0021] 4. This application employs an imaging time compensation mechanism to calculate solar altitude angle, illumination, color temperature, and seasonal information based on timestamps and geographical location, generating an environmental compensation vector. During the fusion stage, a gating network dynamically modulates ship characteristics to achieve adaptive compensation for illumination, season, and weather. Specifically, monthly information is mapped to a seasonal ternary code to eliminate absolute periodicity interference and improve cross-seasonal robustness. Attached Figure Description

[0022] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0023] Figure 1 A flowchart illustrating a low-resolution optical ship identification method according to an embodiment of this application is shown schematically.

[0024] Figure 2 This schematically illustrates a flowchart of pooling processing of angle-calibrated ship features according to an embodiment of this application;

[0025] Figure 3 A schematic diagram illustrating the structure of a low-resolution optical ship identification device according to an embodiment of this application is shown.

[0026] Figure 4 A block diagram of an electronic device for a low-resolution optical ship identification method according to an embodiment of this application is shown schematically. Detailed Implementation

[0027] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0031] Figure 1 A flowchart illustrating a low-resolution optical ship identification method according to an embodiment of this application is shown.

[0032] like Figure 1 As shown, the low-resolution optical ship identification method in this embodiment includes steps S110 to S150.

[0033] In step S110, the geographical location information of the ship to be identified, a low-resolution ship image containing the ship to be identified, and the imaging angle and imaging time of the ship image are obtained. The ship image contains ship features associated with the ship to be identified.

[0034] For example, the geographical location information of the ship to be identified is the spatial coordinates of its location, including longitude and latitude.

[0035] For example, ship images are optical remote sensing images containing the ships to be identified. The data type is an image matrix. The sources of ship images include optical remote sensing satellites, drones or reconnaissance aircraft, and optoelectronic monitoring equipment.

[0036] The imaging angle is the three-dimensional spatial attitude angle of the ship to be identified relative to the sensor during the shooting, including roll angle, pitch angle, and yaw angle. For example, if the roll angle is 0°, the sea surface is calm and the ship does not roll to the left or right; if the pitch angle is -30°, the UAV is looking down at the ship from a 30° angle above; if the yaw angle is 45°, the angle between the ship's bow direction and the shooting direction is 45°.

[0037] For example, the imaging time can be obtained from the metadata of the ship image, and it exists in the form of a timestamp.

[0038] For example, ship features are structural information in ship images used to distinguish different ships. They exist in the form of feature vectors / feature maps, including low-level (edges / textures), mid-level (parts), and high-level (global semantics) features.

[0039] For example, low-level features include the deck edges and sides of the mast; mid-level features include the mast area and the island area; and high-level features include the overall length-to-width ratio and the ship's outline shape.

[0040] In step S120, based on the preset feature enhancement module, the ship features in the ship image are enhanced. Based on the imaging angle of the ship image, a deformable convolution operation is performed on the enhanced ship image to obtain the angle-calibrated ship features. The angle-calibrated ship features are then pooled to obtain feature descriptors.

[0041] For example, a feature enhancement module is a pre-designed neural network module used to enhance key structural features in an image, such as suppressing background noise and enhancing discriminative regions like the bridge and mast.

[0042] The imaging angle is used to guide deformable convolution for deformation calibration. For example, when the pitch angle is -30°, the vertical direction of the ship to be identified in the ship image is compressed, resulting in a lower height. When the yaw angle is 45°, the ship to be identified in the ship image is in a side view, resulting in a certain offset of the island superstructure.

[0043] Deformable convolution is a convolution operation in which the sampling position of the convolution kernel can be dynamically adjusted to adapt the receptive field to match the actual spatial orientation of the ship. For example, in a deformable convolution operation, the sampling position of the convolution kernel is dynamically shifted according to the imaging angle.

[0044] For example, in the imaging angle, the pitch angle is -30° and the yaw angle is 45°. The offset output by the lightweight convolutional neural network (CNN) sub-network is that the sampling point is offset 2 pixels to the right and 1 pixel upward.

[0045] Pooling is a dimensionality reduction and aggregation operation performed on feature maps, compressing a three-dimensional feature map into a one-dimensional feature vector.

[0046] For example, feature descriptors are used to characterize fixed-dimensional one-dimensional feature vectors that represent ship discrimination information.

[0047] In step S130, based on the imaging time of the ship image and the geographical location information of the ship to be identified, an environmental compensation feature of the ship image is generated. The environmental compensation feature includes the offset caused by the solar information and seasonal information corresponding to the imaging time.

[0048] For example, the imaging time is the specific moment when the ship's image was taken, used to estimate the sun's position in the sky and the current season.

[0049] For example, geographic location information is the spatial coordinates of the location of the ship to be identified, used to determine the orientation of the observation point on the ground plane, and is a key parameter for calculating the solar altitude angle.

[0050] For example, solar information refers to environmental parameters related to the sun, including solar altitude angle and light intensity.

[0051] For example, seasonal information refers to the seasonal category to which a ship image belongs, used to compensate for differences in sea background and atmospheric environment caused by seasonal changes. When it is autumn, the sea surface is darker green and may be accompanied by light fog.

[0052] Environmental compensation features are one-dimensional feature vectors generated by integrating solar and seasonal information to describe imaging environmental conditions.

[0053] In step S140, the feature descriptor and the environmental compensation feature are fused to obtain the fused feature.

[0054] When using feature descriptors alone, environmental changes can cause feature drift for the same ship, affecting recognition accuracy. To address this issue, environmental compensation features are fused with feature descriptors, allowing environmental compensation information to guide the modulation of feature descriptors and achieve environmental adaptation.

[0055] For example, fusion is the operation of combining feature vectors from two or more different sources into a new feature vector. Fusion can be performed by feature concatenation, weighted summation, or through neural network layers.

[0056] In step S150, the fused features are compared with multiple reference ship features pre-stored in the ship feature library to determine the target ship features that meet the preset reference conditions from the multiple reference ship features. The target ship features are used to identify the identification information of the ship to be identified.

[0057] The fused features include ship structure information and environment adaptation information.

[0058] The ship feature database stores reference feature vectors for known ships, with each record corresponding to a known ship. Reference ship features are feature vectors stored in the ship feature database and used as comparison benchmarks.

[0059] The target ship features are reference features selected from the ship feature library that best match the fused features.

[0060] For example, when comparing fused features and reference ship features, pre-similarity or Euclidean distance can be used for comparison.

[0061] Based on the low-resolution optical ship recognition method provided in this application, real-time robust recognition of low-resolution ship images is achieved on a mobile terminal through a triple mechanism of feature enhancement, angle adaptive calibration, and environmental compensation. Specifically, the attention mechanism strengthens key local structures, overcoming detail loss without requiring super-resolution; deformable convolution dynamically adjusts the receptive field according to the imaging angle, ensuring consistent features of the same ship from different viewpoints, eliminating the need to store multi-angle templates; and illumination and seasonal information are inferred based on imaging time and location, adaptively enhancing effective features and suppressing interference in adverse environments such as backlighting and fog. Finally, the fused features after environmental compensation are compared with the ship feature database, significantly improving the recognition accuracy and robustness across viewpoints, seasons, and complex environments.

[0062] According to an embodiment of this application, ship features in a ship image are enhanced based on a preset feature enhancement module, including: extracting ship features in the ship image that are easily submerged by background noise; and enhancing ship features that are easily submerged by background noise using an attention mechanism; wherein the attention mechanism includes a channel attention mechanism, a spatial attention mechanism, and a combination of a channel attention mechanism and a spatial attention mechanism.

[0063] Background noise can be interference information in a ship image that is unrelated to the ship to be identified, such as sea waves, clouds, and light reflection. Local structural features of the ship may become blurred and difficult to identify due to low resolution or environmental interference, such as mast edges, island portholes, and deck edges.

[0064] Attention mechanisms are a technique that allows neural networks to automatically focus on important regions or channels and ignore irrelevant information, used to dynamically allocate computing resources and focus on key features.

[0065] For example, channel attention mechanisms weight each channel of the feature map, enhancing important channels and suppressing irrelevant channels; spatial attention mechanisms weight the spatial location of the feature map, focusing on important regions and suppressing background regions; the combination of the two is to use channel attention and spatial attention in series or in parallel.

[0066] For example, enhancing ship features in ship images can be achieved through the following methods.

[0067] First, a primary feature map F is extracted from low-resolution ship images using a lightweight mobile convolutional neural network model. , where F is the local structural feature, R is the image, and H, W, and C are the height, width, and number of channels of the primary feature map, respectively.

[0068] Second, a channel-space dual-path attention mechanism is used to enhance features and suppress the sea surface background.

[0069] The channel attention mechanism module enhances feature channels related to local structure (such as texture and edge channels) by performing global average pooling (GAP) and global max pooling (GMP) on the primary feature map F to generate channel description vectors.

[0070]

[0071]

[0072] In the formula, This is the channel description vector for GAP. This is the channel description vector for GMP.

[0073] Channel weights are generated through a shared fully connected layer (including dimensionality reduction and dimensionality expansion). :

[0074]

[0075] In the formula, It is the sigmoid activation function. is the ReLU activation function, and FC is a fully connected layer.

[0076] Channel weighted features for:

[0077]

[0078] The spatial attention mechanism module aims to acquire the precise spatial location of local structural features, highlighting local structural regions of the ship while de-emphasizing smooth regions with low information content. (Along the channel axis...) Perform max pooling and average pooling, and concatenate them into a two-channel feature map. :

[0079]

[0080] In the formula, concat means concatenation. They are respectively for Perform max pooling and average pooling.

[0081] Spatial weights are generated using a 7×7 convolutional layer. :

[0082]

[0083] In the formula, It is the sigmoid activation function. It is a 7×7 convolutional layer.

[0084] Spatial weighted features are :

[0085]

[0086] To prevent the attention mechanism module from suppressing the original effective information, a residual structure is introduced:

[0087]

[0088] In the formula, This is the final enhanced feature map. F represents the feature map after processing by the attention mechanism, and F is the primary feature map. The weights are learnable and have an initial value of approximately 0.2.

[0089] Based on the low-resolution optical ship recognition method provided in this application embodiment, a preset feature enhancement module is used to adaptively enhance local structural features (such as island, mast, and deck edge) in ship images that are easily submerged by background noise through channel attention and spatial attention mechanisms. Without increasing the image resolution, the response intensity of key discriminative features is effectively restored, overcoming the problem of detail loss caused by low resolution.

[0090] According to an embodiment of this application, based on the imaging angle of a ship image, a deformable convolution operation is performed on the enhanced ship image to obtain angle-calibrated ship features. This includes: generating a spatial offset corresponding to the imaging angle of the ship image, where the imaging angle includes the roll angle, pitch angle, and yaw angle of the ship to be identified when the ship image is captured; adjusting the sampling position of the convolution kernel in the deformable convolution operation based on the spatial offset to match the spatial orientation of the ship to be identified; and extracting features from the enhanced ship image through the adjusted deformable convolution operation to obtain angle-calibrated ship features.

[0091] For example, roll angle is the angle of rotation of a ship about its own longitudinal axis (the line connecting the bow and stern); pitch angle is the angle of rotation of a ship about its transverse axis (the line connecting the port and starboard sides) or the angle of view of a sensor relative to the ship; yaw angle is the horizontal angle between the ship's heading and the sensor's line of sight.

[0092] For example, the receptive field is the size of the input image region corresponding to a single location on the feature map.

[0093] For example, based on the imaging angle of a ship image, a spatial offset corresponding to the imaging angle is generated, specifically using the following formula:

[0094]

[0095] In the formula, This is the spatial offset. These represent the roll angle, pitch angle, and yaw angle, respectively, and g is a lightweight CNN subnetwork consisting of two 3×3 convolutional layers and a ReLU activation function.

[0096] For example, the specific formula used in deformable convolution operations is:

[0097]

[0098] In the formula, Y is the output feature map after angle calibration. To output the current position coordinates of the feature map, Let X be the convolution kernel weights for the k-th sampling point, and let X be the final enhanced feature map. Predefined sampling offsets for standard convolution. This is the dynamically learned spatial offset.

[0099] The low-resolution optical ship recognition method provided in this application dynamically generates a spatial offset based on the ship's imaging angle, and adjusts the sampling position of the convolution kernel based on the spatial offset to match the deformed receptive field with the ship's spatial orientation, thereby adapting to geometric deformation under different viewpoints. Through angle adaptive calibration, the same ship can generate consistent feature representations at different angles, eliminating the need to store massive multi-angle templates, significantly reducing the feature library size, and avoiding a decrease in recognition accuracy due to excessive angle deviation.

[0100] Figure 2 The flowchart illustrating the pooling process of angle-calibrated ship features according to an embodiment of this application is shown.

[0101] like Figure 2 As shown, in some embodiments, pooling the angle-calibrated ship features includes steps S210 to S260.

[0102] In step S210, basic feature maps are extracted from the angle-calibrated ship features through a preset backbone network.

[0103] For example, the ship features after angle calibration are feature maps that have undergone deformable convolution calibration and whose spatial deformation has been corrected.

[0104] For example, the backbone network can be a pre-designed lightweight convolutional neural network, such as a depthwise separable convolutional network.

[0105] In step S220, the basic feature map is subjected to generalized average pooling along the spatial dimension to obtain global features.

[0106] In step S230, the basic feature map is uniformly divided into multiple horizontal strips along the height direction.

[0107] For example, horizontal stripes are multiple horizontal regions that are evenly divided along the height direction of the basic feature map.

[0108] In step S240, generalized average pooling is performed on each horizontal strip to obtain multiple local features.

[0109] In step S250, the global features and multiple local features are concatenated sequentially to obtain a multi-granularity descriptor.

[0110] For example, concatenation is the process of joining multiple vectors end-to-end along the channel dimension to form a new vector.

[0111] In step S260, the multi-granularity descriptors are sequentially weighted and batch normalized to obtain feature descriptors.

[0112] Batch normalization (BN) standardizes features so that the distribution has a mean of 0 and a variance of 1. Weight standardization (WS) normalizes the weight matrix, constraining the norm of the weights.

[0113] For example, pooling is performed on the ship features after angle calibration to obtain feature descriptors, which specifically include:

[0114] First, a depthwise separable convolutional network is used to construct the backbone network:

[0115]

[0116] In the formula, As a backbone network, For depthwise convolution, This is a pointwise convolution.

[0117] Second, let the input image be... Lightweight convolutional backbone network Extracting basic feature maps :

[0118]

[0119]

[0120] In the formula, Here are the network parameters, and H and W are the feature maps, respectively. The height and width are given by S, where S is the downsampling step size, and h, w, and c are the height, width, and number of channels of the base feature map, respectively.

[0121] Third, the basic feature map Generalized average pooling is applied along the spatial dimension to obtain the global feature vector. :

[0122]

[0123] In the formula, , For a set of spatial coordinates, Indicates the position of the k-th channel. The value of , where P is the pooling parameter and P>0. It represents a c-dimensional real vector.

[0124] Fourth, the basic feature map Divide the area into m horizontal strips evenly along the vertical direction:

[0125]

[0126] In the formula, This represents the k-th horizontal stripe. This represents the starting row index of the k-th horizontal stripe. This represents the index of the terminating row of the k-th horizontal stripe; in middle, The first colon indicates the selection range along the height direction, the second colon indicates the entire selection along the width direction, and the third colon indicates the selection of all channels.

[0127] Perform generalized average pooling (GeM) independently on each horizontal strip to obtain local feature vectors. :

[0128]

[0129] global feature vectors and local feature vectors Concatenate into multi-granularity descriptors :

[0130]

[0131] In the formula, This represents the first local eigenvector. Represents the m-th local feature vector, a multi-granularity descriptor. Shape A one-dimensional vector, c, is the basic feature map. The number of channels.

[0132] Fifth, the robustness of multi-granularity descriptors is improved through batch normalization and weight standardization to obtain feature descriptors. :

[0133]

[0134] In the formula, BN represents batch normalization, and W represents weight standardization, which satisfies the weight standardization constraint. b is the bias vector.

[0135] For example, a lightweight ship re-identification (ReID) feature extraction network is used to generate feature descriptors. A triple-supervision mechanism is employed for joint optimization in the loss function of the lightweight ReID feature extraction network, as detailed below:

[0136] First, cross-entropy loss (ID Loss) Used to enhance category discrimination capabilities:

[0137]

[0138] In the formula, C represents the total number of ship categories. The true class label of the sample is given, and softmax is the normalization exponential function. This is the weight matrix for the classification layer.

[0139] Second, Triplet Loss. Used to increase the distance between classes:

[0140]

[0141] In the formula, The feature vector of the anchor sample is the feature of the ship to be identified. The feature vector of a positive sample is the feature vector of a ship of the same type as the anchor sample. The feature vector of the negative sample is the feature vector of the ship that is different from the anchor sample. This represents the boundary margin.

[0142] Third, Center Loss. Used to reduce intra-class discrepancies:

[0143]

[0144] In the formula, y is the true class label of the current sample. Let y be the category center vector of category y.

[0145] In summary, the total loss function for:

[0146]

[0147] In the formula, For cross-entropy loss weights, For triple loss weights, The loss weight is centered.

[0148] The low-resolution optical ship recognition method provided in this application combines global GeM pooling with horizontal strip GeM pooling to capture local structural features such as masts, islands, and decks while preserving the overall semantics of the ship. It adaptively adjusts the pooling intensity using a learnable parameter p to enhance the response of salient features at low resolution and suppress noise. Independent horizontal strip pooling makes the descriptors naturally robust to local occlusion and image truncation. Dual normalization, employing batch normalization and weight normalization, eliminates feature amplitude differences and improves cross-environment generalization capabilities. A lightweight design based on depthwise separable convolution meets the real-time inference requirements of mobile terminals.

[0149] According to an embodiment of this application, the solar information includes illumination information and solar altitude angle information; based on the imaging time of the ship image and the geographical location information of the ship to be identified, environmental compensation features of the ship image are generated, including: determining the solar altitude angle information based on the imaging time and the geographical location information of the ship to be identified; determining illumination information based on the solar altitude angle information; determining seasonal information based on the imaging time; and generating environmental compensation features based on the solar altitude angle information, illumination information, and seasonal information.

[0150] For example, the solar altitude angle is the angle between the sunlight and the horizontal plane at the ship's location. Illumination information is the illumination intensity calculated from the solar altitude angle. Seasonal information is the seasonal category code inferred from the month information in the imaging time.

[0151] For example, environmental compensation features are feature vectors generated by integrating solar altitude angle, illumination information, and seasonal information.

[0152] For example, based on the angle information of the solar altitude angle Classification of light levels As shown below:

[0153]

[0154] In the formula, when the solar altitude angle is less than 0, the illumination level is dark; when the solar altitude angle is between 0 and 15, the illumination level is low; when the solar altitude angle is between 15 and 45, the illumination level is medium; and when the solar altitude angle is greater than 45, the illumination level is high.

[0155] Lighting information is obtained by vectorizing the lighting levels. :

[0156]

[0157] For example, environmental compensation characteristics for:

[0158]

[0159] In the formula, For illumination information, Here, W represents seasonal information, W is the projection matrix, b is the bias term, and concat represents concatenation.

[0160] Based on the low-resolution optical ship recognition method provided in this application embodiment, an imaging time compensation mechanism is used to calculate solar altitude angle, illumination, color temperature, and seasonal information based on timestamps and geographical locations, generating an environmental compensation vector. During the fusion stage, ship features are dynamically modulated through a gating network to achieve adaptive compensation for illumination, season, and weather. Specifically, monthly information is mapped to a seasonal ternary code to eliminate absolute periodicity interference and improve cross-seasonal robustness.

[0161] According to an embodiment of this application, the solar altitude angle is determined based on the imaging time and the geographical location information of the ship to be identified, including: calculating the solar declination angle and hour angle using an astronomical algorithm based on the imaging time; obtaining the longitude and latitude of the ship to be identified based on the geographical location information of the ship to be identified; and determining the solar altitude angle based on the solar declination angle, hour angle, longitude, and latitude.

[0162] The solar declination angle is the angle between the sun's rays and the Earth's equatorial plane.

[0163] The hour angle is the angle between the meridian of the observation point and the solar meridian.

[0164] For example, the angle information for determining the solar altitude angle is based on the solar declination angle, hour angle, longitude, and latitude. The following formula is used for calculation:

[0165]

[0166]

[0167]

[0168] In the formula, lat is latitude, lon is longitude, and H is hour angle. The solar declination angle is given by , UTC is Universal Time, and N is the day of the year in which the image was formed.

[0169] Based on the low-resolution optical ship identification method provided in the embodiments of this application, the accurate solar altitude angle provides a reliable input for subsequent illumination intensity, enabling environmental compensation features to effectively reflect actual illumination conditions and significantly improving the robustness of ship identification in complex scenarios such as changes in dawn and dusk, seasonal changes, and different latitudes.

[0170] According to an embodiment of this application, determining seasonal information based on imaging time includes: determining the season type of the imaging time based on the month information in the imaging time; encoding the season type into a ternary season vector, and using the ternary season vector as seasonal information.

[0171] For example, encoding season types as ternary season vectors The following formula is used for calculation:

[0172]

[0173] In the formula, m represents the month.

[0174] Based on the low-resolution optical ship recognition method provided in this application embodiment, discrete months are mapped to four seasonal categories: spring, summer, autumn, and winter. This eliminates the interference of the absolute periodicity of months on feature extraction and avoids the neural network from incorrectly associating months with similar values ​​but different seasons (such as November and December) or months with large differences in values ​​but the same season (such as December and January).

[0175] According to an embodiment of this application, the feature descriptor and the environmental compensation feature are fused to obtain the fused feature, including: generating a feature gating vector based on the environmental compensation feature through a preset gating network; and multiplying the feature gating vector with the feature descriptor to obtain the fused feature.

[0176] For example, the pre-defined gating network is a pre-designed lightweight neural network used to map environmental compensation features into gating vectors.

[0177] For example, the feature gating vector is a modulation vector generated from environmental compensation features.

[0178] For example, multiplying the feature gating vector by the feature descriptor yields the fused features, specifically using the following formula:

[0179]

[0180]

[0181] In the formula, , Here, are the gate weight matrix and the bias term, respectively, and g is the feature gate vector. The function is , As a feature of fusion, For feature descriptors, This is a characteristic of environmental compensation.

[0182] Based on the low-resolution optical ship recognition method provided in this application embodiment, the environmental compensation features are dynamically mapped to feature gating vectors through a gating network, and adaptive fusion is achieved by multiplying them element by element with the ship feature descriptor. This enables adaptive soft selection of each channel of the feature descriptor, enhances effective features and suppresses interference features in harsh environments, and does not increase feature dimensions or computational overhead, thus significantly improving the robustness of ship recognition under complex and variable conditions.

[0183] According to an embodiment of this application, the fused feature is compared with multiple reference ship features pre-stored in a ship feature library to determine the target ship feature that meets the preset reference conditions from the multiple reference ship features, including: calculating the similarity between the fused feature and any one of the multiple reference ship features; when the similarity is greater than a preset similarity threshold, determining any one of the reference ship features as the target ship feature.

[0184] For example, similarity is a numerical value that measures the degree of similarity between fused features and reference ship features. A preset similarity threshold is a pre-defined critical value for determining whether a match exists.

[0185] For example, similarity calculations can use cosine similarity or Euclidean distance.

[0186] For example, similarity is calculated using the following formula:

[0187]

[0188] In the formula, To obtain a similarity score with the i-th ship, Let i be the reference ship characteristics of the i-th ship. This is a feature of fusion.

[0189] For example, when similarity If the similarity is greater than 0.6, the ship to be identified is determined to be the type of the i-th ship; when the similarity is greater than 0.6, the ship to be identified is determined to be the type of the i-th ship. If the value is less than 0.6, the ship to be identified will be marked as an unknown ship.

[0190] The low-resolution optical ship recognition method provided in this application embodiment achieves flexible matching decisions by setting a preset similarity threshold. The threshold can be adjusted according to task requirements to adapt to different application scenarios. Based on lightweight feature descriptor row similarity calculation, the amount of computation for a single comparison is extremely low, supporting fast retrieval of large-scale feature databases. The entire comparison process does not require re-extraction of features or complex calculations, meeting the real-time recognition requirements of mobile terminals.

[0191] Figure 3 A schematic block diagram of a low-resolution optical ship identification device according to an embodiment of this application is shown.

[0192] like Figure 3 As shown, the low-resolution optical ship identification device 300 of this embodiment includes an acquisition module 310, an adjustment module 320, a compensation module 330, a fusion module 340, and an identification module 350.

[0193] The acquisition module 310 is used to acquire the geographical location information of the ship to be identified, a low-resolution image of the ship containing the ship to be identified, and the imaging angle and imaging time of the ship image. The ship image contains ship features associated with the ship to be identified. In one embodiment, the acquisition module 310 can be used to perform step S110 described above, which will not be repeated here.

[0194] The adjustment module 320 is used to enhance the ship features in the ship image based on a preset feature enhancement module. Based on the imaging angle of the ship image, a deformable convolution operation is performed on the enhanced ship image to obtain angle-calibrated ship features. The angle-calibrated ship features are then pooled to obtain feature descriptors. In one embodiment, the adjustment module 320 can be used to execute step S120 described above, which will not be repeated here.

[0195] The compensation module 330 is used to generate environmental compensation features for the ship image based on the imaging time and the geographical location information of the ship to be identified. The environmental compensation features include an offset caused by solar information and seasonal information corresponding to the imaging time. In one embodiment, the compensation module 330 can be used to perform step S130 described above, which will not be repeated here.

[0196] The fusion module 340 is used to fuse the feature descriptor with the environmental compensation feature to obtain the fused feature. In one embodiment, the fusion module 340 can be used to perform step S140 described above, which will not be repeated here.

[0197] The identification module 350 compares the fused features with multiple reference ship features pre-stored in a ship feature library to determine the target ship features that meet preset reference conditions from the multiple reference ship features. The target ship features are used to identify the identification information of the ship to be identified. In one embodiment, the identification module 350 can be used to perform step S150 described above, which will not be repeated here.

[0198] According to embodiments of this application, any multiple modules among the acquisition module 310, adjustment module 320, compensation module 330, fusion module 340, and identification module 350 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the acquisition module 310, adjustment module 320, compensation module 330, fusion module 340, and identification module 350 can be at least partially implemented as hardware circuitry, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or implemented in hardware or firmware by any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the acquisition module 310, adjustment module 320, compensation module 330, fusion module 340 and identification module 350 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0199] Figure 4 A block diagram of an electronic device for a low-resolution optical ship identification method according to an embodiment of this application is shown schematically.

[0200] like Figure 4 As shown, an electronic device 400 according to an embodiment of this application includes a processor 401, which can perform various appropriate actions and processes according to a program stored in a read-only memory 402 or a program loaded from a storage portion 408 into a random access memory 403. The processor 401 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 401 may also include onboard memory for caching purposes. The processor 401 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0201] Random access memory 403 stores various programs and data required for the operation of electronic device 400. Processor 401, read-only memory 402, and random access memory 403 are interconnected via bus 404. Processor 401 executes various steps of the method flow according to embodiments of this application by executing programs stored in read-only memory 402 and / or random access memory 403. It should be noted that the programs may also be stored in one or more memories other than read-only memory 402 and random access memory 403. Processor 401 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0202] According to embodiments of this application, the electronic device 400 may further include an input / output interface 405, which is also connected to a bus 404. The electronic device 400 may also include one or more of the following components connected to the input / output interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card, such as a local area network card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to the input / output interface 405 as needed. A removable medium 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 410 as needed so that computer programs read from it can be installed into the storage section 408 as needed.

[0203] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0204] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 402 described above, and / or random access memory 403, and / or one or more memories other than read-only memory 402 and random access memory 403.

[0205] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined or combined in various ways without departing from the spirit and teachings of this application. All such combinations or combinations fall within the scope of this application.

[0206] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A low-resolution optical ship identification method, characterized in that, include: The system acquires the geographical location information of the ship to be identified, a low-resolution image of the ship to be identified, and the imaging angle and imaging time of the ship image, wherein the ship image contains ship features associated with the ship to be identified. Based on the preset feature enhancement module, the ship features in the ship image are enhanced. Based on the imaging angle of the ship image, a deformable convolution operation is performed on the enhanced ship image to obtain the angle-calibrated ship features. The angle-calibrated ship features are then pooled to obtain feature descriptors. Based on the imaging time of the ship image and the geographical location information of the ship to be identified, an environmental compensation feature of the ship image is generated. The environmental compensation feature includes an offset caused by solar information and seasonal information corresponding to the imaging time. The feature descriptor is fused with the environmental compensation feature to obtain the fused feature; The fused features are compared with multiple reference ship features pre-stored in a ship feature library to determine target ship features that meet preset reference conditions from the multiple reference ship features. The target ship features are used to identify the identification information of the ship to be identified.

2. The method according to claim 1, characterized in that, The preset feature enhancement module enhances the ship features in the ship image, including: Extract ship features from the ship images that are easily obscured by background noise; An attention mechanism is employed to enhance the ship features that are easily drowned out by background noise; The attention mechanism includes a channel attention mechanism, a spatial attention mechanism, and a combination of the channel attention mechanism and the spatial attention mechanism.

3. The method according to claim 1, characterized in that, The process of performing a deformable convolution operation on the enhanced ship image based on the imaging angle of the ship image to obtain angle-calibrated ship features includes: Based on the imaging angle of the ship image, a spatial offset corresponding to the imaging angle is generated. The imaging angle includes the roll angle, pitch angle and yaw angle of the ship to be identified when the ship image is being captured. Based on the spatial offset, the sampling position of the convolution kernel in the deformable convolution operation is adjusted so that the convolution receptive field matches the spatial orientation of the ship to be identified. The enhanced ship image is subjected to feature extraction using an adjusted deformable convolution operation to obtain the angle-calibrated ship features.

4. The method according to claim 1, characterized in that, The process of pooling the angle-calibrated ship features to obtain feature descriptors includes: The basic feature map is extracted from the ship features after angle calibration through a preset backbone network; The basic feature map is subjected to generalized average pooling along the spatial dimension to obtain global features; The basic feature map is evenly divided into multiple horizontal stripes along the height direction; Generalized average pooling is applied to each horizontal strip to obtain multiple local features; The global features and the multiple local features are concatenated sequentially to obtain a multi-granularity descriptor; The multi-granularity descriptors are sequentially weighted and batch normalized to obtain the feature descriptors.

5. The method according to claim 1, characterized in that, The solar information includes illumination information and solar altitude angle information; The step of generating environmental compensation features for the ship image based on the imaging time of the ship image and the geographical location information of the ship to be identified includes: Based on the imaging time and the geographical location information of the ship to be identified, the angle information of the solar altitude angle is determined; The illumination information is determined based on the solar altitude angle information; The seasonal information is determined based on the imaging time; The environmental compensation feature is generated based on the solar altitude angle information, the illumination information, and the seasonal information.

6. The method according to claim 5, characterized in that, Based on the imaging time and the geographical location information of the ship to be identified, the angle information of the solar altitude angle is determined, including: Based on the imaging time, the solar declination angle and hour angle are calculated using astronomical algorithms; Based on the geographical location information of the ship to be identified, the longitude and latitude of the ship to be identified are obtained; The solar altitude angle is determined based on the solar declination angle, the hour angle, the longitude, and the latitude.

7. The method according to claim 5, characterized in that, The determination of seasonal information based on the imaging time includes: Based on the month information in the imaging time, the season type of the imaging time is determined; The seasonal type is encoded as a ternary seasonal vector, and the ternary seasonal vector is used as the seasonal information.

8. The method according to claim 1, characterized in that, The feature descriptor is fused with the environmental compensation feature to obtain the fused feature, including: Based on the aforementioned environmental compensation features, a feature gating vector is generated through a preset gating network; The fused feature is obtained by multiplying the feature gating vector by the feature descriptor.

9. The method according to claim 1, characterized in that, The step of comparing the fused features with multiple reference ship features pre-stored in a ship feature library to determine the target ship features that meet preset reference conditions from the multiple reference ship features includes: Calculate the similarity between the fused feature and any one of the plurality of reference ship features; When the similarity is greater than a preset similarity threshold, any reference ship feature is determined to be the target ship feature.

10. A low-resolution optical ship identification device, characterized in that, include: The acquisition module is used to acquire the geographical location information of the ship to be identified, a low-resolution ship image containing the ship to be identified, and the imaging angle and imaging time of the ship image, wherein the ship image has ship features associated with the ship to be identified. The adjustment module is used to enhance the ship features in the ship image based on the preset feature enhancement module, perform deformable convolution operation on the enhanced ship image based on the imaging angle of the ship image to obtain the angle-calibrated ship features, and perform pooling processing on the angle-calibrated ship features to obtain feature descriptors. The compensation module is used to generate environmental compensation features of the ship image based on the imaging time of the ship image and the geographical location information of the ship to be identified. The environmental compensation features include an offset caused by solar information and seasonal information corresponding to the imaging time. The fusion module is used to fuse the feature descriptor with the environmental compensation feature to obtain the fused feature; The identification module is used to compare the fused features with multiple reference ship features pre-stored in a ship feature library to determine the target ship features that meet preset reference conditions from the multiple reference ship features. The target ship features are used to identify the identification information of the ship to be identified.

Images