A ship name recognition method and system based on similarity clustering and time sequence consistency analysis

Abstract

This invention provides a method and system for ship name recognition based on similarity clustering and temporal consistency analysis, belonging to the field of computer vision technology. The method includes: acquiring ship images and labeling only the ship name region, with the rest marked as ignored regions; training a dedicated recognition model based on an OCR pre-trained model; acquiring continuous image sequences of ships; identifying candidate texts and their confidence scores frame by frame; filtering according to ship name structure rules; calculating string similarity and clustering; performing temporal consistency analysis, calculating stability scores, and removing outlier clusters; and selecting the optimal cluster to output the text with the highest frequency and confidence score. The system includes modules for dataset construction, model training, image acquisition, single-frame recognition, rule filtering, similarity clustering, temporal analysis, and result output. This invention effectively suppresses interference in maritime scenarios, improves recognition accuracy and stability, reduces computational complexity, and is applicable to various ship name recognition scenarios such as fixed monitoring, shipborne, airborne, and low-light nighttime conditions.

Description

Technical Field

[0001] This invention relates to the field of computer vision and intelligent recognition technology, and more specifically, to a method and system for ship name recognition based on similarity clustering and temporal consistency analysis. Background Technology

[0002] A ship's name is its unique identifier, and its accurate identification plays a crucial role in maritime supervision and port operations. Traditional ship name identification methods primarily rely on manual observation, radar monitoring, and AIS systems, which suffer from limitations such as low efficiency, high error rates, and unreliable information. Furthermore, AIS systems are susceptible to risks such as being manually shut down or illegally exchanged, leading to unrecognizable ship names. Therefore, text recognition technologies based on computer vision and deep learning are gradually being applied and developed in the field of ship name identification. However, existing text recognition technologies still have significant shortcomings in real-world maritime scenarios. (1) The interference of light and water mist in maritime scenarios can easily cause noise in ship name recognition, and there is a lack of screening mechanism for ship name structure, making it difficult to distinguish ship name from irrelevant text. (2) Existing technologies are mostly designed for single-frame image recognition, without utilizing the temporal characteristics of continuous image sequences. Single-frame results are easily affected by instantaneous interference, including lens fog, slight swaying of the ship, changes in light intensity, and instantaneous obstruction of the ship name area by birds or floating objects. This can lead to misidentification or omission in single-frame recognition. Furthermore, single-frame recognition lacks scientific clustering and temporal consistency analysis, resulting in poor recognition stability. (3) Many existing ship name recognition technologies do not distinguish between ship names and irrelevant text areas during training, which will further affect the recognition accuracy. Summary of the Invention

[0003] The purpose of this invention is to address the problems of current ship identification methods, such as weak anti-interference, low accuracy, and poor stability, which make it difficult to meet the requirements of accurate identification. This invention provides a ship name identification method and system based on similarity clustering and temporal consistency analysis. This method performs similarity clustering on candidate texts identified in continuous ship image sequences, merging potentially inconsistent identification results to avoid the impact of instability in single-frame image acquisition on the final result. Simultaneously, a stability scoring mechanism is incorporated to select the optimal candidate text cluster from multiple frames, and the final output result is determined further through intra-cluster frequency statistics, thereby significantly improving the stability and overall accuracy of ship name identification results.

[0004] To achieve the above-mentioned technical features, the objective of this invention is as follows: The first aspect of this invention provides a ship name recognition method based on similarity clustering and temporal consistency analysis, comprising: S1. Collect a dataset of ship images, and only annotate the text regions and content corresponding to the ship names to obtain annotation data; mark the remaining text regions as ignored regions. S2. OCR-based text detection and recognition pre-trained model: Use labeled data to train the ship name text detection and recognition model; S3. Obtain a continuous image sequence of the vessel to be identified; S4. Perform text detection and recognition on each frame of the image sequence to obtain candidate texts and their corresponding confidence levels; S5. Based on the preset ship name structure rules, the candidate texts are filtered to obtain a set of valid candidate ship names; S6. Calculate the string similarity of the valid candidate ship names and construct a similarity matrix. Based on the similarity matrix, perform clustering to obtain candidate text clusters. S7. Perform consistency analysis on the clustering results based on time series, calculate stability index and remove outlier clusters; S8. Select the candidate text cluster with the highest stability and output the text with the highest frequency in the cluster as the ship name recognition result; if there are multiple texts with the same frequency, select the text with the highest confidence.

[0005] Preferably, step S1 further includes: performing effective image screening and watermark cropping preprocessing on the acquired ship images.

[0006] Preferably, in step S2, PaddleOCR's pre-trained text detection and text recognition models are fine-tuned to train the ship name detection model and the ship name recognition model, respectively. During the fine-tuning training process using ship name image inputs, data augmentation strategies are employed to improve model robustness. For text labeled as an ignored region, it is not included in the loss calculation during model training.

[0007] Preferably, in step S3, continuous images containing the ship's hull and name area are collected by a fixed monitoring camera, shipborne high-definition equipment, or airborne inspection equipment; the video stream is framed at preset time intervals to obtain an image sequence.

[0008] Preferably, the ship name structure rules in step S5 include at least one of the following: character type rules, length rules, and format rules.

[0009] Preferably, in step S6, string similarity is calculated using an edit distance-based method, and the similarity calculation formula is as follows: ; in, Representing text and Edit distance, Indicates the text length; when the similarity is greater than a preset threshold, candidate texts are grouped into the same cluster.

[0010] Preferably, a cluster center mechanism is introduced during clustering. New texts are only calculated to have similarity with each cluster center. If the similarity meets the threshold, the text is assigned to the corresponding cluster; otherwise, a new cluster is created.

[0011] Preferably, the temporal consistency analysis in step S7 includes at least one of the following: the frequency of occurrence of candidate text clusters in consecutive frames, the magnitude of change of candidate texts in the time series, and the change in text similarity between adjacent frames.

[0012] Preferably, the stability score calculation formula in step S7 is: ; in, Indicates the frequency of occurrence of candidate text clusters. Indicators representing time-series consistency This represents the mean confidence level. , , These are the weight parameters.

[0013] Preferably, in step S8, the candidate text clusters are sorted according to their stability scores: ; in, This represents the candidate text cluster number with the highest stability score; From the candidate text clusters The text that appears most frequently is selected as the final recognition result: ; in, Representing text In candidate text clusters The number of times it appears in the text.

[0014] Another aspect of the present invention provides a ship name recognition system based on similarity clustering and temporal consistency analysis. The system is used to implement the recognition method and includes: The dataset building module is used to collect ship image datasets, label the ship name text area, and mark the remaining areas as ignored areas; The model training module is used to train a ship name text detection and recognition model based on an OCR pre-trained model and using labeled data. The image acquisition module is used to acquire a continuous sequence of images of the vessel to be identified. The single-frame recognition module is used to perform text detection and recognition on an image sequence frame by frame, and output candidate text and confidence scores. The rule filtering module is used to filter and obtain a set of valid candidate ship names based on ship name structure rules; The similarity clustering module is used to calculate the similarity of candidate texts, construct a similarity matrix, and perform clustering. The time series analysis module is used to perform time series consistency analysis on clustering results, calculate stability indices, and remove outlier clusters. The result output module is used to select the cluster with the highest stability and output the text with the highest frequency and the best confidence in the cluster as the recognition result.

[0015] Preferably, the model training module uses a pre-trained detection and recognition model of PaddleOCR for fine-tuning, ignoring regions that are not included in the loss calculation.

[0016] Preferably, the similarity clustering module calculates similarity using edit distance and optimizes clustering efficiency using a cluster center mechanism.

[0017] Preferably, the time series analysis module calculates a stability score based on a weighted average of occurrence frequency, time series continuity, and average confidence level.

[0018] Preferably, the image acquisition module supports video frame extraction, image scaling, noise reduction, and enhanced preprocessing.

[0019] The present invention has the following beneficial effects: 1. This invention performs similarity clustering on candidate texts identified in a continuous sequence of ship images, merging potentially disparate identification results to avoid the impact of instability in single-frame image acquisition on the final result. Simultaneously, a stability scoring mechanism is used to select the optimal candidate text cluster from multiple frames, and the final output result is determined further through an intra-cluster frequency statistics strategy, thereby significantly improving the stability and overall accuracy of ship name recognition results.

[0020] 2. By introducing a multi-dimensional consistency analysis method based on occurrence frequency, temporal continuity, and confidence level, it is possible to effectively identify and eliminate occasional erroneous recognition results, reduce the impact of noisy text, occlusion interference, and complex background on recognition results, and make the system more robust in practical application environments.

[0021] 3. Based on the text recognition results, multiple rules such as character type, text length, and ship naming format are introduced to filter candidate texts. This can effectively remove interfering texts that do not conform to the characteristics of ship names, thereby improving the quality of the candidate text set and providing a more reliable data foundation for subsequent similarity clustering and time series consistency analysis.

[0022] 4. During the clustering process, by introducing a cluster center mechanism, new candidate texts only need to perform similarity calculations with the cluster center, reducing unnecessary pairwise comparisons, lowering computational complexity, and improving processing efficiency, thereby meeting the real-time or near real-time processing needs of video streams or continuous image scenarios.

[0023] 5. By annotating only the ship name region during the data annotation stage and setting non-target regions as ignored regions, they are not included in the loss calculation during model training, thereby reducing the interference of irrelevant text and improving the model's ability to recognize target text and its generalization performance. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Figure 1 This is a flowchart illustrating a ship name recognition method based on similarity clustering and temporal consistency analysis as described in this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] Example 1: Please see Figure 1 As shown, this embodiment provides a ship name recognition method based on similarity clustering and temporal consistency analysis, including the following steps: S1. Collect a dataset of ship images, annotating only the text regions and content corresponding to the ship names, and marking the remaining text regions as ignored regions; specifically including the following steps: S101. Acquire image data containing ships; S102. Accurately label the text areas of ship names in the image; S103. Mark the non-ship name text region as "###" ignored region so that it does not participate in the loss calculation during model training; By using the above methods, the interference of irrelevant text on model training is reduced, and the model's ability to recognize target text is improved.

[0028] S2. PaddleOCR-based pre-trained text detection and recognition model: The text detection and recognition model is trained using the above-mentioned labeled data; specifically, it includes the following steps: S201. Divide the labeled data into a training set and a validation set; S202. Train the model parameters based on the training set; S203. During training, the ignored region is not included in the loss calculation; S204. Evaluate the model performance using the validation set and adjust the model parameters.

[0029] S3. Obtain a continuous image sequence of the vessel to be identified; specifically including the following steps: S301. Capture continuous image data using a camera device; S302. If the acquired data is a video stream, then the video stream is processed by frame extraction according to a preset time interval to obtain an image sequence. S303. Perform preprocessing on the image, including scaling, denoising, or enhancement, to improve the accuracy of subsequent text recognition.

[0030] S4. Perform text detection and recognition on each frame of the image sequence to obtain candidate text and corresponding confidence scores; specifically including the following steps: S401, perform text region detection on the image to obtain multiple text candidate regions; S402, perform character recognition on each text candidate region to obtain candidate text; S403, assign a corresponding confidence level to each candidate text; This yields a candidate text set for continuous image data of each ship.

[0031] S5. Based on the ship name structure rules, filter the candidate texts, remove texts that do not meet the conditions, and obtain a set of valid candidate texts; specifically, this includes the following steps: S501, based on character type rules, only retain text containing Chinese characters or a combination of Chinese characters and numbers; S502, filter according to text length rules and remove text with abnormal length; S503, based on format rules, filter and retain text that conforms to the naming rules of ships; To obtain a set of valid candidate texts.

[0032] S6. Calculate pairwise string similarity for the valid candidate texts, construct a similarity matrix, and cluster the candidate texts based on the similarity matrix to obtain multiple candidate text clusters; the specific method is as follows: (1) For any two candidate texts and Calculate edit distance; (2) Calculate string similarity based on edit distance, wherein the similarity is defined as: ; in, Representing text and Edit distance, Indicates text length; (3) Construct a similarity matrix for candidate texts: ; (4) Based on similarity threshold Clustering is performed to group candidate texts into the same class when the following conditions are met: ; (5) Further, each candidate text cluster determines a cluster center text. New candidate text Upon arrival, similarity is calculated only with the centers of each cluster: ; (6) When the similarity is greater than the preset threshold, the candidate text is assigned to the corresponding cluster: ; Otherwise, create a new candidate text cluster.

[0033] S7. Based on the temporal order of the image sequences, perform consistency analysis on each candidate text cluster, calculate its stability index, and remove candidate text clusters with stability below a preset threshold; the specific method is as follows: S701, Calculate the frequency of occurrence of each candidate text cluster in the image sequence: ; in, Indicates the first k The number of times each candidate text cluster appears N This indicates the total number of frames.

[0034] S702, Timing Continuity Calculation: ; in, This indicates the longest consecutive occurrence length of a candidate text cluster in the time series; S703, Calculate the average confidence level of candidate text clusters: ; in, Indicates the confidence level of the corresponding candidate text; S704, calculate the stability score based on the above indicators; ; in, Indicates the frequency of occurrence of candidate text clusters. Indicators representing time-series consistency This represents the mean confidence level. Wherein, , , Let be the weight parameters, and satisfy: ; S8. Based on the stability index of the candidate text clusters, select the candidate text cluster with the highest stability and output the text with the highest frequency in the cluster as the final recognition result of the ship name. If there are multiple texts with the highest frequency in the cluster, select the text with the highest confidence as the final recognition result. The specific method is as follows: S801. Sort the candidate text clusters according to their stability scores: ; in, This represents the candidate text cluster number with the highest stability score; S802, from the candidate text cluster The text that appears most frequently is selected as the final recognition result: ; in, Representing text In candidate text clusters The number of times it appears in; S803. If there are multiple most frequently occurring texts in the candidate text cluster, the text with the highest confidence level shall be selected as the final recognition result.

[0035] Example 2: Combination Figure 1 As shown, this embodiment provides a ship name recognition method based on similarity clustering and temporal consistency analysis, including the following steps: Step S1: Collect a dataset of ship images, label the text regions and content corresponding to the ship names, and mark the remaining text as ignored regions.

[0036] 8287 1920×1080 images of navigable vessels were captured from fixed monitoring camera locations along an inland waterway. Watermarks were cropped from the images, and the PPOCRLabel annotation tool was used to annotate the vessel name area and corresponding text content. Non-vessel name text was marked as "###" and ignored. The ignored annotations will not be included in the loss calculation when training the PaddleOCR model.

[0037] Step S2: Train a text detection and recognition model based on the labeled data.

[0038] The labeled dataset is divided into a training set and a validation set in a ratio of 8:2. Based on the text detection pre-trained model PP-OCRv5_server_det and the text recognition pre-trained model PP-OCRv5_server_rec of PaddleOCR, secondary development is carried out to train the ship name detection model and the ship name recognition model respectively.

[0039] Step S3: Obtain a continuous image sequence of the ship to be recognized.

[0040] The entry of the ship is recognized through a pre-configured ship recognition algorithm, and then the fixed-point monitoring camera captures multiple consecutive sequence pictures of the ship to be recognized from entry to departure at a frequency of 3 seconds each time. When capturing pictures at night, after the ship is recognized to enter, the fill light is turned on to capture a clear sequence of ship pictures.

[0041] Step S4: Perform text detection and recognition on each frame of the image sequence to obtain candidate texts and corresponding confidence levels.

[0042] Input the continuous image sequence frame by frame into the trained ship name detection model to locate all text candidate regions in the image; then input each text region into the ship name recognition model to complete character recognition, screen the candidate texts with a confidence level greater than the threshold, and finally obtain a preliminary candidate text set and corresponding confidence levels (example: "Chuan Shunqing Cargo 6890" (0.93), "Chuan Shunqing Cargo 689" (0.89), etc.).

[0043] Step S5: Screen the candidate texts based on the ship name structure rules, remove the texts that do not meet the conditions, and obtain a set of valid candidate ship names.

[0044] Screen the candidate texts according to the preset ship name structure rules. The candidate texts need to meet the following naming rules: (1) The text starts with Chinese characters; (2) The text consists only of Chinese characters and numbers; (3) The text length is 2 to 10 characters; (4) The text contains at least one numeric character.

[0045] Eliminate all interfering texts that do not meet the rules to obtain a set of valid candidate ship names (such as "Chuan Shunqing Cargo 6890", "Chuan Shunqing Cargo 689", etc.).

[0046] Step S6: Calculate the pairwise string similarity of the valid candidate texts, construct a similarity matrix, and cluster and group the candidate texts based on the similarity matrix to obtain multiple candidate text clusters.

[0047] (1) For any two candidate texts and Calculate the edit distance; (2)Calculate the string similarity based on the edit distance. The similarity calculation formula is as follows: ; Among them, represents the edit distance between the text and , represents the text length; (3)Construct a similarity matrix for the candidate texts: ; (4)Based on the similarity threshold perform clustering grouping, and divide the candidate texts that meet the following conditions into the same category: 0.8; (5)Furthermore, determine a cluster center text for each candidate text cluster. When a new candidate text arrives, only calculate the similarity with each cluster center: ; (6)When the similarity is greater than the preset threshold, classify the candidate text into the corresponding cluster: ; Calculate the string similarity between pairwise valid candidate texts using the edit distance. The calculation formula is: ; Construct a similarity matrix based on the calculation results, perform clustering grouping on the valid candidate texts according to the preset similarity threshold T = 0.85, introduce a cluster center mechanism to optimize the clustering efficiency, and finally obtain 2 candidate text clusters: Cluster 1 {"Chuan Shunqing Goods 6890", "Chuan Shunqing Goods 689"}, Cluster 2 {"Chuan Shunqing Goods 6819"}.

[0048] Step S7: Based on the time order of the image sequence, perform consistency analysis on each candidate text cluster, calculate its stability index, and eliminate the candidate text clusters with stability lower than the preset threshold; Based on the time order of consecutive images, perform temporal consistency analysis on the 2 candidate text clusters, and calculate the multi-dimensional indexes of each cluster respectively: Occurrence frequency , temporal consistency index , confidence mean , and calculate the stability score of each cluster according to the following formula: ; The calculation results are as follows: the stability score of cluster 1 is 0.862, and the stability score of cluster 2 is 0.34. Cluster 2 with a stability score lower than the preset threshold of 0.6 is determined as an abnormal cluster and is screened out, and only cluster 1 is retained.

[0049] Step S8: According to the stability index of the candidate text clusters, select the candidate text cluster with the highest stability, and output the text with the highest frequency in the cluster as the final recognition result of the ship name.

[0050] Select the candidate text cluster 1 with the highest stability according to the stability score, and count the occurrence frequencies of each text in cluster 1. If there are texts with the same occurrence frequency, select the text with the highest confidence. For example, "Chuan Shunqing Cargo 6890" appears the most times, and "Chuan Shunqing Cargo 6890" is output as the final recognition result of the ship name.

[0051] Meanwhile, ablation experiments were also carried out in the embodiment. The similarity clustering and temporal consistency analysis modules of a ship name recognition method based on similarity clustering and temporal consistency analysis proposed in this invention were ablated. Among them, the ship name recognition method based on deep learning only uses the ship name detection model and ship name recognition model trained in step S2 for recognition, without using continuous image sequences, and the final comparison results are shown in Table 1 below: Table 1 Comparison of calculation results of different methods

[0052] The results of the above ablation experiments show that without introducing similarity clustering and temporal consistency analysis, the accuracy rate of the ship name recognition method based on deep learning in the embodiment is 74.55%. After introducing similarity clustering and temporal consistency analysis, the accuracy rate is increased to 85.45%, an increase of about 10.9 percentage points. This result shows that this invention effectively reduces the influence of the quality of single-frame ship images and recognition errors through similarity clustering of the recognition results of multiple consecutive images and comprehensive determination in combination with temporal consistency, improves the stability and reliability of the recognition results, and thus significantly improves the overall recognition performance.

[0053] Embodiment 3: This embodiment provides a ship name recognition system based on similarity clustering and temporal consistency analysis, and the system is used to implement the recognition method, including: A dataset construction module, which is used to collect a ship image dataset, label the ship name text area, and mark the remaining areas as ignored areas; A model training module, which is used to train a ship name text detection and recognition model based on the OCR pre-training model using the labeled data; An image acquisition module, which is used to obtain a continuous image sequence of the ship to be recognized; The single-frame recognition module is used to perform text detection and recognition on an image sequence frame by frame, and output candidate text and confidence scores. The rule filtering module is used to filter and obtain a set of valid candidate ship names based on ship name structure rules; The similarity clustering module is used to calculate the similarity of candidate texts, construct a similarity matrix, and perform clustering. The time series analysis module is used to perform time series consistency analysis on clustering results, calculate stability indices, and remove outlier clusters. The result output module is used to select the cluster with the highest stability and output the text with the highest frequency and the best confidence in the cluster as the recognition result.

[0054] Preferably, the model training module uses a pre-trained detection and recognition model of PaddleOCR for fine-tuning, ignoring regions that are not included in the loss calculation.

[0055] Preferably, the similarity clustering module calculates similarity using edit distance and optimizes clustering efficiency using a cluster center mechanism.

[0056] Preferably, the time series analysis module calculates a stability score based on a weighted average of occurrence frequency, time series continuity, and average confidence level.

[0057] Preferably, the image acquisition module supports video frame extraction, image scaling, noise reduction, and enhanced preprocessing.

[0058] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A ship name recognition method based on similarity clustering and temporal consistency analysis, characterized in that, include: S1. Collect a dataset of ship images, and only annotate the text regions and content corresponding to the ship names to obtain annotation data; mark the remaining text regions as ignored regions. S2. OCR-based text detection and recognition pre-trained model: Use labeled data to train the ship name text detection and recognition model; S3. Obtain a continuous image sequence of the vessel to be identified; S4. Perform text detection and recognition on each frame of the image sequence to obtain candidate texts and their corresponding confidence levels; S5. Based on the preset ship name structure rules, the candidate texts are filtered to obtain a set of valid candidate ship names; S6. Calculate the string similarity of the valid candidate ship names and construct a similarity matrix. Based on the similarity matrix, perform clustering to obtain candidate text clusters. S7. Perform consistency analysis on the clustering results based on time series, calculate stability index and remove outlier clusters; S8. Select the candidate text cluster with the highest stability and output the text with the highest frequency in the cluster as the ship name recognition result; if there are multiple texts with the same frequency, select the text with the highest confidence.

2. The ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 1, characterized in that, Step S1 also includes: performing effective image screening and watermark cropping preprocessing on the collected ship images.

3. The ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 1, characterized in that, In step S2, PaddleOCR's pre-trained text detection and text recognition models are fine-tuned to train the ship name detection model and the ship name recognition model respectively. During the fine-tuning training process using ship name image inputs, data augmentation strategies are employed to improve model robustness. For text labeled as an ignored region, it is not included in the loss calculation during model training.

4. The ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 1, characterized in that, In step S3, continuous images containing the ship's hull and name area are collected using a fixed monitoring camera, shipborne high-definition equipment, or airborne inspection equipment; the video stream is framed at preset time intervals to obtain an image sequence.

5. The ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 1, characterized in that, The ship name structure rules mentioned in step S5 include at least one of the following: character type rules, length rules, and format rules.

6. The ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 1, characterized in that, In step S6, string similarity is calculated using an edit distance-based method. The similarity calculation formula is as follows: ； in, Representing text and Edit distance, Indicates the text length; when the similarity is greater than a preset threshold, candidate texts are grouped into the same cluster.

7. The ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 1, characterized in that, During clustering, a cluster center mechanism is introduced. New texts are only calculated to have similarity with each cluster center. If the similarity meets the threshold, the text is assigned to the corresponding cluster; otherwise, a new cluster is created.

8. The ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 1, characterized in that, The temporal consistency analysis in step S7 includes at least one of the following: the frequency of occurrence of candidate text clusters in consecutive frames, the magnitude of change of candidate texts in the time series, and the change in text similarity between adjacent frames.

9. The ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 1, characterized in that, The formula for calculating the stability score in step S7 is as follows: ； in, Indicates the frequency of occurrence of candidate text clusters. Indicators representing time-series consistency This represents the mean confidence level. , , These are the weight parameters.

10. A ship name recognition method based on similarity clustering and temporal consistency analysis according to claim 9, characterized in that, In step S8, the candidate text clusters are sorted according to their stability scores: ； in, This represents the candidate text cluster number with the highest stability score; From the candidate text clusters The text that appears most frequently is selected as the final recognition result: ； in, Representing text In candidate text clusters The number of times it appears in the text.

11. A ship name recognition system based on similarity clustering and temporal consistency analysis, characterized in that, The system is used to implement the identification method according to any one of claims 1-10, comprising: The dataset building module is used to collect ship image datasets, label the ship name text area, and mark the remaining areas as ignored areas; The model training module is used to train a ship name text detection and recognition model based on an OCR pre-trained model and using labeled data. The image acquisition module is used to acquire a continuous sequence of images of the vessel to be identified. The single-frame recognition module is used to perform text detection and recognition on an image sequence frame by frame, and output candidate text and confidence scores. The rule filtering module is used to filter and obtain a set of valid candidate ship names based on ship name structure rules; The similarity clustering module is used to calculate the similarity of candidate texts, construct a similarity matrix, and perform clustering. The time series analysis module is used to perform time series consistency analysis on clustering results, calculate stability indices, and remove outlier clusters. The result output module is used to select the cluster with the highest stability and output the text with the highest frequency and the best confidence in the cluster as the recognition result.

12. A ship name recognition system based on similarity clustering and temporal consistency analysis according to claim 10, characterized in that, The model training module uses a pre-trained detection and recognition model of PaddleOCR for fine-tuning, ignoring regions that are not included in the loss calculation.

13. A ship name recognition system based on similarity clustering and temporal consistency analysis according to claim 10, characterized in that, The similarity clustering module calculates similarity using edit distance and optimizes clustering efficiency using a cluster center mechanism.

14. A ship name recognition system based on similarity clustering and temporal consistency analysis according to claim 10, characterized in that, The time series analysis module calculates a stability score based on a weighted average of occurrence frequency, time series continuity, and average confidence level.

15. A ship name recognition system based on similarity clustering and temporal consistency analysis according to claim 10, characterized in that, The image acquisition module supports video frame extraction, image scaling, noise reduction, and enhanced preprocessing.

Images