Model training method and device for point of interest name recognition

Abstract

Embodiments of the present application provide a model training method and device for point of interest name recognition. The method comprises: obtaining a plaque image including a sample plaque; inputting the plaque image into a preset initial large model to obtain a plurality of candidate point of interest names of the sample plaque; determining a total reward value of each candidate point of interest name according to a preset multi-dimensional reward function; wherein the multi-dimensional reward function includes at least two dimensions of consistency reward, semantic similarity reward, format specification reward, information richness reward, and length reasonableness reward; performing training processing on the initial large model according to the total reward value of the candidate point of interest name and a preset GRPO algorithm to obtain a target large model; wherein the target large model is used to recognize a point of interest name in a plaque image. The method is used to improve the accuracy of point of interest name recognition.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a model training method and apparatus for point of interest name recognition. Background Technology

[0002] Points of Interest (POI) names are core identifying information in street view maps, POI location annotations, and location services. Their accuracy directly impacts map data quality, retrieval matching precision, and user positioning experience. In large-scale street view data collection and updates, relying on manual annotation to obtain POI names is costly, inefficient, and fails to meet the demands of large-scale, real-time data updates. Therefore, there is an urgent need for efficient and accurate automated identification of POI names in street view images.

[0003] In related technologies, point-of-interest (POI) name recognition schemes mainly fall into two categories. One type first extracts text candidates from street view images using Optical Character Recognition (OCR) technology, then uses a deep learning model to correct, sort, or generate the candidate text to obtain the POI name. However, OCR recognition is susceptible to image blurring, occlusion, and noise, leading to missing key text or the introduction of interfering information, resulting in lower accuracy in subsequent name generation. The other type treats POI name recognition as a sequence generation task, directly modeling visual and textual information together to output POI names from street view images. However, this type of model is typically optimized using character or token-level cross-entropy loss, which deviates from the final evaluation metrics such as sequence-level name accuracy. The model is prone to generating fictitious entities based on data distribution, causing name illusion problems and resulting in lower accuracy in POI name recognition.

[0004] In summary, there is an urgent need for a solution that can improve the accuracy of point of interest name recognition. Summary of the Invention

[0005] The model training method and apparatus for point of interest name recognition provided in this application are used to improve the accuracy of point of interest name recognition.

[0006] In a first aspect, embodiments of this application provide a model training method for point-of-interest (POI) name recognition, including:

[0007] Obtain images of the hanging tags, including the sample tags;

[0008] The image of the sign is input into a preset initial large model to obtain multiple candidate point of interest names for the sample sign;

[0009] The total reward value for each candidate point of interest name is determined according to a preset multi-dimensional reward function; wherein, the multi-dimensional reward function includes at least two dimensions from consistency reward, semantic similarity reward, format standardization reward, information richness reward, and length reasonableness reward.

[0010] Based on the total reward value of the candidate point of interest names and the preset GRPO algorithm, the initial large model is trained to obtain the target large model; wherein, the target large model is used to identify the point of interest names in the hanging image.

[0011] In one possible implementation, determining the total reward value for each candidate point of interest name according to a preset multi-dimensional reward function includes:

[0012] For each candidate point of interest name, the sub-reward value of the candidate point of interest name in each dimension is determined according to the multi-dimensional reward function;

[0013] The total reward value for a candidate point of interest is determined based on the sub-reward values ​​of that candidate point of interest name in each dimension.

[0014] In one possible implementation, the multi-dimensional reward function includes a semantic similarity reward dimension, and determining the sub-reward value of the candidate point of interest name in each dimension according to the multi-dimensional reward function includes:

[0015] Based on the candidate point of interest name and the preset name truth value of the sample sign, a first parameter and a second parameter are determined; wherein, the first parameter is used to measure the content coverage of the candidate point of interest name to the name truth value; and the second parameter is used to measure the matching accuracy between the candidate point of interest name and the name truth value.

[0016] The embedding vector of the sign image and the embedding vector of the candidate point of interest name are similar to each other to obtain the third parameter;

[0017] Based on the first parameter, the second parameter, and the third parameter, determine the sub-reward value of the candidate point of interest name in the semantic similarity reward dimension.

[0018] In one possible implementation, the candidate point of interest name includes corresponding inference text information; the multi-dimensional reward function includes a format-standard reward dimension, and determining the sub-reward value of the candidate point of interest name in each dimension according to the multi-dimensional reward function includes:

[0019] Based on the candidate point of interest name and the preset symbol usage rules, a fourth parameter is determined; wherein, the fourth parameter is used to measure the degree of symbol standardization of the candidate point of interest name;

[0020] The fifth parameter is determined based on the inference text information corresponding to the candidate point of interest name; wherein, the fifth parameter is used to measure whether the initial large model performs inference;

[0021] Based on the fourth and fifth parameters, determine the sub-reward value of the candidate point of interest name in the format normativity reward dimension.

[0022] In one possible implementation, determining the sub-reward value of the candidate point of interest name in each dimension according to the multi-dimensional reward function includes:

[0023] When the multi-dimensional reward function includes a consistency reward dimension, the sub-reward value of the candidate point of interest name in the consistency reward dimension is determined based on the candidate point of interest name and the preset truth value of the name on the sample tag; or,

[0024] When the multi-dimensional reward function includes an information richness reward dimension, the sub-reward value of the candidate point of interest name in the information richness reward dimension is determined based on the candidate point of interest name and the preset information category; or,

[0025] When the multi-dimensional reward function includes a length reasonableness reward dimension, the sub-reward value of the candidate point of interest name in the length reasonableness reward dimension is determined based on the character length of the candidate point of interest name.

[0026] In one possible implementation, the candidate interest point names have confidence levels; the step of training the initial large model based on the total reward value of the candidate interest point names and a preset GRPO algorithm to obtain the target large model includes:

[0027] Based on the confidence level of the candidate point of interest names, the candidate point of interest names are grouped to obtain multiple sets of candidate names;

[0028] For each set of candidate names, the relative advantage value of each candidate point of interest name is determined based on the total reward value of each candidate point of interest name in that set.

[0029] The loss function value is determined based on the confidence level of the candidate point of interest name, the relative advantage value of the candidate point of interest name, and the loss function of the preset GRPO algorithm.

[0030] Based on the loss function value, the model parameters of the initial large model are updated in reverse until the updated initial large model meets the preset conditions, thus obtaining the target large model.

[0031] In one possible implementation, the step of grouping the candidate point-of-interest names according to their confidence levels to obtain multiple sets of candidate names includes:

[0032] The uncertainty of the candidate point of interest name is determined based on the confidence level of the candidate point of interest name;

[0033] The candidate point of interest names whose uncertainty is greater than a preset threshold and the candidate point of interest names whose uncertainty is less than or equal to the preset threshold are mixed and grouped according to a preset ratio to obtain multiple sets of candidate names.

[0034] In one possible implementation, the method further includes:

[0035] Acquire street view images;

[0036] The street view image is subjected to target detection processing to obtain the target sign image;

[0037] Based on the target tag image and the target large model, determine the names of the points of interest in the target tag image.

[0038] Secondly, embodiments of this application provide a model training apparatus for point-of-interest (POI) name recognition, comprising:

[0039] The acquisition module is used to acquire images of the hanging tags, including sample hanging tags;

[0040] The processing module is used to input the hanging image into a preset initial large model to obtain multiple candidate interest point names of the sample hanging;

[0041] The determination module is used to determine the total reward value of each candidate point of interest name according to a preset multi-dimensional reward function; wherein, the multi-dimensional reward function includes at least two dimensions from consistency reward, semantic similarity reward, format standardization reward, information richness reward, and length reasonableness reward;

[0042] The training module is used to train the initial large model based on the total reward value of the candidate point of interest names and the preset GRPO algorithm to obtain the target large model; wherein, the target large model is used to identify the point of interest names in the hanging image.

[0043] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0044] The memory stores computer-executed instructions;

[0045] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0046] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0047] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0048] The model training method and apparatus for point-of-interest (POI) name recognition provided in this application acquires images of signs containing sample signs, generates multiple candidate POI names using an initial large model, quantifies the overall quality of each candidate POI name using a multi-dimensional reward function, and completes model training using the GRPO algorithm to obtain the target large model. This approach eliminates reliance on traditional OCR, avoids recognition errors caused by image quality defects, and ensures that the model optimization objective remains consistent with the actual name accuracy index through the multi-dimensional reward function. Simultaneously, the GRPO algorithm enables stable and efficient model updates, effectively suppressing name illusion, and significantly improving the accuracy and robustness of POI name recognition in sign images. This meets the practical needs of automated recognition and updating of large-scale street scene data. Attached Figure Description

[0049] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0050] Figure 1 A flowchart illustrating the model training method for point-of-interest (POI) name recognition provided in this application. Figure 1 ;

[0051] Figure 2 A flowchart illustrating the model training method for point-of-interest (POI) name recognition provided in this application. Figure 2 ;

[0052] Figure 3 Flowchart of the Point of Interest Name Recognition Method Provided in this Application Figure 1 ;

[0053] Figure 4 A schematic diagram of the structure of the model training device for point of interest name recognition provided in this application;

[0054] Figure 5 A schematic diagram of the structure of the electronic device provided in this application.

[0055] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0056] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0057] In related technologies, the identification methods for point of interest names are mainly divided into two categories: two-stage feature fusion and end-to-end generation models.

[0058] (1) Two-stage feature fusion method

[0059] Phase 1: OCR Text Recall

[0060] This phase aims for high recall (retrieving as many text fragments as possible related to the real Points of Interest (POIs) names) by extracting as many relevant text fragments as possible from the images. Deep learning models (such as BERT) are used to extract semantic features from the text, and ensemble learning models such as Light Gradient Boosting Machine (LightGBM) are combined to fuse multi-source heterogeneous features, ensuring that no real POI text is missed.

[0061] Phase Two: Precise Candidate Ranking

[0062] The text results recalled in the first stage are arranged and combined to generate a set of candidate names; then, manually designed numerical features (such as text lines, coordinates and size information of the sign) and BERT semantic features are combined, and sorting and classification are completed by LightGBM. The combination with the highest confidence is selected as the final POI name.

[0063] However, this approach breaks down the task into two separate steps: OCR recall and accurate ranking, leading to a vicious cycle of "information flow disruption." In the first stage, the system relies solely on the OCR engine for text extraction. However, the OCR engine itself has serious limitations; it cannot distinguish between key entities and irrelevant noise, nor can it assess the reliability of its own recognition results. This directly results in the fatal weakness of "relying on high-quality OCR recall." When the input image is blurry or occluded, the OCR may miss key text or incorrectly identify a large amount of irrelevant text. These recalled low-quality candidate texts constitute a huge, noisy candidate set. The ranking model in the second stage is forced to search within this vast space of invalid combinations, resulting in exponentially increasing computational complexity and extreme susceptibility to noise interference. Ultimately, it struggles to find the correct name combination, leading to low accuracy in POI names.

[0064] (2) End-to-end generative model approach

[0065] This approach treats POI name generation as a sequence generation task, using a unified generative model to complete the entire process. A typical example is the VL-Bert model, which can simultaneously fuse visual image and textual information to directly output the target POI name from street view images.

[0066] However, in this approach, the model typically aims to minimize the label-level cross-entropy loss, and the optimization process is based on the prediction of each character or token. However, for POI name generation tasks, the final evaluation metric is the accuracy of the entire name sequence (such as F1-score), leading to a target misalignment problem. This gap between "label-level" optimization and "sequence-level" evaluation metrics prevents the model from effectively learning the ultimate goal of "generating a complete, correct, and non-illusionary name" during training. Therefore, when the input image information is incomplete, the model lacks sufficient constraints and is prone to "imagining" non-existent entities based on patterns in the training data, thus producing "illusions."

[0067] In summary, the accuracy of POI name recognition in related technologies is relatively low.

[0068] Therefore, based on an in-depth analysis of the limitations of the two-stage method and the end-to-end generation model, the inventors realized that to fundamentally improve the accuracy and reliability of POI name recognition, it is necessary to get rid of the strong dependence on OCR recognition results and avoid the error transmission caused by low-quality text candidates. It is also necessary to build an optimization mechanism that is highly consistent with the actual task objectives, so that the model can learn and update around the overall quality of the name. To address this, the inventors proposed using a multimodal initial large model to directly generate multiple candidate POI names from the sign image, eliminating the reliance on OCR for text candidates and reducing the impact of image quality defects on recognition results from the source. Simultaneously, a multi-dimensional reward function is constructed, incorporating consistency, semantic similarity, format standardization, information completeness, and length rationality, to quantitatively evaluate the overall quality of each candidate name, forming a precise supervisory signal matching the sequence-level name accuracy index. Furthermore, the GRPO algorithm is combined to stably optimize the initial large model based on the total reward value of each candidate name, continuously enhancing the model's ability to generate high-quality, high-confidence names during training. This effectively suppresses low-quality outputs such as fictitious, non-standard, and missing information, thus simultaneously solving the error accumulation problem caused by traditional methods relying on OCR, as well as the issues of misaligned optimization targets and the potential for illusions in end-to-end generation models. This significantly improves the accuracy and practicality of POI name recognition.

[0069] The execution subject of this application embodiment can be an electronic device with processing capabilities, such as a computer, server, etc., and this application embodiment is not limited thereto.

[0070] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0071] Figure 1 A flowchart illustrating the model training method for point-of-interest (POI) name recognition provided in this application. Figure 1 ,like Figure 1 As shown, the method includes:

[0072] S101. Obtain the image of the hanging tag, including the sample hanging tag.

[0073] For example, sample signage refers to physical signs, logos, or other signs that contain real POI name information in a street view scene.

[0074] The image containing the tag is an image that can be obtained from street view capture devices, public image datasets, or locally stored data, and is used to provide real visual input for model training.

[0075] In one example, an electronic device can extract an image region containing only the sample sign by performing preprocessing operations such as object detection and region cropping on a street view image, thereby reducing the interference of irrelevant background on model training.

[0076] S102. Input the image of the hanging sign into the preset initial large model to obtain the names of multiple candidate points of interest for the sample hanging sign.

[0077] For example, the preset initial large model is a multimodal large model with image-text joint understanding and sequence generation capabilities, which can directly extract visual features and generate text from the sign image. It should be noted that the embodiments of this application do not limit the type of large model, such as a general vision-language pre-trained (Visual-Linguistic BERT, VL-BERT) model.

[0078] Candidate point of interest names refer to multiple alternative names generated by the initial large model based on the sign image, used to represent the entity identifier corresponding to the sign. It should be noted that the embodiments of this application do not limit the number of candidate point of interest names.

[0079] In one example, an electronic device can input a sign image into an initial large model, which then generates a sequence based on the text, layout, and visual features in the image, outputting multiple candidate point of interest names with different descriptions and combinations.

[0080] S103. Determine the total reward value for each candidate point of interest name based on the preset multi-dimensional reward function.

[0081] For example, a multi-dimensional reward function refers to a function that quantifies and scores the quality of candidate point-of-interest (POI) names from multiple evaluation dimensions, including at least two dimensions such as consistency reward, semantic similarity reward, format conformity reward, information richness reward, and length reasonableness reward. Specifically, the consistency reward measures the degree of matching between the candidate POI name and the real name corresponding to the sample sign, reflecting the overall consistency between the candidate POI name and the real name. The semantic similarity reward measures the semantic similarity between the candidate POI name and the real name, reflecting the effectiveness of the candidate POI name in restoring the content and meaning of the real name. The format conformity reward measures whether the candidate POI name conforms to preset specifications in terms of symbol usage, structural composition, and expression, representing the legality and compliance of the name. The information richness reward measures the degree to which the candidate POI name contains key information in the real name, reflecting whether core content such as brand, category, location, and branches is complete and without omissions. The length reasonableness reward measures whether the character length of the candidate POI name is within a reasonable range, avoiding names that are too long or too short, thus failing to conform to the conventional expression habits of real signs.

[0082] The total reward value is a comprehensive evaluation value obtained by weighting and integrating the scores of each dimension, and is used to characterize the degree of fit between the candidate point of interest name and the real name.

[0083] In one example, for each candidate point of interest name, according to the mapping relationship of the preset multi-dimensional reward function, its sub-reward values ​​in at least two dimensions such as consistency, semantic similarity, format standardization, information completeness, and length reasonableness are calculated. Then, the sub-reward values ​​are weighted and summed to obtain the total reward value corresponding to the candidate point of interest name.

[0084] S104. Based on the total reward value of the candidate interest point names and the preset GRPO algorithm, train the initial large model to obtain the target large model.

[0085] For example, the Group Relative Policy Optimization (GRPO) algorithm is a reinforcement learning optimization algorithm based on the relative advantages of grouping, which can combine reward values ​​to achieve stable updates of model parameters.

[0086] Large target models are used to identify points of interest (POIs) in sign images. A large target model is a model that, after training and convergence, can directly perform end-to-end identification of unknown sign images and output accurate POI names.

[0087] In one example, the candidate interest point names are grouped based on the total reward value of the candidate interest point names and the confidence of the candidate interest point names output by the initial large model. Then, the relative advantage of the group is determined according to the GRPO algorithm and the total reward value of each candidate interest point name. The loss function corresponding to the GRPO algorithm is constructed. The parameters of the initial large model are updated in reverse using the loss function. After the model converges, a target large model that can be used for actual recognition scenarios is obtained.

[0088] The model training method for point-of-interest (POI) name recognition provided in this application acquires images of signs containing sample signs, generates multiple candidate POI names using an initial large model, quantifies the overall quality of each candidate POI name using a multi-dimensional reward function, and completes model training using the GRPO algorithm to obtain the target large model. This approach eliminates reliance on traditional OCR, avoids recognition errors caused by image quality defects, and ensures that the model optimization objective remains consistent with the actual name accuracy index through the multi-dimensional reward function. Simultaneously, the GRPO algorithm enables stable and efficient model updates, effectively suppressing name illusion, and significantly improving the accuracy and robustness of POI name recognition in sign images. This method can meet the practical needs of automated recognition and updating of large-scale street scene data.

[0089] Figure 2A flowchart illustrating the model training method for point-of-interest (POI) name recognition provided in this application. Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, a detailed description of the model training method for point-of-interest (POI) name recognition is provided. This method includes:

[0090] S201. Obtain the image of the hanging tag, including the sample hanging tag.

[0091] It should be noted that this step is similar to the aforementioned step S101, and will not be repeated here.

[0092] S202. Input the image of the hanging sign into the preset initial large model to obtain the names of multiple candidate points of interest for the sample hanging sign.

[0093] It should be noted that this step is similar to the aforementioned step S102, and will not be repeated here.

[0094] S203. For each candidate point of interest name, determine the sub-reward value of the candidate point of interest name in each dimension according to the multi-dimensional reward function.

[0095] For example, the sub-reward value refers to the quantitative score obtained by the candidate point of interest name under a single evaluation dimension.

[0096] In one example, the multi-dimensional reward function includes the calculation logic for the sub-reward value of each dimension. Then, for each candidate point of interest name, the calculation logic for the sub-reward value of the candidate point of interest name in the at least two evaluation dimensions is performed independently to obtain the sub-reward value of the candidate point of interest name in the at least two evaluation dimensions.

[0097] In some possible implementations, when the multi-dimensional reward function includes a semantic similarity reward dimension, the first and second parameters are determined based on the candidate point of interest name and the preset name truth value of the sample sign; the similarity calculation is performed on the embedding vector of the sign image and the embedding vector of the candidate point of interest name to obtain the third parameter; and the sub-reward value of the candidate point of interest name in the semantic similarity reward dimension is determined based on the first, second, and third parameters.

[0098] The first parameter measures the content coverage of candidate point-of-interest (POI) names to ground truth names, reflecting the inclusion of key information in the real names. The second parameter measures the matching accuracy between candidate POI names and ground truth names, reflecting the accuracy of effective content within the candidate POI names. Ground truth names refer to the real, standard POI names corresponding to the sample signs, serving as a benchmark for evaluating candidate POI names. The embedding vector of the sign image is a vector representation obtained after feature encoding of the sign image, used to characterize the visual semantic features of the image. The embedding vector of the candidate POI name is a vector representation obtained after text encoding of the candidate POI name, used to characterize the textual semantic features of the name. The third parameter is the similarity value between the embedding vector of the sign image and the embedding vector of the candidate POI name, used to characterize the semantic matching degree between the image and text. The sub-reward value of the semantic similarity reward dimension is a score calculated based on the above three parameters, used to characterize the quality score of the candidate POI name in the semantic similarity dimension.

[0099] Specifically, text matching calculations are performed between candidate point of interest names and their ground truth values ​​to obtain a first parameter representing the degree of content coverage and a second parameter representing the accuracy of matching. Embedding vectors of the sign image and candidate point of interest names are extracted using a feature encoding model, and a third parameter is obtained by calculating the cosine similarity between the two embedding vectors. The first, second, and third parameters are then weighted and fused to obtain the sub-reward value corresponding to the candidate point of interest name in the semantic similarity reward dimension.

[0100] For example, the first parameter can be determined based on Recall-Oriented Understudy for Gisting Evaluation (ROUGE). Understandably, ROUGE can be used to measure the extent to which the generated text covers the information in the reference text (i.e., recall). First parameter It can be represented as:

[0101] ;

[0102] Among them, ROUGE-1 is a recall metric based on unigrams, which calculates the proportion of single words matching the real name in the candidate point of interest name to the unigrams of the real name, reflecting the coverage of the candidate point of interest name with the basic vocabulary of the real name; ROUGE-2 is a recall metric based on bigrams, which calculates the proportion of two consecutive word combinations matching the real name in the candidate point of interest name to the bigrams of the real name, reflecting the coverage and coherence of the phrase-level semantics of the real name by the candidate point of interest name; ROUGE-L is a recall metric based on the longest common subsequence (LCS), which calculates the recall based on the length of the longest common subsequence between the candidate point of interest name and the real name, reflecting the degree of fit between the candidate point of interest name and the real name in terms of overall structure and semantic flow. , , The weighting coefficients are used to adjust the contribution of each sub-indicator to the overall score. They can be preset according to task requirements or learned through data, for example, all can be 1 / 3. This weighted fusion formula integrates recall indicators from three levels: unary words, bigrams, and longest common subsequence. It comprehensively measures the content coverage of candidate point-of-interest names to the real names from three dimensions: vocabulary, phrase, and overall structure. This avoids the limitations of a single indicator focusing only on local matching, making the quantification of the first parameter more accurate and comprehensive. It provides a more reliable basis for evaluating content coverage for the multi-dimensional reward function, thereby improving the accuracy of judging the quality of candidate point-of-interest names during model training and helping to guide the model to generate listing names that better match the key information of the real names.

[0103] Furthermore, the second parameter can be determined based on the Bilingual Evaluation Understudy (BLEU). Understandably, BLEU can be used to measure the accuracy of the generated text against the reference text. The second parameter... It can be represented as:

[0104] ;

[0105] Among them, BLEU-1 is a precision metric based on unigrams, which calculates the proportion of single words matching the real name in the candidate point of interest names out of the unigrams in the candidate point of interest names, reflecting the accuracy of the basic vocabulary in the candidate point of interest names; BLEU-2 is a precision metric based on bigrams, which calculates the proportion of two consecutive word combinations matching the real name in the candidate point of interest names out of the bigrams in the candidate point of interest names, reflecting the accuracy and coherence of phrase-level expressions in the candidate point of interest names; BLEU-3 is a precision metric based on trigrams, which calculates the proportion of three consecutive word combinations matching the real name in the candidate point of interest names out of the trigrams in the candidate point of interest names, further refining the precision of phrase-level expressions; BLEU-4 is a precision metric based on 4-grams, which calculates the proportion of four consecutive word combinations matching the real name in the candidate point of interest names out of the 4-grams in the candidate point of interest names, reflecting the matching accuracy of the candidate point of interest names on longer semantic segments; , , , These are weighting coefficients used to adjust the contribution of each sub-indicator to the overall score. They can be preset according to task requirements or obtained through data learning, for example... , , , The values ​​can be 0.1, 0.1, 0.2, and 0.6 respectively. This weighted fusion formula integrates precision metrics from four levels: unary, bigram, trigram, and quaternary words. It comprehensively measures the matching accuracy between candidate point-of-interest names and real names from multiple dimensions, including vocabulary, phrases, and longer semantic fragments. This avoids the limitations of a single metric focusing only on local matching, making the quantification of the second parameter more accurate and comprehensive. It provides a more reliable basis for evaluating matching accuracy for the multi-dimensional reward function, thereby improving the accuracy of judging the quality of candidate point-of-interest names during model training. This helps guide the model to generate signage names that more accurately reflect the real names, effectively suppressing the generation of fictitious or erroneous names.

[0106] For the third parameter, the cosine similarity between the embedding vector of the tag image and the embedding vector of the candidate point of interest name can be calculated using the Contrastive Language–Image Pre-training (CLIP) model to obtain the third parameter. .

[0107] Furthermore, the first, second, and third parameters can be weighted and fused according to the following formula to obtain the sub-reward value corresponding to the semantic similarity reward dimension of the candidate interest point name. :

[0108] ;

[0109] in, , , These are weighting coefficients used to adjust the contribution of each sub-indicator to the overall score. They can be preset according to task requirements or obtained through data learning, for example... =0.3, =0.3, =0.4.

[0110] This approach integrates three dimensions: text content coverage, text matching accuracy, and image-text semantic matching. It comprehensively quantifies the semantic similarity of candidate point-of-interest names from two levels: intra-text matching and cross-modal image-text matching. This avoids the one-sidedness of single-dimensional evaluation and makes the sub-reward values ​​of semantic similarity reward more accurate and comprehensive. It provides a more reliable semantic-level evaluation basis for multi-dimensional reward functions, thereby improving the accuracy of judging the quality of candidate point-of-interest names during model training. This helps guide the model to generate sign names that not only fit the key information of real names but also have a high degree of consistency with the visual information of images, effectively improving the semantic accuracy and robustness of sign name recognition.

[0111] In some possible implementations, the candidate point of interest name includes the corresponding inference text information; the multi-dimensional reward function includes a format normative reward dimension, and a fourth parameter is determined based on the candidate point of interest name and the preset symbol usage rules; a fifth parameter is determined based on the inference text information corresponding to the candidate point of interest name; and a sub-reward value of the candidate point of interest name in the format normative reward dimension is determined based on the fourth and fifth parameters.

[0112] The reasoning text information refers to the reasoning process, explanations, or logical judgments generated by the initial large model during the generation of candidate point-of-interest (POI) names. This text characterizes whether the model makes reasonable inferences based on image content. The format standardization reward dimension quantifies the standardization of candidate POI names in terms of expression and generation logic from two aspects: symbol usage and reasoning rationality. Preset symbol usage rules refer to the pre-defined usage rules for punctuation, parentheses, separators, and special characters in the nameplate, constraining the format legality of the names. The fourth parameter measures the standardization of symbols in the candidate POI names, reflecting whether the symbol usage conforms to the preset rules and whether there are garbled characters, redundant symbols, or missing necessary symbols. The fifth parameter measures whether the initial large model performs reasonable reasoning, characterizing whether the model generates names based on logical deduction from image visual content rather than arbitrarily. The sub-reward value of the format standardization reward dimension is a quantitative score obtained by comprehensively calculating the fourth and fifth parameters, used to comprehensively characterize the overall standardization of candidate POI names in terms of format and reasoning.

[0113] Specifically, the symbols in the candidate point of interest names are detected to determine whether they conform to the preset symbol usage rules, thus obtaining the fourth parameter of the standardization degree of the corresponding symbols. At the same time, the reasoning text information corresponding to the candidate point of interest names is parsed to determine whether the model makes reasonable inferences based on the content of the hanging image, thus obtaining the fifth parameter of the reasonableness of the corresponding inferences. The fourth and fifth parameters are weighted and fused to obtain the sub-reward value corresponding to the format standardization reward dimension of the candidate point of interest name.

[0114] For example, the fourth parameter can be obtained by parsing the candidate point of interest names according to the following correspondence. :

[0115] ;

[0116] Among them, standard parentheses refer to the standard parenthesis form used to distinguish core names from additional information (such as branches and regions), namely parentheses "()"; non-standard separation methods refer to other symbols or expressions used to separate core names from additional information, such as square brackets "[]" or spaces; no parentheses are used at all, which means that no parenthesis symbols are used to separate information in the candidate point of interest name, and it is presented in plain text form.

[0117] This segmented assignment formula precisely quantifies the standardization of bracket usage in candidate point-of-interest (POI) names, categorizing bracket usage into standard, non-standard, and no bracket usage, and assigning differentiated scores. This forms a clear quantitative basis for evaluating bracket usage in symbol standardization assessment. This approach makes the calculation of the fourth parameter of the format standardization reward dimension more targeted and operable, accurately identifying non-standard bracket usage in candidate POI names. It provides a refined symbol standardization evaluation index for the multi-dimensional reward function, guiding the model to learn the standard bracket usage habits of sign names during training, effectively improving the format standardization of generated candidate POI names, and making the final output POI names more consistent with the actual expression of street view signs.

[0118] Furthermore, the inference text information corresponding to the candidate point of interest names can be parsed according to the following correspondence to obtain the fifth parameter. :

[0119] ;

[0120] Here, "generated text" refers to the overall inference text information generated synchronously when the initial large model outputs candidate interest point names; "inference label" refers to a specific text identifier pre-defined for the model to identify its executed inference logic, such as "Inference basis:", "Derivation logic:", etc. <think>…think>” etc. Using this binary judgment formula, the fifth parameter is obtained based on whether the generated text corresponding to the candidate point of interest name contains a preset inference label, thus completing the quantitative judgment of whether the initial large model performs inference. This method achieves a simple, efficient, and unambiguous quantitative calculation of the fifth parameter, providing a clear and practical basis for judging inference behavior for the format standardization reward dimension. It can clearly distinguish whether the model generates names based on logical reasoning or generates them arbitrarily without basis, guiding the model to form output habits of generating names and labeling inference during training, avoiding illogical fictitious generation, further improving the evaluation system of format standardization, and enhancing the logical rationality and interpretability of the model's generated nameplates.

[0121] Furthermore, the fourth and fifth parameters can be weighted and fused according to the following formula to obtain the sub-reward value corresponding to the format normativity reward dimension of the candidate point of interest name. :

[0122] ;

[0123] in, It is a weighting coefficient with a value range of [0,1]. It is used to adjust the contribution ratio of the fifth parameter in the overall score. It can be preset or adaptively adjusted according to the actual needs of the sign name recognition task. For example, it can be set to 0.3.

[0124] This weighted fusion formula integrates the fifth parameter, which measures the model's reasoning behavior, with the fourth parameter, which measures the standardization of symbol usage, using normalized weighting to obtain a comprehensive format standardization score. This score is then used as a sub-reward value for the format standardization reward dimension, achieving an integrated quantitative evaluation of the format standardization of candidate point of interest names. This effectively guides the model to consider both the standardization of symbol usage in the signage names and the rationality of reasoning in the generation process during training, significantly improving the format regularity of the final generated signage names and the interpretability of the generation logic.

[0125] In some possible implementations, when the multi-dimensional reward function includes a consistency reward dimension, the sub-reward value of the candidate point of interest name in the consistency reward dimension is determined based on the candidate point of interest name and the preset name truth value of the sample listing.

[0126] In this context, "name truth value" refers to the standard, authentic point-of-interest (POI) name pre-labeled on the sample listings, serving as the sole reference standard for comparing the consistency of candidate POI names. The sub-reward value of the consistency reward dimension is a quantitative score calculated based on the degree of matching between the candidate POI name and the name truth value, directly reflecting the degree of correspondence between the candidate POI name and the authentic name. Specifically, text matching calculations are performed between the candidate POI name and the name truth value, and the sub-reward value corresponding to that candidate POI name in the consistency reward dimension is obtained based on the character matching ratio, semantic overlap, or complete match status.

[0127] For example, name truth values ​​can include an optimal name and zero or more acceptable names. The optimal name represents the standard and most accurate name truth value for labeling, while acceptable names represent semantically equivalent and reasonably expressed variant names. The hit rate between candidate point-of-interest names and name truth values ​​can be determined according to the following correspondence, yielding sub-reward values ​​for the consistency reward dimension. :

[0128] ;

[0129] Where p represents the name of the candidate point of interest; C represents the optimal name in the true name set; C represents the acceptable name. This hierarchical scoring formula uses the optimal name as the core reference while also accommodating acceptable and reasonable name variations. Differentiated scores are used to accurately quantify the consistency of candidate point-of-interest (POI) names. This approach strictly constrains the core accuracy of candidate POI names relative to the true names while avoiding excessive penalties for reasonable variations due to differences in expression. This improves the flexibility and robustness of consistency evaluation, provides more accurate supervision signals for model training that fit real-world application scenarios, and guides the model to generate signage names that accurately match the true names while possessing reasonable expressive flexibility, effectively balancing the rigor and practicality of name recognition.

[0130] When the multi-dimensional reward function includes a consistency reward dimension, the candidate point of interest (POI) names are compared for consistency with the real names by using the preset true names of the sample listings as a reference, and the corresponding sub-reward values ​​are calculated. This achieves direct quantification of the degree to which candidate POI names match the real names. This method provides the most intuitive accuracy supervision signal for model training, quickly distinguishing between candidate POI names that match or deviate from the real names. It allows the model to prioritize name generation methods that are highly consistent with the real names during training, improving the accuracy of the model's output listing names from the source. This provides fundamental and crucial accuracy evaluation support for the multi-dimensional reward function.

[0131] In some possible implementations, when the multi-dimensional reward function includes an information richness reward dimension, the sub-reward value of the candidate point of interest name in the information richness reward dimension is determined based on the candidate point of interest name and the preset information category.

[0132] The information richness reward dimension measures the degree to which candidate point-of-interest names contain the core information of the corresponding entity, and is the evaluation dimension representing the comprehensiveness of name information in the multi-dimensional reward function. The preset information category refers to the core information type pre-defined for the listed name, used to define the scope of key information that the name must contain. The sub-reward value of the information richness reward dimension is a quantitative score calculated based on the result of the candidate point-of-interest name matching the preset information category, used to reflect the completeness of the information coverage of the candidate point-of-interest name. Specifically, text parsing and information recognition are performed on the candidate point-of-interest name to determine whether it contains the preset information category. A quantitative score is then calculated based on the information category matching result to obtain the sub-reward value corresponding to the information richness reward dimension for that candidate point-of-interest name.

[0133] For example, the aforementioned information categories may include multiple subcategories such as proprietary brands (e.g., traditional brands like Li Ji and Xie Ji), business scope (e.g., stone pot dishes, beef restaurants, noodle shops, stir-fries), and branch names (e.g., Gaoxin branch, Qingbaijiang branch, xx Road branch). The electronic device can judge the candidate point of interest names according to the following correspondence to determine whether they hit each subcategory in the preset information categories. Then, it can perform a weighted sum based on the hit status of each subcategory to obtain the sub-reward value corresponding to the information richness reward dimension of the candidate point of interest name. .

[0134] ;

[0135] ;

[0136] ;

[0137] ;

[0138] In other words, the existence of subcategories such as proprietary brands, business scope, and branch names in candidate point-of-interest names can be determined using the above formula. Then, the determination results can be averaged to obtain the sub-reward value corresponding to the information richness reward dimension of the candidate point-of-interest name. .in, The brand presence score is a binary indicator used to measure whether candidate point of interest names contain proprietary brand information. A proprietary brand refers to the exclusive brand identity of the entity to which the sign is located. The business scope existence score is a binary indicator used to measure whether the candidate point of interest name contains business scope information. Business scope refers to the business type or service scope of the listed entity. The branch name existence score is a binary indicator used to measure whether candidate point of interest names contain branch name information. A branch name refers to the specific store identifier of the corresponding entity.

[0139] By binarizing the existence of three core information categories—brand, business scope, and branch names—and then averaging the results to obtain an information richness score, this method achieves a precise quantitative evaluation of the information completeness of candidate point-of-interest (POI) names. This approach decomposes information completeness into three core dimensions, accurately identifying whether candidate POI names contain key information and comprehensively reflecting the comprehensiveness of information coverage through averaging. This avoids excessive penalties caused by missing a single dimension, providing precise information-level monitoring signals for the multi-dimensional reward function. It effectively distinguishes between POI names with complete and missing information, providing information-level monitoring signals for model training. This guides the model to retain key information when generating names, avoiding the omission of core information, further enriching the evaluation dimensions of the multi-dimensional reward function, and improving the information completeness and practical application value of the model's output signage names.

[0140] In some possible implementations, when the multi-dimensional reward function includes a length reasonableness reward dimension, the sub-reward value of the candidate point of interest name in the length reasonableness reward dimension is determined based on the character length of the candidate point of interest name.

[0141] The length reasonableness reward dimension measures whether the character length of candidate point-of-interest (POI) names conforms to the standard expression of sign names in street view scenarios. It is an evaluation dimension representing the compliance of name length in the multi-dimensional reward function. Character length refers to the total number of Chinese characters, letters, numbers, and standard punctuation marks contained in the candidate POI name, not the number of tokens. The sub-reward value of the length reasonableness reward dimension is a quantitative score calculated based on the degree of matching between the character length of the candidate POI name and a preset reasonable length range, reflecting the reasonableness of the name length. Specifically, the character length corresponding to the candidate POI name is determined, compared with the preset reasonable length range, and the sub-reward value corresponding to the length reasonableness reward dimension is calculated based on the comparison result.

[0142] For example, the sub-reward value corresponding to the length reasonableness reward dimension of the candidate point of interest name can be determined according to the following correspondence. :

[0143] ;

[0144] Here, len(s) represents the character length of the candidate point of interest name s. This method effectively constrains the length of the candidate point of interest names to a reasonable range that conforms to practical applications, avoiding information redundancy caused by excessively long names or the loss of key information due to excessively short names. It further enriches the evaluation dimensions of the multi-dimensional reward function, provides length-level supervision signals for model training, guides the model to generate point of interest names with regular lengths that conform to the conventional expression habits of street view sign names, and improves the standardization and practicality of the model output results.

[0145] S204. Determine the total reward value of the candidate point of interest name based on the sub-reward value of the candidate point of interest name in each dimension.

[0146] For example, the total reward value is the overall value obtained by weighted fusion of the sub-reward values ​​of each dimension, used to comprehensively characterize the overall quality and credibility of the candidate point of interest names. Specifically, at least two dimension sub-reward values ​​corresponding to each candidate point of interest name are weighted and summed, and the result of the weighted summation is used as the total reward value of the candidate point of interest name. It should be noted that the embodiments of this application do not limit the number of dimensions; taking a sub-reward value including 5 dimensions as an example, the total reward value is... It can be represented as:

[0147] ;

[0148] in, , , , , These are the weighting coefficients corresponding to the sub-reward values ​​of each dimension, used to adjust the contribution of each evaluation dimension to the total reward value. They can be preset according to task requirements or obtained through data learning, for example, they can be set to 0.35, 0.30, 0.10, 0.05, and 0.20 respectively.

[0149] By introducing five dimensions—consistency reward, semantic similarity reward, format conformity reward, length appropriateness reward, and information richness reward—this approach systematically addresses the generation bias problem caused by traditional methods relying solely on a single signal (such as OCR confidence or language model probability). Specifically, the consistency reward ensures that the generated name is logically consistent with the standard answer type; the semantic similarity reward uses embedding space metrics (such as CLIP) to determine the semantic matching degree between candidate POI names and image content or context, effectively suppressing illusions; the format conformity reward forces the output to conform to platform naming conventions (such as disabling special symbols and unifying the order of Chinese and English characters) through regularity rules or finite state machines; the length appropriateness reward avoids generating names that are too short (such as "shop") or redundantly long (such as including the full address); and the information richness reward encourages the model to include key distinguishing information in the name. These five dimensions together constitute an interpretable, configurable, and business-aligned evaluation loop, ensuring that the generated results are not only accurate but also directly meet the map data entry standards, thereby significantly reducing the manual review rate and improving the end-to-end automation level. Meanwhile, since all scores can be achieved by rule bases, lightweight embedded models, or statistical metrics, there is no need to train an independent reward model, which greatly reduces system complexity and deployment costs.

[0150] S205. Based on the confidence level of the candidate point of interest names, group the candidate point of interest names to obtain multiple sets of candidate names.

[0151] For example, the confidence score is the output probability value when the initial large model generates the corresponding candidate interest point name, which reflects the model's certainty about the candidate interest point name. The candidate name set refers to the grouping result formed after dividing multiple candidate interest point names according to preset rules.

[0152] In one example, an electronic device can determine the corresponding uncertainty based on the confidence level of the candidate point of interest name, and mix the candidate point of interest names with high uncertainty and low uncertainty into the same group according to a preset ratio to obtain multiple sets of candidate names.

[0153] Specifically, the uncertainty of candidate point of interest names is determined based on the confidence level of the candidate point of interest names; the candidate point of interest names with uncertainty greater than a preset threshold and the candidate point of interest names with uncertainty less than or equal to the preset threshold are mixed and grouped according to a preset ratio to obtain multiple sets of candidate names.

[0154] Uncertainty is a quantitative indicator derived from confidence level, used to characterize the degree of uncertainty in the model's generation of the candidate point of interest name. The preset threshold is a pre-defined uncertainty threshold used to distinguish between candidate point of interest names with high and low uncertainty. The preset ratio is the proportion of high-uncertainty to low-uncertainty candidate point of interest names in the mixed grouping. The candidate name set is the grouping result formed by mixing candidate point of interest names with different uncertainties according to the preset ratio; each group contains candidate point of interest names with different uncertainty levels.

[0155] For example, the confidence level of a candidate point of interest name can be converted into the uncertainty level of the candidate point of interest name using the following formula:

[0156] ;

[0157] in, Indicates the name of the candidate point of interest. The uncertainty is used to quantify the degree of uncertainty when the model generates the name. This indicates the generation of candidate interest point names under a large model with parameter θ, given a tag image input I. The conditional probability is the confidence level of the large model in the candidate interest point name. The higher the probability, the more certain the large model is about the generated result.

[0158] Furthermore, a preset threshold of 0.5 can be set to filter out candidate point of interest names with high uncertainty > 0.5 and candidate point of interest names with low uncertainty ≤ 0.5. Then, they are mixed according to a preset ratio of high:low = 1:1, with each group of 4 candidate point of interest names including 2 high uncertainty and 2 low uncertainty, and finally multiple sets of candidate names are obtained.

[0159] Alternatively, the candidate point of interest names can be sorted in descending order according to their uncertainty, and then two candidate point of interest names with high uncertainty are selected in the first order, and two candidate point of interest names with low uncertainty are selected in the last order. That is, Top-2 + Bottom-2 form a set of candidate names, in order to avoid low-quality samples dominating the mean within the group.

[0160] By employing an uncertainty-aware grouping mechanism that mixes candidate point-of-interest names with high and low uncertainty in a pre-defined ratio, each group of candidate names covers samples with different confidence levels. This preserves the model's learning samples of high-confidence names while incorporating low-confidence samples to be optimized, providing a richer sample distribution for the subsequent GRPO algorithm to calculate the relative advantage value. This allows the model to simultaneously optimize the generated results of high and low confidence during training, improving the stability and comprehensiveness of model updates, further suppressing name illusion, and enhancing the model's ability to learn and optimize candidate point-of-interest names of different qualities.

[0161] S206. For each set of candidate names, determine the relative advantage value of each candidate point of interest name based on the total reward value of each candidate point of interest name in the set of candidate names.

[0162] For example, the relative advantage value refers to the relative superiority or inferiority of different candidate point of interest names within the same group based on the comparison of total reward values, and is used to characterize the quality ranking of candidate point of interest names in the same group.

[0163] In one example, within the same set of candidate names, the total reward values ​​of each candidate point of interest are compared horizontally, and the relative advantage value for each candidate point of interest is calculated based on the difference in total reward values. Specifically, the relative advantage value for each candidate point of interest within the group can be calculated using the following formula.

[0164] ;

[0165] in, This represents the relative advantage value of the i-th candidate point of interest name in the t-th candidate name set; Let represent the total reward value of the i-th candidate point of interest name; mean(r) represents the average total reward value of all candidate point of interest names in the current group, used to characterize the average quality level of candidate point of interest names in this group; std(r) represents the standard deviation of the total reward value of all candidate point of interest names in the current group, used to characterize the dispersion of the quality of candidate names in this group; the relative advantage value, by dividing the difference between the total reward value and the mean within the group by the standard deviation within the group, eliminates the influence of the difference in reward value distribution between different groups, making the comparison of advantage values ​​more fair and consistent. Furthermore, it can accurately distinguish the relative merits of different candidate point of interest names within the same group, guiding the model to prioritize name generation methods with high advantage values ​​during training, improving the targeting and effectiveness of model updates, further suppressing name illusion, and enhancing the model's ability to learn and optimize candidate point of interest names of different qualities.

[0166] S207. Determine the loss function value based on the confidence level of the candidate interest point names, the relative advantage value of the candidate interest point names, and the preset loss function of the GRPO algorithm.

[0167] For example, the loss function of the GRPO algorithm is a function used to measure the difference between the model's generated results and the optimization objective. The loss function value is a quantitative result calculated by the loss function and is used to guide the updating of model parameters.

[0168] In one example, an electronic device can convert the confidence level of candidate point-of-interest (POI) names into uncertainty. From this uncertainty, an uncertainty modulation factor can be derived. The uncertainty modulation factor and relative advantage value of each POI name are then input into the loss function of the GRPO algorithm to calculate the corresponding loss function value. The uncertainty modulation factor dynamically adjusts the contribution weight of each POI name in the loss function based on its uncertainty, achieving differentiated weighting for POI names of different qualities.

[0169] Specifically, the electronic device can first determine the uncertainty of the candidate point of interest name based on the confidence level of each candidate point of interest name according to the aforementioned step S205. Furthermore, the uncertainty is converted into an uncertainty modulation factor according to the following formula:

[0170] ;

[0171] in, Indicates the name of the candidate point of interest. Uncertainty modulation factor; This represents the attenuation coefficient, a hyperparameter greater than 0, used to control the attenuation strength of the uncertainty modulation factor. The larger the value, the more significant the impact of uncertainty on the uncertainty modulation factor; for example, a value of 0.5 can be used.

[0172] In this way, during GRPO training, candidate interest point names with high uncertainty (usually results generated by the model with poor quality and prone to name illusion) are assigned a small uncertainty modulation factor to reduce their impact on the loss function and prevent the model from being misled by low-quality samples; while candidate interest point names with low uncertainty (results generated by the model with high quality and high confidence) are assigned an uncertainty modulation factor close to 1, so that they dominate the training and guide the model to optimize towards high-quality generation.

[0173] Furthermore, the uncertainty modulation factor and relative advantage value of each candidate interest point name are used as inputs and substituted into the loss function corresponding to the GRPO algorithm for calculation, to obtain the corresponding loss function value. For example, this loss function can be expressed as:

[0174] ;

[0175] in, It is a loss function based on the GRPO algorithm, used to guide the parameters of the initial large model. Update; T is the total number of candidate name sets; t is the group index, t∈[1,T]; It is the set of candidate names for group t. is the number of candidate point of interest names in the candidate name set of group t; i is the candidate index within the group, representing the i-th candidate point of interest name in group t; It is the relative advantage value of the name of the i-th candidate point of interest in the t-th group; It is the sampling probability ratio of the name of the i-th candidate point of interest in the t-th group; It is the clipping factor, and clip is the probability ratio clipping function; It is the KL divergence weight (penalty coefficient). The current model Compared with the reference model The KL divergence is used to constrain the model update magnitude.

[0176] Optional, It can be calculated based on the Sigmoid decay function, as shown in the following formula:

[0177] ;

[0178] in, The KL divergence base weights are preset hyperparameters that represent the maximum effectiveness of the KL divergence regularization term. The upper limit. `Sql` is a sigmoid non-linear activation function. Its core function is to map any real number to the (0,1) interval, achieving a smooth transition of weights and avoiding abrupt changes. `K` is a slope hyperparameter used to control the steepness of the sigmoid function's change, determining the rate of weight adjustment. The larger `k` is, the more rapidly the weights change with the number of training steps. `Step` is the total number of training steps, the preset total number of iterations for model training, serving as the time base for weight adjustment. `t` is the current training step, an iteration number variable that increases with the training process, representing the training progress and can be used to control the central position of decay, i.e., when a significant decrease begins. Through this sigmoid-type dynamic weight formula, adaptive adjustment of the KL divergence regularization term weights is achieved. In the early stages of training, the KL divergence constraint is weakened, allowing the model parameters to update rapidly to learn the core generation rules of point of interest names, avoiding excessive regularization that restricts initial convergence. In the later stages of training, the KL divergence constraint is strengthened, strictly limiting the model update amplitude and preventing the model from deviating from the reference model distribution, thus preventing overfitting or name illusion. Meanwhile, the non-linear characteristics of the Sigmoid function ensure a smooth transition of weights, completely avoiding training oscillations caused by fixed or abrupt weight changes. This significantly improves the stability and convergence efficiency of the GRPO algorithm, further optimizes the accuracy and robustness of the model's generated point of interest names, and effectively suppresses the name illusion phenomenon.

[0179] Using the loss function described above, all T candidate name sets are traversed using the group index t, and then all candidates within each group are traversed using the candidate index i within the group. For each candidate, the product of the pruned probability ratio and the relative advantage value is calculated. After combining the uncertainty modulation factor, the KL divergence regularization term is subtracted, and finally the average of the results for all groups and all candidates is taken to obtain the loss function value.

[0180] This approach fundamentally optimizes the stability and efficiency of reinforcement learning training. Compared to traditional proximal policy optimization (PPO), which relies on absolute reward values, in fine-grained tasks like POI name generation, the reward differences between different candidates are small (e.g., 0.88 vs 0.92), leading to unstable policy gradient directions and slow training convergence. This application uses relative advantage within a group (i.e., candidate score minus the group average) as the optimization signal, ensuring that policy updates focus only on "relatively better" samples, insensitive to the absolute scale of rewards, thus improving training stability. Furthermore, low-resolution image generation often involves a large number of low-confidence samples among the candidates. Random grouping would lower the group mean, causing the advantage of truly high-quality samples to be underestimated. To address this, this application proposes an uncertainty-aware grouping mechanism. Candidates are ranked according to the generation uncertainty output by the policy model, and a high-low mixed pairing strategy is used for grouping, ensuring that each group contains both high-confidence and low-confidence samples, preventing low-quality samples from dominating the baseline. Simultaneously, a gradient modulation factor, also known as an uncertainty modulation factor, is designed to automatically attenuate the gradient contribution of low-confidence samples, allowing policy updates to focus on high-confidence, high-scoring samples. Furthermore, a KL constraint with sigmoid decay is introduced, penalizing only when the policy deviates excessively from the reference policy, preventing collapse while preserving exploratory capabilities. These mechanisms work together to result in faster convergence, lower variance, stronger robustness to noisy inputs, and ultimately more reliable and consistent POI names.

[0181] Understandably, through the above steps, the basic GRPO algorithm is improved using an uncertainty-aware grouping mechanism, resulting in an improved GRPO algorithm (e.g., GRPO+). This mechanism divides and groups high- and low-confidence samples based on the generation uncertainty of candidate POI names, thereby optimizing the distribution structure of training samples. Simultaneously, the GRPO loss function is improved based on this grouping pattern. The entire training framework does not require additional training of a separate reward model; it can construct a reliable reward signal solely based on rule scoring or lightweight model scoring based on dimensions such as consistency, semantic similarity, format standardization, length rationality, and information richness. During training, standardized advantage values ​​are calculated based on the relative advantages of candidate POI names within the same group, thus mitigating training fluctuations caused by different sample distributions and significantly improving the stability of the overall training process. Furthermore, through uncertainty-related gradient modulation strategies, the model gradient automatically focuses on high-quality, high-confidence generated samples, weakening the interference of low-quality, low-confidence samples on model updates, ultimately achieving more stable and accurate POI name optimization and generation.

[0182] S208. Based on the loss function value, update the model parameters of the initial large model in reverse until the updated initial large model meets the preset conditions, and obtain the target large model.

[0183] For example, preset conditions refer to training termination conditions such as model convergence, loss function value stabilization, or reaching a preset number of iterations. The target large model refers to the optimized model that, after training, can directly perform end-to-end recognition of the sign image and output accurate point of interest names.

[0184] In one example, backpropagation is performed based on the calculated loss function value to iteratively update the model parameters of the initial large model until the model converges and meets the preset training conditions, ultimately obtaining a target large model that can be used for actual recognition scenarios.

[0185] The model training method for point-of-interest (POI) name recognition provided in this application acquires images of signs containing sample signs, generates multiple candidate POI names using an initial large model, calculates the sub-reward value and total reward value of each candidate POI name using a multi-dimensional reward function, groups the candidates based on their confidence level and calculates their relative advantage value, and iteratively updates the model parameters based on the loss function of the GRPO algorithm, ultimately obtaining the target large model. This method abandons the traditional OCR recognition step, avoiding the error propagation problem caused by image quality defects from the source. By quantifying the overall quality of the names through a multi-dimensional reward function, the model optimization goal is kept consistent with the actual recognition requirements. At the same time, by leveraging the training mechanism of the GRPO algorithm based on confidence grouping and relative advantage comparison, the ability to generate high-reward, high-confidence names is strengthened during training, while the output of low-quality, fictitious names is constrained. This makes the model update process more stable and efficient, effectively suppressing the name illusion problem caused by incomplete image information, and significantly improving the accuracy, standardization, and robustness of POI name recognition in sign images. It can adapt to the practical application needs of automated recognition and updating of large-scale street scene data.

[0186] In summary, this application firstly employs the GRPO reinforcement learning framework, fundamentally solving the problems of "target misalignment" and "information flow fragmentation." It models the entire generation process as a Markov decision process, using a multi-dimensional reward function to quantify the generated candidate interest point names against their truth values ​​across multiple dimensions, including consistency rewards, semantic similarity rewards, format standardization rewards, length reasonableness rewards, and information richness rewards. This ensures that the model's learning objective is completely consistent with the final evaluation metric; that is, the model no longer pursues local optima for each character but strives to maximize cumulative rewards, i.e., generating complete names that best match the truth values, fundamentally avoiding the possibility of generating illusions.

[0187] Secondly, this application introduces a dynamic grouping and pairing mechanism based on uncertainty. This mechanism no longer treats all candidate point-of-interest (POI) names as equivalent, but instead uses the confidence level of the model output to quantify the uncertainty of each candidate PIO name. Furthermore, instead of simple averaging or aggregation within groups, it employs a high-low pairing hybrid strategy. This proactively and consciously mixes high-quality candidates from high-confidence groups with potentially effective candidates from low-confidence groups, avoiding the "bad money drives out good" phenomenon and mitigating the risk of the "within-group mean" being dragged down by a large number of low-quality samples in traditional methods. By pairing high-confidence samples with low-confidence samples, the reliability of high-confidence samples can be fully utilized to guide and correct the predictions of low-confidence samples, thereby improving the overall decision-making quality and maximizing information value. This ensures that the "knowledge" of high-confidence samples is not drowned out by the noise of the low-confidence group, while also providing effective "anchors" for low-confidence samples to help them escape ambiguity.

[0188] Finally, the penalty coefficient of the KL constraint is reduced using a sigmoid function. This provides the model with a dynamic exploration-utilizing balancing mechanism. In the early stages of training, a larger... The β value encourages the model to actively explore various generation paths to discover potential excellent strategies; as training progresses, the β value gradually decays, and the model converges to a more stable and reliable strategy. This dynamic adjustment gives the model strong generalization ability, enabling it to adapt to new scenarios beyond the training data and solving the problem of weak generalization ability in traditional methods.

[0189] This application, through profound insights into the two fundamental problems of "information flow disruption" and "target misalignment," constructs a novel paradigm that is self-evolving, highly robust, and strongly generalizable. Specifically, it builds a POI name generation framework that does not require a reward model and is based on multi-dimensional reward scoring and GRPO relative advantage optimization. This framework jointly scores candidate point-of-interest names and standard answers through multiple dimensions, including consistency rewards, semantic similarity rewards, format standardization rewards, length reasonableness rewards, and information richness rewards. At the same time, it introduces an uncertainty-aware grouping mechanism and combines KL constraints with Sigmoid decay and uncertainty modulation factors to attenuate the gradient of low-confidence samples. It designs an improved GRPO algorithm loss function based on gradient modulation, ultimately achieving high-precision and highly robust POI name generation in low-precision source image scenarios.

[0190] Figure 3 Flowchart of the Point of Interest Name Recognition Method Provided in this Application Figure 1 ,like Figure 3 As shown, the method includes:

[0191] S301. Acquire street view image.

[0192] For example, street view images refer to visual images containing information about entities such as city streets, shops, and buildings, acquired by a data acquisition device, used to provide visual input in a real-world scene. The data acquisition device may be, for example, a vehicle-mounted camera, a drone, or a mobile device; this embodiment of the application does not limit the scope of the data acquisition. For instance, an electronic device can acquire a street view image to be processed by calling a street view data interface, reading locally stored image files, or receiving a real-time video stream.

[0193] S302. Perform target detection processing on the street view image to obtain the target sign image.

[0194] For example, object detection processing refers to the process of analyzing street scene images using computer vision algorithms to identify and locate the areas containing the sign entity. The target sign image is a partial image cropped from the street scene image that contains only the target sign entity, used to eliminate background interference and focus on the visual features of the sign itself. For instance, the street scene image can be input into a pre-trained object detection model. The model identifies the detection box positions of the sign entity through operations such as feature extraction, candidate box generation, classification, and regression. Then, based on the coordinates of the detection boxes, the street scene image is cropped to obtain a partial image containing only the target sign.

[0195] S303. Based on the target sign image and the target large model, determine the names of the points of interest in the target sign image.

[0196] For example, the point of interest name is the name of the entity represented by the target sign, such as a shop name or organization name. For instance, the target sign image can be input into the aforementioned trained target model, which outputs the point of interest name with the highest confidence, and then stored in a database after formal verification.

[0197] The point-of-interest (POI) name recognition method provided in this application first acquires a street view image containing entity information, then locates and crops the target sign image from the street view image through object detection, and finally uses a multimodal large model to understand and generate the corresponding POI name. This method first focuses on the sign entity through object detection, eliminating interference from complex backgrounds in the street view image and improving the targeting and accuracy of subsequent large model processing. Then, leveraging the powerful image understanding capabilities of the multimodal large model, it directly generates POI names from the sign image, avoiding the dependence of traditional OCR technology on text clarity. It can effectively handle blurry, distorted, or artistic sign text, significantly improving the robustness and accuracy of street view POI recognition, while simplifying the recognition process and improving overall processing efficiency.

[0198] In summary, the target large model trained as described above significantly improves the accuracy of POI name generation (especially under low-precision source images); the generated results are semantically reasonable, formatted correctly, of appropriate length, and complete in information, conforming to map business standards; the training process converges faster and the strategy updates are more stable, reducing oscillations and overfitting; the manual review rate is greatly reduced, enabling highly automated POI data production; it has strong robustness to low-quality inputs such as blurry, occluded, and garbled characters; and it eliminates the need to train an independent reward model (RM), reducing deployment complexity and computational overhead.

[0199] Figure 4 A schematic diagram of the structure of the model training device for point of interest name recognition provided in this application is shown below. Figure 4 As shown, the model training device 400 for point of interest name recognition provided in this embodiment includes:

[0200] The acquisition module 401 is used to acquire the image of the hanging tag, including the sample hanging tag;

[0201] The processing module 402 is used to input the hanging image into a preset initial large model to obtain multiple candidate interest point names of the sample hanging;

[0202] The determination module 403 is used to determine the total reward value of each candidate point of interest name according to a preset multi-dimensional reward function; wherein, the multi-dimensional reward function includes at least two dimensions from consistency reward, semantic similarity reward, format standardization reward, information richness reward, and length reasonableness reward;

[0203] The training module 404 is used to train the initial large model based on the total reward value of the candidate point of interest names and the preset GRPO algorithm to obtain the target large model; wherein, the target large model is used to identify the point of interest names in the hanging image.

[0204] In one possible implementation, the determining module 403 is configured to:

[0205] For each candidate point of interest name, the sub-reward value of the candidate point of interest name in each dimension is determined according to the multi-dimensional reward function;

[0206] The total reward value for a candidate point of interest is determined based on the sub-reward values ​​of that candidate point of interest name in each dimension.

[0207] In one possible implementation, the multi-dimensional reward function includes a semantic similarity reward dimension, and the determination module 403 is used for:

[0208] Based on the candidate point of interest name and the preset name truth value of the sample sign, a first parameter and a second parameter are determined; wherein, the first parameter is used to measure the degree of content coverage of the candidate point of interest name with the name truth value; and the second parameter is used to measure the matching accuracy between the candidate point of interest name and the name truth value.

[0209] The similarity between the embedding vector of the sign image and the embedding vector of the candidate point of interest name is calculated to obtain the third parameter;

[0210] Based on the first parameter, the second parameter, and the third parameter, determine the sub-reward value of the candidate point of interest name in the semantic similarity reward dimension.

[0211] In one possible implementation, the candidate point of interest name includes the corresponding inference text information; the multi-dimensional reward function includes a format-standard reward dimension; and the determination module 403 is used for:

[0212] The fourth parameter is determined based on the candidate point of interest name and the preset symbol usage rules; the fourth parameter is used to measure the degree of symbol standardization of the candidate point of interest name.

[0213] The fifth parameter is determined based on the inference text information corresponding to the candidate point of interest name; the fifth parameter is used to measure whether the initial large model performs inference.

[0214] Based on the fourth and fifth parameters, determine the sub-reward value of the candidate point of interest name in the format normativity reward dimension.

[0215] In one possible implementation, the determining module 403 is configured to:

[0216] When the multi-dimensional reward function includes a consistency reward dimension, the sub-reward value of the candidate point of interest name in the consistency reward dimension is determined based on the candidate point of interest name and the preset truth value of the name on the sample listing; or,

[0217] When the multi-dimensional reward function includes an information richness reward dimension, the sub-reward value of the candidate point of interest name in the information richness reward dimension is determined based on the candidate point of interest name and the preset information category; or,

[0218] When the multi-dimensional reward function includes a length reasonableness reward dimension, the sub-reward value of the candidate point of interest name in the length reasonableness reward dimension is determined based on the character length of the candidate point of interest name.

[0219] In one possible implementation, training module 404 is used for:

[0220] Based on the confidence level of the candidate point of interest names, the candidate point of interest names are grouped to obtain multiple sets of candidate names;

[0221] For each set of candidate names, the relative advantage value of each candidate point of interest name is determined based on the total reward value of each candidate point of interest name in that set.

[0222] The loss function value is determined based on the confidence level of the candidate interest point names, the relative advantage value of the candidate interest point names, and the loss function of the preset GRPO algorithm.

[0223] Based on the loss function value, the model parameters of the initial large model are updated in reverse until the updated initial large model meets the preset conditions, thus obtaining the target large model.

[0224] In one possible implementation, the training module 404 is used for:

[0225] The uncertainty of the candidate point of interest names is determined based on the confidence level of the candidate point of interest names.

[0226] The candidate point of interest names whose uncertainty is greater than a preset threshold and the candidate point of interest names whose uncertainty is less than or equal to the preset threshold are mixed and grouped according to a preset ratio to obtain multiple sets of candidate names.

[0227] In one possible implementation, the device further includes an identification module for:

[0228] Acquire street view images;

[0229] The street view image is processed by target detection to obtain the target sign image;

[0230] Based on the target tag image and the target large model, determine the names of the points of interest in the target tag image.

[0231] The model training device for point of interest name recognition provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0232] Figure 5 A schematic diagram of the structure of the electronic device provided in this application. Figure 5 As shown, the electronic device 500 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the electronic device 500 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus.

[0233] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0234] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0235] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0236] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0237] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0238] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0239] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0240] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0241] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0242] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0243] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0244] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0245] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0246] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0247] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.< / think>

Claims

1. A model training method for point-of-interest (POI) name recognition, characterized in that, include: Obtain images of the hanging tags, including the sample tags; The image of the sign is input into a preset initial large model to obtain multiple candidate point of interest names for the sample sign; The total reward value for each candidate point of interest name is determined according to a preset multi-dimensional reward function; wherein, the multi-dimensional reward function includes at least two dimensions from consistency reward, semantic similarity reward, format standardization reward, information richness reward, and length reasonableness reward. Based on the total reward value of the candidate point of interest names and the preset GRPO algorithm, the initial large model is trained to obtain the target large model; wherein, the target large model is used to identify the point of interest names in the hanging image.

2. The method according to claim 1, characterized in that, The step of determining the total reward value for each candidate point of interest name based on a preset multi-dimensional reward function includes: For each candidate point of interest name, the sub-reward value of the candidate point of interest name in each dimension is determined according to the multi-dimensional reward function; The total reward value for a candidate point of interest is determined based on the sub-reward values ​​of that candidate point of interest name in each dimension.

3. The method according to claim 2, characterized in that, The multi-dimensional reward function includes a semantic similarity reward dimension. Determining the sub-reward value of the candidate interest point name in each dimension based on the multi-dimensional reward function includes: Based on the candidate point of interest name and the preset name truth value of the sample sign, a first parameter and a second parameter are determined; wherein, the first parameter is used to measure the content coverage of the candidate point of interest name to the name truth value; and the second parameter is used to measure the matching accuracy between the candidate point of interest name and the name truth value. The embedding vector of the sign image and the embedding vector of the candidate point of interest name are similar to each other to obtain the third parameter; Based on the first parameter, the second parameter, and the third parameter, determine the sub-reward value of the candidate point of interest name in the semantic similarity reward dimension.

4. The method according to claim 2, characterized in that, The candidate point of interest name includes corresponding inference text information; the multi-dimensional reward function includes a format-standard reward dimension, and determining the sub-reward value of the candidate point of interest name in each dimension according to the multi-dimensional reward function includes: Based on the candidate point of interest name and the preset symbol usage rules, a fourth parameter is determined; wherein, the fourth parameter is used to measure the degree of symbol standardization of the candidate point of interest name; The fifth parameter is determined based on the inference text information corresponding to the candidate point of interest name; wherein, the fifth parameter is used to measure whether the initial large model performs inference; Based on the fourth and fifth parameters, determine the sub-reward value of the candidate point of interest name in the format normativity reward dimension.

5. The method according to claim 2, characterized in that, The step of determining the sub-reward value of the candidate point of interest name in each dimension according to the multi-dimensional reward function includes: When the multi-dimensional reward function includes a consistency reward dimension, the sub-reward value of the candidate point of interest name in the consistency reward dimension is determined based on the candidate point of interest name and the preset truth value of the name on the sample tag; or, When the multi-dimensional reward function includes an information richness reward dimension, the sub-reward value of the candidate point of interest name in the information richness reward dimension is determined based on the candidate point of interest name and the preset information category; or, When the multi-dimensional reward function includes a length reasonableness reward dimension, the sub-reward value of the candidate point of interest name in the length reasonableness reward dimension is determined based on the character length of the candidate point of interest name.

6. The method according to claim 1, characterized in that, The candidate point of interest names have confidence levels; the process of training the initial large model based on the total reward value of the candidate point of interest names and a preset GRPO algorithm to obtain the target large model includes: Based on the confidence level of the candidate point of interest names, the candidate point of interest names are grouped to obtain multiple sets of candidate names; For each set of candidate names, the relative advantage value of each candidate point of interest name is determined based on the total reward value of each candidate point of interest name in that set. The loss function value is determined based on the confidence level of the candidate point of interest name, the relative advantage value of the candidate point of interest name, and the loss function of the preset GRPO algorithm. Based on the loss function value, the model parameters of the initial large model are updated in reverse until the updated initial large model meets the preset conditions, thus obtaining the target large model.

7. The method according to claim 6, characterized in that, The candidate point of interest names are grouped according to their confidence levels to obtain multiple sets of candidate names, including: The uncertainty of the candidate point of interest name is determined based on the confidence level of the candidate point of interest name; The candidate point of interest names whose uncertainty is greater than a preset threshold and the candidate point of interest names whose uncertainty is less than or equal to the preset threshold are mixed and grouped according to a preset ratio to obtain multiple sets of candidate names.

8. The method according to any one of claims 1-7, characterized in that, The method further includes: Acquire street view images; The street view image is subjected to target detection processing to obtain the target sign image; Based on the target tag image and the target large model, determine the names of the points of interest in the target tag image.

9. A model training device for point-of-interest (POI) name recognition, characterized in that, include: The acquisition module is used to acquire images of the hanging tags, including sample hanging tags; The processing module is used to input the hanging image into a preset initial large model to obtain multiple candidate interest point names of the sample hanging; The determination module is used to determine the total reward value of each candidate point of interest name according to a preset multi-dimensional reward function; wherein, the multi-dimensional reward function includes at least two dimensions from consistency reward, semantic similarity reward, format standardization reward, information richness reward, and length reasonableness reward; The training module is used to train the initial large model based on the total reward value of the candidate point of interest names and the preset GRPO algorithm to obtain the target large model; wherein, the target large model is used to identify the point of interest names in the hanging image.

10. An electronic device / computer-readable storage medium / computer program product, characterized in that, The electronic device includes: a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8; The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-8; The computer program product includes a computer program that, when executed by a processor, is used to implement the method as described in any one of claims 1-8.

Images