Artificial intelligence-based ship identity intelligent identification method and system

Abstract

The application provides an intelligent ship identity recognition method and system based on artificial intelligence, and the method comprises the following steps: training original samples through structure constraint enhancement, and constructing a model conforming to real maritime imaging conditions; generating stable target sequences by using joint segmentation and identification bearing area guide mechanism on input videos; continuously updating the frames by tracking feature extraction model in a global coordinate system to realize target tracking; detecting texts on the tracking sequences, and suppressing single-frame errors by using the probability fusion of the same character positions in multiple frames; and finally, matching the recognition results with an identity database index to complete verification. The application significantly improves the accuracy and stability of ship identity recognition in complex environments, and is mainly used in port monitoring, maritime supervision and ship automatic identification scenes.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, and in particular relates to a method and system for intelligent identification of ship identity based on artificial intelligence. Background Technology

[0002] With the increasing demands for port supervision, waterway monitoring, and maritime traffic safety management, image- or video-based ship identification technology has become an important component of intelligent maritime supervision systems. Existing technologies typically use visible markings such as ship names and numbers on the hull surface as the primary identification targets, employing methods such as target detection, text localization, and character recognition to determine ship identity. However, in practical applications, ships are often located in open waters, port shorelines, or complex waterways. The imaging process is easily affected by factors such as changing viewing angles, ship rolling, sea surface reflection, day-night lighting differences, rain and fog, partial obstruction, and fluctuations in camera resolution. This leads to unstable hull region extraction, drift of the marking area, discrete text detection results, and amplified character recognition errors. Especially in continuous video scenarios, existing solutions mostly treat hull segmentation, target tracking, and character recognition as independent processing steps, lacking unified constraints around the ship name marking area and temporal continuity. This causes positional deviations from previous steps to accumulate in subsequent processing, ultimately resulting in inaccurate ship name region extraction, significant fluctuations in single-frame recognition results, incomplete recognition of multilingual or multi-line characters, and insufficient reliability of identity matching. Meanwhile, although some methods introduce image enhancement or deep learning models, their enhancement process often does not constrain the structural stability of ship name characters. There is still a significant gap between the training samples and real ship scenarios, making it difficult to guarantee the long-term stable application of the model in complex maritime environments. Therefore, there is an urgent need for a ship identity intelligent recognition scheme that can simultaneously take into account the quality of offline modeling and the continuity of online recognition. Summary of the Invention

[0003] This invention discloses a method and system for intelligent identification of ship identities based on artificial intelligence, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the first aspect of the present invention provides a ship identification intelligent recognition method based on artificial intelligence, the method comprising: S1: Perform structural constraint enhancement on the original ship image samples to construct training samples that conform to the real maritime imaging conditions. Based on this, complete the joint training of the ship hull region segmentation model, the tracking feature extraction model, and the text detection and character recognition model. At the same time, establish an identity database index corresponding to the standard ship name. S2: For the input video sequence to be identified, the joint segmentation model is used to extract the ship region, and a marker-carrying region guidance and temporal response stabilization mechanism is introduced to generate a stable target image sequence; S3: Use the tracking feature extraction model to update the continuous bounding boxes in the original frame global coordinate system of the target image sequence to achieve target tracking; S4: Perform text region detection on the tracked image sequence, and use probability fusion of the same character position in multiple frames to suppress single-frame errors and generate fusion recognition results; The fused identification results are matched with the identity database index to complete the ship's identity verification.

[0005] Furthermore, the structural constraint enhancement of the original ship image samples includes: Calculate the geometric arrangement structure description value of the ship name characters in the original image; Perform enhancement operations consistent with the actual scene, including small-angle rotation, perspective stretching, local blurring, local brightness enhancement, and overall brightness compression; Simultaneously apply the same geometric transformation to the character annotation box and the ship area box; Calculate the structural description value of the enhanced image and compare the difference with the original structural description value; If the difference is less than the preset allowable range for structural changes, the enhanced sample is retained; otherwise, it is discarded.

[0006] Furthermore, the introduced identifier-bearing area guidance and timing response stabilization mechanism includes: The joint segmentation model is used to simultaneously output a probability map of the ship area and a probability map of the identified carrying area. The joint response map is generated by weighting the probability map of the ship area based on the probability map of the identification carrying area; The joint response map of the previous frame, after center displacement alignment, is fused with the joint response map of the current frame, and a difference suppression term is added to generate a time-consistent response. The main body region of the ship is extracted based on the time-consistent response, and the minimum bounding rectangle is calculated to generate the target image sequence.

[0007] Furthermore, the step of updating the continuous bounding boxes in the original frame global coordinate system of the target image sequence using the tracking feature extraction model includes: The smallest bounding rectangle of the first frame is used as the initial target box; Starting from the second frame, candidate boxes are generated in the original frame global coordinate system of the current frame, using the target box updated in the previous frame as the reference box. The feature vectors of the reference box and the candidate box are extracted using a tracking feature extraction model, and the appearance similarity is calculated. A comprehensive score is generated by combining appearance similarity, mask fitting ratio, and spatial displacement constraints. The candidate bounding box with the highest comprehensive score is selected and a smooth update based on time series is performed to generate the target bounding box for the current frame.

[0008] Furthermore, the tracking feature extraction model consists of a three-layer convolutional structure. The first layer is used to extract the hull edge and waterline information, the second layer is used to extract the side texture and local structure, and the third layer is used to extract the overall hull outline and orientation information.

[0009] Furthermore, the method of suppressing single-frame errors by utilizing probabilistic fusion at the same character position across multiple frames includes: The tracked image sequence is input into the text detection and character recognition model, which outputs the character probability distribution sequence at each character position in each frame. The weights of each frame are determined based on the trajectory parameters of the target bounding box in the global coordinate system of the original frame, with larger weights for smaller displacements. The probability of characters at the same position in each frame is weighted and superimposed to generate a fusion recognition result.

[0010] Furthermore, the matching of the fused recognition result with the identity database index includes: Calculate the edit distance between the recognition result and the standard ship names in the database; Calculate the average fusion probability of the selected character at each position in the recognition result; An identity matching score is constructed based on edit distance and average fusion probability. The database entry with the highest score that exceeds the preset verification threshold is determined as the final matching result; otherwise, it is marked as pending review.

[0011] Furthermore, the training process of the joint segmentation model includes: Ship branches in the joint segmentation model are supervised using ship region labeling; The identifier-carrying branch of the joint segmentation model is supervised by the strip region after character annotation merging.

[0012] Furthermore, the identity database index contains a unified mapping relationship for standard ship names, ship codes, and information on homonyms, abbreviations, and language variants.

[0013] A second aspect of the present invention provides an artificial intelligence-based intelligent identification system for ship identities, the system comprising: The offline modeling module is used to perform structural constraint enhancement on the original ship image samples, construct training samples that conform to the real maritime imaging conditions, and complete the joint training of the hull region segmentation model, the tracking feature extraction model, and the text detection and character recognition model based on these samples. At the same time, it establishes an identity database index corresponding to the standard ship name. The online recognition module is used to extract the ship region from the input video sequence to be recognized using a joint segmentation model, and introduces a marker-carrying area guidance and temporal response stabilization mechanism to generate a stable target image sequence. The target tracking module is used to update the continuous bounding boxes in the global coordinate system of the original frame of the target image sequence using a tracking feature extraction model, thereby achieving target tracking. The character recognition and verification module is used to detect text regions in the tracked image sequence and suppress single-frame errors by using probability fusion of the same character position in multiple frames to generate fused recognition results. The fused recognition results are then matched with the identity database index to complete the ship identity verification.

[0014] The beneficial technical effects of the present invention are at least as follows: To address the aforementioned issues, this invention provides an AI-based intelligent ship identity recognition method and system. First, in the offline phase, structural constraint enhancement is applied to the original ship image samples. By maintaining the consistency of the geometric arrangement of ship name characters, training samples that better reflect real maritime imaging conditions are constructed. Based on this, joint training of the hull region segmentation model, tracking feature extraction model, and text detection and character recognition model is completed. Simultaneously, an identity database index corresponding to standard ship names is established. Subsequently, in the online phase, the video sequence to be recognized sequentially undergoes hull region extraction, continuous target tracking, text region detection, character probability fusion, and identity matching. The hull region extraction process incorporates a marker-carrying area guidance and temporal response stabilization mechanism. The target tracking process employs a continuous bounding box update method under the global coordinate system of the original frame. The character recognition process utilizes probability fusion at the same character position across multiple frames to suppress single-frame errors. Finally, the identity database is combined to complete ship identity verification. Attached Figure Description

[0015] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0016] Figure 1 This is a flowchart of the intelligent ship identification method based on artificial intelligence according to the present invention.

[0017] Figure 2 This is a framework diagram of the intelligent ship identity recognition system based on artificial intelligence according to the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] In one or more embodiments, such as Figure 1As shown, an intelligent identification method for ship identity based on artificial intelligence is disclosed, the method comprising the following: S1: Perform structural constraint enhancement on the original ship image samples to construct training samples that conform to the real maritime imaging conditions. Based on this, complete the joint training of the hull region segmentation model, the tracking feature extraction model, and the text detection and character recognition model. At the same time, establish an identity database index corresponding to the standard ship name.

[0020] Specifically, this step, as the offline modeling stage, involves structural constraint enhancement, supervised data processing, and model training on the original ship image samples to form the model parameter set and identity database index required for subsequent online recognition. The input includes three parts of data. This represents the original set of ship images, which are derived from frame extraction results of port shore-based high-definition cameras, fixed waterway monitoring equipment, and drone inspection video streams. The images retain complete information about the ship's hull, ship name area, and surrounding environment, and cover scenes such as daytime strong light, nighttime lighting, rain and fog, front lighting, backlighting, and oblique side views. Indicates and The manually labeled information corresponds to each frame, and the labeled content includes the location of the ship area, the location of the entire ship name area, the location of the character's bounding rectangle, and the character's arrangement order; This represents the original ship identity database, containing standard ship names, ship codes, and information on variant spellings, abbreviations, and language variations of the same name. During processing, the boundary coordinates of the character frames in each image are first read in the labeled order. The width, height, and center position of each character frame are calculated. Then, the character spacing is obtained based on the center coordinates of adjacent characters. This process constructs a structural description value representing the geometric arrangement of the current ship name characters. ; in, The structural description value representing the current ship name; This indicates the number of characters in the ship's name, which was directly counted by manual annotation; Indicates the first The width of each character box is obtained by the difference between the coordinates of its left and right boundaries; Indicates the first The height of each character box is obtained by the difference between the coordinates of its upper and lower boundaries; Indicates the first The distance between the center of each character and the center of its adjacent characters is calculated from the coordinates of the centers of the adjacent characters. This represents the average distance between the centers of all adjacent characters, calculated from all... The average is calculated. This structural description value retains two types of information: the aspect ratio of the characters and the spacing between characters, reflecting the basic arrangement pattern of the ship names in the image. Taking a ship name image captured from a port oblique perspective as an example, if a segment of the ship name contains 10 characters, most of which have similar aspect ratios and the center distance between adjacent characters remains relatively balanced, then... It will fall within a stable range; when the image only experiences changes in brightness, local glare, or slight blurring, the sharpness of character boundaries will decrease, but the relative positions between characters usually maintain their original pattern. The changes were minor.

[0021] Based on this structural descriptor value, enhancement operations consistent with the actual scene are performed on the original image. Enhancement methods include small-angle rotation to simulate camera installation deviation and ship sway, perspective stretching to simulate the observation effect of shore-based equipment on a slanted ship, local blurring to simulate speed changes and momentary lens defocusing, local brightness enhancement to simulate sea surface reflection and side metal reflection, and overall brightness compression while retaining locally bright areas to simulate nighttime searchlight illumination. For several frames from the same video sequence that are temporally adjacent, the same set of enhancement parameters is used for synchronous processing during enhancement, ensuring consistency in brightness, blur, and perspective changes between adjacent frames within the same sequence, retaining only the true inter-frame differences of the original video itself. This allows subsequent video-level segmentation and tracking training to be based on a consistent image distribution. During enhancement, the character bounding boxes and ship region boxes undergo the same geometric transformation synchronously, thus maintaining a strict correspondence between the enhanced image and the enhanced annotations. After each enhancement, a new structural descriptor value is recalculated using the synchronously transformed character bounding boxes. The enhancement results are then filtered using the following formula. ; in, The structural description value of the enhanced image is calculated in the same way as... It's the same, except the input is replaced with an enhanced character box; This represents the structural description value of the original image before enhancement; This represents the permissible range of structural changes, pre-defined using a small number of representative samples, used to distinguish between samples that conform to actual imaging changes and anomalous samples that disrupt character arrangement. In a typical nighttime airway sample set, after local highlight enhancement in some images, parts of the character strokes are covered by light spots, but the positional relationship of the character frames remains stable. The calculated... and Samples with small differences are retained; in another set of images, when the perspective stretching parameter is too large, adjacent character boxes are over-compressed, causing the center distance of some characters to deviate significantly from the original proportion. and When the difference exceeds a threshold, the corresponding sample is discarded. After this processing, the remaining image samples not only carry the complex imaging features of real seafaring conditions but also maintain the spatial consistency of ship name characters.

[0022] After the sample selection is completed, a set of offline model parameters is trained using the retained samples. The aforementioned It consists of three parts: first, a joint segmentation model, which simultaneously outputs a ship region probability map and an identifier carrying region probability map when inputting a ship image; second, a tracking feature extraction model, which extracts appearance features suitable for candidate box matching from local images of adjacent frames; and third, a text detection and character recognition model, which outputs the text region and the probability distribution of characters at each character position. During training, ship region annotations are used to supervise the ship branch of the joint segmentation model, the merged character annotation strip region is used to supervise the identifier carrying branch of the joint segmentation model, consecutive frame samples are used to supervise the tracking feature extraction model, and character sequence annotations are used to supervise the text detection and character recognition model. Simultaneously, the original ship identity database is used... Standardize and organize the information, unify the format of ship names, aliases, abbreviations, and language mappings, and create an identity database index for subsequent online matching. After the above processing, this step outputs... and ,in This represents the training parameter set used for subsequent online ship area extraction, target tracking, and identifier recognition. This represents the normalized index of the ship identity database.

[0023] S2: For the input video sequence to be identified, the joint segmentation model is used to extract the ship region, and a marker-carrying region guidance and temporal response stabilization mechanism is introduced to generate a stable target image sequence.

[0024] Specifically, this step serves as the initial processing step in the online recognition phase, with the input being the video sequence to be recognized. and the model parameter set output in step 1 .in, This represents a sequence of video frames of the vessel to be identified acquired in real time, with each frame still in the original frame's global coordinate system; subscript Indicates the first video in the same video sequence Frames, adjacent The frame is determined directly by the video frame order. During processing, the current frame is input into the joint segmentation model trained in step 1. The shared encoder extracts multi-scale features from local texture to overall contour, and then the two decoding branches output the ship area probability map and the identification carrying area probability map, respectively. The former is denoted as... , representing the probability that each pixel in the current frame belongs to the ship region; the latter is denoted as , representing the probability that each pixel in the current frame belongs to the ship name character carrying area. Since the identifier carrying area is usually located on the outer plating of the hull or the outer wall of the superstructure, this probability map can serve as a structural prior in hull segmentation. Based on the weighted fusion approach commonly used in image segmentation, the original ship region probability map is enhanced with structure guidance to obtain a joint response map. ; in, Represents the joint response graph of the current frame; The probability map of ship regions output by the joint segmentation model; This is the probability map of the identifier carrying area output by the joint segmentation model, with values ​​ranging from 0 to 1; The adjustment coefficient controls the enhancement level of the marker-bearing region's response to hull segmentation. This formula originates from an image weighted enhancement model, whose basic form is a product of the signal and weight factors, upon which the following is introduced... As a structural guiding term, it concentrates the hull segmentation results towards the identified load-bearing area. In the specific calculation process, for example, a pixel point... The value in the range is 0.6. If the pixel is located in the high-signal-carrying response area and ,set up After enhancement If the pixel is far from the identifier carrying area and ,but This process ensures that the hull segmentation results depend not only on the outer contour but also on the spatial information carried by the text.

[0025] Considering the dynamic disturbances in the marine environment such as wave reflection, localized highlights, and floating objects, and the continuous motion of the ship target within a short period, a time consistency constraint is introduced to optimize the segmented response. This method originates from the exponential weighting and differential suppression model in time-series signal processing. Before fusion, the response of the previous frame is spatially aligned to separate the actual ship displacement from the background disturbance. Specifically, the joint response map of the current frame is first... Joint response map with the previous frame Each pixel is compared with a preset response threshold, and pixels exceeding the threshold are extracted to form binary candidate regions. Connectivity analysis is then performed on these binary candidate regions, and the connected region with the largest area is selected as the hull candidate region for both the current and previous frames. The difference in center coordinates between the two regions is calculated as the inter-frame displacement. Subsequently, Translate by this displacement to obtain the aligned response of the previous frame. The current frame is then fused with the aligned response of the previous frame, and a difference suppression term is added to obtain a time-consistent response. ; in, Indicates the time consistency response of the current frame; Represents the joint response graph of the current frame; Represents the joint response graph of the previous frame. The result after center displacement alignment; The weight of the current frame; This is the difference suppression coefficient. In this formula, the first two terms are weighted averages used to preserve continuous target information; the third term is the difference suppression term used to weaken rapidly changing areas. Taking a set of consecutive frames as an example, if the responses of a certain hull region in the current frame and the previous frame after alignment are 0.85 and 0.8 respectively, then the difference is 0.05. , Sometimes, For reflective areas on the sea surface, if the response difference between two frames is significant, for example, 0.7 and 0.2, then the difference is 0.5. This will be significantly weakened. The temporal consistency here only applies to the segmented response layer to stabilize the probability distribution of the hull region; it does not handle target-level tracking or bounding box updates.

[0026] In obtaining Then, these regions are compared with a preset threshold. Pixels exceeding the threshold are extracted to form candidate regions, and connected component analysis is used to select the connected region with the largest area as the main body region of the ship. Connected component analysis is based on pixel adjacency relationships, grouping adjacent high-response pixels into the same region and calculating the region's area and shape features. During this process, [the analysis is combined with...] The spatial distribution of candidate regions is used to further filter them, prioritizing connected regions with a high overlap with the identified carrying area to avoid misidentifying nearby ships or shore structures as targets. Finally, the minimum bounding rectangle is calculated for the filtered regions to obtain the ship's location, and this location is then used to extract information from the original frame. Extract the corresponding local image from the middle to form the output. Simultaneously, the coordinates of the ship region binary mask and the minimum bounding rectangle in the original frame's global coordinate system are uniformly recorded as follows: .in, This represents a sequence of partial images of the ship's hull obtained by cropping frames. This represents the ship region mask and its global coordinate rectangle for the corresponding frame.

[0027] S3: Use the tracking feature extraction model to update the continuous bounding boxes in the original frame global coordinate system of the target image sequence to achieve target tracking.

[0028] Specifically, the input for this step is the output of step 2. , and the model parameter set output in step 1 It is used for continuous frame tracking and dynamic updating of target boxes in the ship area. The sequence of partial images of the hull is extracted from each original frame according to the minimum bounding rectangle obtained in step 2. The main body of the hull and the area carrying the logo occupy the main position in the image. It contains the binary mask of the ship region corresponding to each frame and the coordinates of the minimum bounding rectangle in the original frame's global coordinate system. In the first frame processing, using... The global coordinate rectangle in the original frame is used as the initial target bounding box; in subsequent frame processing, the generation and updating of the target bounding box, the calculation of candidate box displacements, and trajectory recording are all performed in the global coordinate system of the original frame. The local image in the image is only used as the local observation area corresponding to the global target box for feature extraction. This step inherits the "stable distribution of the identifier-bearing area" formed in step 2 in space, and further establishes the "continuous and stable change of the target box" in time.

[0029] During processing, the first frame is... The provided minimum bounding rectangle global coordinates are used as the initial bounding box, and a local image is cropped from the corresponding original frame based on these global coordinates as the first frame reference template. Starting from the second frame, features are extracted from the local images corresponding to the bounding boxes in the current frame and the previous frame. Specifically, the local images are input into the tracking feature extraction model trained in step 1. This model consists of a three-layer convolutional structure, with each layer using a 3×3 convolutional kernel and non-linear activation. The first layer extracts the hull edge and waterline information, the second layer extracts the side texture and local structure, and the third layer extracts the overall hull contour and orientation information. Finally, a fixed-length feature vector is obtained through global averaging and convergence. Starting from the second frame, the updated bounding box from the previous frame is used... As a reference frame, and in the original frame global coordinate system of the current frame, with Using the center of the reference frame as a reference, multiple candidate boxes are generated within its neighborhood. These candidate boxes are obtained by making small translations in the horizontal and vertical directions and scaling their width and height proportionally. Each candidate box is then mapped to the corresponding local image region of the current frame, cropped, and input into the same feature extraction model to obtain candidate feature vectors. The similarity between the reference feature and the candidate feature is calculated through a normalized inner product to obtain an appearance consistency measure.

[0030] After obtaining the appearance similarity score, spatial displacement constraints and mask fitting constraints are introduced to form a comprehensive score. This scoring model is based on the classic multi-feature weighted fusion idea, and adds a mask fitting term to make the score reflect not only appearance similarity but also the degree of consistency between the candidate bounding box and the actual hull area. The comprehensive score is defined as follows: ; in, The overall score for the candidate boxes; The appearance similarity is obtained by the normalized inner product of the reference feature and the candidate feature; To determine the mask fitting ratio, after mapping the current candidate bounding box to the ship region mask in the current frame, the area within the candidate bounding box belongs to... The ratio of the number of pixels in the ship region mask to the total number of pixels in the candidate boxes is calculated. This is the Euclidean distance between the center of the current candidate box and the center of the target box in the previous frame in the original frame global coordinate system. This distance is calculated by the difference between the global center coordinates of the two frames. This is the upper limit of the normalized distance in the global coordinate system of the original frame, and its value is determined based on the maximum possible displacement of the ship in adjacent frames. and These represent the weights for the appearance and masking terms, respectively. This formula is derived from a linear weighted model and incorporates a normalized displacement penalty term to keep the score within a reasonable range. Taking a continuous channel monitoring image as an example, let the appearance similarity between a candidate bounding box and the reference template be... Its overlap ratio with the hull mask The center displacement is 0.3 times the maximum allowable displacement, that is... ,Pick , The overall score is If another candidate box has a slightly higher appearance similarity of 0.85, but a mask fitting ratio of only 0.60 and a displacement ratio of 0.6, then its score is... Ultimately, the former was selected as the target bounding box for the current frame.

[0031] After selecting the candidate bounding box with the highest score, a time-series-based smoothing update is performed to ensure that the changes in the target box across consecutive frames conform to the ship's motion patterns. This update method originates from the exponential smoothing model in time series analysis. Its basic idea is that the current state is obtained by weighting the current observation with the historical state, specifically expressed as follows: ; in, The updated bounding box parameters for the current frame include the center coordinates and width and height in the original frame's global coordinate system; The parameters of the candidate bounding boxes selected by the comprehensive score in the current frame are in the coordinate system of... Consistent; The bounding box parameters of the previous frame are also in the original frame's global coordinate system. This is for updating coefficients. The formula is derived from a first-order recursive filtering model. In actual calculations, if the target box parameters of the previous frame... The center is Width and height are Current candidate box parameters The center is Width and height are ,Pick The updated center is After the update, the width and height are... This update method can suppress abnormal fluctuations in a single frame while maintaining the target response speed, such as momentary shifts caused by wave occlusion or localized highlights.

[0032] Through the above scoring selection and smoothing update process, a stable sequence of target boxes in consecutive frames is obtained. Finally, based on... Extract the corresponding local image in the global coordinate system of the original frame to form the output. Simultaneously, the target bounding box parameters of each frame are recorded in chronological order to form... This stores the trajectory parameters of the target bounding box in the global coordinate system of the original frame. Since the target bounding box has undergone both spatial and temporal constraints in consecutive frames, the output... Maintain consistency in position, scale, and orientation. It provides complete motion trajectory information.

[0033] S4: Perform text region detection on the tracked image sequence, and use probability fusion of the same character position in multiple frames to suppress single-frame errors and generate fusion recognition results; match the fusion recognition results with the identity database index to complete ship identity verification.

[0034] Specifically, the input for this step is the output of step 3. , and the model parameter set output in step 1 with identity database index It is used to complete the identification and authentication of ship markings. The image sequence of ships is aligned by continuous frame tracking, and the position, scale and marked areas of the ship in the images remain stable in the time dimension. The input includes global trajectory parameters of the target bounding box for each frame. These parameters provide the spatial location of the hull region in each frame and its changing trend across consecutive frames. This input format allows the recognition process to directly utilize the features of "spatial alignment + temporal continuity," extending single-frame character recognition to multi-frame joint recognition. Since ship markings are typically located on the hull side or the outer wall of the superstructure, they gradually enter, leave, or are partially obscured as the ship moves across consecutive frames. Therefore, the recognition stage needs to consider both the local sharpness of the characters and their consistency over time.

[0035] During processing, first in Text region detection is performed within the target bounding box region of each frame. Specifically, the image is input into the text detection and character recognition model trained in step 1. This model first outputs a text probability map through a multi-layer convolutional and feature fusion structure, then crops and uniformly scales the detected text regions, outputting a character probability distribution sequence for each frame and each character position. In other words, for the same character position, the model provides a probability distribution for each character category in each frame. To utilize complementary information between consecutive frames, the character probabilities at the same character position in each frame are weighted and superimposed during the multi-frame fusion stage. This method originates from the expectation-weighted model in probability theory, and introduces weights based on the stability of the target bounding box, allowing stable frames to contribute more to the final result. The specific calculation is as follows: ; in, Indicates the character at a certain position Fusion probability across all frames; This indicates the number of frames participating in the fusion, and this number is determined by... The length of consecutive frames in the middle is determined; Indicates the first The character at that position in the frame The recognition probability, which is output by the text detection and character recognition model; Indicates the first The weight of the frame, which is based on The global trajectory parameters of the original frame recorded in the calculation are determined by the displacement difference between the global center of the target box in the current frame and the global centers of the target boxes in the preceding and following frames. The smaller the displacement, the more stable the frame, and the greater the corresponding weight. This formula comes from the weighted average model. In this way, the recognition results of the same character position in different frames can be fused, thereby compensating for the missing or incorrect results in single-frame recognition. Taking a practical calculation as an example, suppose the recognition probability of the character "C" at a certain character position in the three frames is as follows: , , The corresponding weight is , , Then there is This result is more stable than the single-frame probability and is more affected by high-weight frames.

[0036] After obtaining the complete recognition string, the recognition result is compared with the identity database index compiled in step 1. The process involves matching. The database stores standard ship names and their corresponding codes. The difference between the identified result and the database entry is calculated through string matching and edit distance calculations. This process is then used to construct an identity matching score. ; in, Indicates the recognition result and the first Matching score for each database entry; Indicates that the currently identified string is related to the first... Edit spacing between standard ship names; Indicates the length of the currently recognized string; This represents the fusion probability of each character position in the final recognized string being selected. The average value; This represents the weight between the edit distance term and the character probability term. This formula is derived from a linear combination of the string similarity normalization model and the probability weighting model, where... Reflects the degree of name matching. This reflects the stability of the recognition results themselves. For a given database entry, if the edit distance... Identify string length Average character probability ,Pick Then there is Combining all candidate entries The entries are sorted, and those with the highest scores exceeding a preset verification threshold are identified as the final matching results. If the highest score does not reach the verification threshold, the vessel's identity is marked as pending verification. Final output. ,in To identify the results and corresponding vessel identity information, including character sequences and matching scores of candidate database entries. The final matching result and its confidence level, or the status pending verification.

[0037] In one or more embodiments, such as Figure 2 As shown, an intelligent identification system for ship identity based on artificial intelligence is disclosed, the system comprising: The offline modeling module is used to perform structural constraint enhancement on the original ship image samples, construct training samples that conform to the real maritime imaging conditions, and complete the joint training of the hull region segmentation model, the tracking feature extraction model, and the text detection and character recognition model based on these samples. At the same time, it establishes an identity database index corresponding to the standard ship name. The online recognition module is used to extract the ship region from the input video sequence to be recognized using a joint segmentation model, and introduces a marker-carrying area guidance and temporal response stabilization mechanism to generate a stable target image sequence. The target tracking module is used to update the continuous bounding boxes in the global coordinate system of the original frame of the target image sequence using a tracking feature extraction model, thereby achieving target tracking. The character recognition and verification module is used to detect text regions in the tracked image sequence and suppress single-frame errors by using probability fusion of the same character position in multiple frames to generate fused recognition results. The fused recognition results are then matched with the identity database index to complete the ship identity verification.

[0038] It is worth noting that the specific workflow of the AI-based intelligent ship identification system provided in this embodiment is the same as that of the AI-based intelligent ship identification method described in the above embodiments, and will not be repeated here.

[0039] This invention also provides an artificial intelligence-based intelligent ship identification device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps described in the above-described embodiments of the artificial intelligence-based intelligent ship identification method, for example... Figure 1 The steps S1 to S4 described above; or, when the processor executes the computer program, it implements the functions of each module in the above system embodiments.

[0040] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the AI-based intelligent ship identification device.

[0041] The AI-based intelligent ship identification device can be a desktop computer, laptop, handheld computer, or cloud server, among other computing devices. This AI-based intelligent ship identification device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the AI-based intelligent ship identification device may also include input / output devices, network access devices, and buses.

[0042] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the AI-based intelligent ship identification device, connecting all parts of the device via various interfaces and lines.

[0043] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the AI-based intelligent ship identification device by running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created according to the operation of the controller, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0044] The modules integrated into the AI-based intelligent ship identification device, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0045] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0046] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A ship identification intelligent recognition method based on artificial intelligence, characterized in that, The method includes: S1: Perform structural constraint enhancement on the original ship image samples to construct training samples that conform to the real maritime imaging conditions. Based on this, complete the joint training of the ship hull region segmentation model, the tracking feature extraction model, and the text detection and character recognition model. At the same time, establish an identity database index corresponding to the standard ship name. S2: For the input video sequence to be identified, the joint segmentation model is used to extract the ship region, and a marker-carrying region guidance and temporal response stabilization mechanism is introduced to generate a stable target image sequence; S3: Use the tracking feature extraction model to update the continuous bounding boxes in the original frame global coordinate system of the target image sequence to achieve target tracking; S4: Perform text region detection on the tracked image sequence, and use probability fusion of the same character position in multiple frames to suppress single-frame errors and generate fusion recognition results; The fused identification results are matched with the identity database index to complete the ship's identity verification.

2. The intelligent ship identification method based on artificial intelligence according to claim 1, characterized in that, The structural constraint enhancement of the original ship image samples includes: Calculate the geometric arrangement structure description value of the ship name characters in the original image; Perform enhancement operations consistent with the actual scene, including small-angle rotation, perspective stretching, local blurring, local brightness enhancement, and overall brightness compression; Simultaneously apply the same geometric transformation to the character annotation box and the ship area box; Calculate the structural description value of the enhanced image and compare the difference with the original structural description value; If the difference is less than the preset allowable range for structural changes, the enhanced sample is retained; otherwise, it is discarded.

3. The intelligent ship identification method based on artificial intelligence according to claim 1, characterized in that, The introduced identifier-bearing area guidance and timing response stabilization mechanism includes: The joint segmentation model is used to simultaneously output a probability map of the ship area and a probability map of the identified carrying area. The joint response map is generated by weighting the probability map of the ship area based on the probability map of the identification carrying area; The joint response map of the previous frame, after center displacement alignment, is fused with the joint response map of the current frame, and a difference suppression term is added to generate a time-consistent response. The main body region of the ship is extracted based on the time-consistent response, and the minimum bounding rectangle is calculated to generate the target image sequence.

4. The intelligent ship identification method based on artificial intelligence according to claim 1, characterized in that, The step of updating the continuous bounding boxes of the target image sequence in the original frame global coordinate system using the tracking feature extraction model includes: The smallest bounding rectangle of the first frame is used as the initial target box; Starting from the second frame, candidate boxes are generated in the original frame global coordinate system of the current frame, using the target box updated in the previous frame as the reference box. The feature vectors of the reference box and the candidate box are extracted using a tracking feature extraction model, and the appearance similarity is calculated. A comprehensive score is generated by combining appearance similarity, mask fitting ratio, and spatial displacement constraints. The candidate bounding box with the highest comprehensive score is selected and a smooth update based on time series is performed to generate the target bounding box for the current frame.

5. The intelligent ship identification method based on artificial intelligence according to claim 4, characterized in that, The tracking feature extraction model consists of a three-layer convolutional structure. The first layer is used to extract the hull edge and waterline information, the second layer is used to extract the side texture and local structure, and the third layer is used to extract the overall hull outline and orientation information.

6. The intelligent ship identification method based on artificial intelligence according to claim 1, characterized in that, The method of suppressing single-frame errors by utilizing probabilistic fusion at the same character position across multiple frames includes: The tracked image sequence is input into the text detection and character recognition model, which outputs the character probability distribution sequence at each character position in each frame. The weights of each frame are determined based on the trajectory parameters of the target bounding box in the global coordinate system of the original frame, with larger weights for smaller displacements. The probability of characters at the same position in each frame is weighted and superimposed to generate a fusion recognition result.

7. The intelligent ship identification method based on artificial intelligence according to claim 1, characterized in that, The process of matching the fused recognition results with the identity database index includes: Calculate the edit distance between the recognition result and the standard ship names in the database; Calculate the average fusion probability of the selected character at each position in the recognition result; An identity matching score is constructed based on edit distance and average fusion probability. The database entry with the highest score that exceeds the preset verification threshold is determined as the final matching result; otherwise, it is marked as pending review.

8. The intelligent ship identification method based on artificial intelligence according to claim 1, characterized in that, The training process of the joint segmentation model includes: Ship branches in the joint segmentation model are supervised using ship region labeling; The identifier-carrying branch of the joint segmentation model is supervised by the strip region after character annotation merging.

9. The intelligent ship identification method based on artificial intelligence according to claim 1, characterized in that, The identity database index contains a unified mapping relationship between standard ship names, ship codes, and information on variant spellings, abbreviations, and language variations of the same name.

10. An intelligent ship identification system based on artificial intelligence, characterized in that, The system is used to execute the AI-based intelligent ship identification method according to any one of claims 1-9, the system comprising: The offline modeling module is used to perform structural constraint enhancement on the original ship image samples, construct training samples that conform to the real maritime imaging conditions, and complete the joint training of the hull region segmentation model, the tracking feature extraction model, and the text detection and character recognition model based on these samples. At the same time, it establishes an identity database index corresponding to the standard ship name. The online recognition module is used to extract the ship region from the input video sequence to be recognized using a joint segmentation model, and introduces a marker-carrying area guidance and temporal response stabilization mechanism to generate a stable target image sequence. The target tracking module is used to update the continuous bounding boxes in the global coordinate system of the original frame of the target image sequence using a tracking feature extraction model, thereby achieving target tracking. The character recognition and verification module is used to detect text regions in the tracked image sequence and suppress single-frame errors by using probability fusion of the same character position in multiple frames to generate fused recognition results. The fused recognition results are then matched with the identity database index to complete the ship identity verification.

Images