Automatic recognition and verification system for handwritten signatures of case documents

Abstract

The application discloses a kind of handwritten signature automation identification and check system for file documents, including handwritten signature capture module, key personnel list generation module, context enhanced constrained signature recognition module, multi-source confidence management and unknown rejection module and file level batch check engine module.Handwritten signature capture module is used to locate and cut handwritten signature area;Key personnel list generation module is used to extract candidate character list;Context enhanced constrained signature recognition module is used to construct large graph, small graph and prompt text and input visual language model to reason;Multi-source confidence management and unknown rejection module are used to calculate confidence and determine final output;File level batch check engine module is used to process entire file and generate check report.The application can realize the automated attribution identification and batch compliance check of handwritten signature identity in file documents.

Description

Technical Field

[0001] This invention belongs to the field of intelligent document processing and computer vision technology, specifically relating to an automated recognition and verification system for handwritten signatures in case files. Background Technology

[0002] In fields such as judicial litigation and financial lending, accurate identification of handwritten signatures in case files is a fundamental task. Currently, this is mainly done manually, page by page. Reviewing a case file of several hundred pages typically takes one to three working days, which is inefficient and prone to errors due to fatigue. Therefore, the industry's demand for automated signature recognition and verification systems is increasingly urgent. As an upstream process, the accuracy of handwritten signature recognition directly affects the reliability of downstream automated analysis tasks such as case file review and compliance verification.

[0003] In existing technologies, one direct approach is to use optical character recognition (OCR) to convert signature images into text. However, OCR essentially performs a mapping from pixels to standard characters. Signatures in case files are often cursive or cursive, with stroke shapes that deviate significantly from regular character shapes, frequently resulting in garbled text. Furthermore, it processes text independently within a single line, neglecting the semantic context of fields like agent signatures on the same page and the information about the parties involved recorded throughout the case file, lacking prior knowledge to effectively constrain the recognition results. Another approach is to use a large-scale visual language model. This type of model takes the entire page image as input and jointly understands the layout and text semantics end-to-end, performing well in print document understanding tasks. However, when faced with cursive signatures, its open-ended text generation paradigm still tends to generate a semantically plausible but actually incorrect name with high confidence, resulting in the illusion problem. Even with constraints through injecting a list of candidates using prompts, the model's underlying visual encoder is not specifically optimized for cursive signature shapes, and the recognition accuracy still falls far short of business requirements. Furthermore, the signature target detection scheme only solves the spatial location problem of where the signature is, and does not involve the identity judgment of who the signature belongs to; the signature authenticity comparison scheme requires pre-storing signature samples for each candidate, but a large number of parties in the case file do not have historical signature archives, making it difficult to implement in practice.

[0004] Therefore, there is an urgent need to provide an automated recognition and verification system for handwritten signatures in case files to address the shortcomings of the existing technologies. Summary of the Invention

[0005] To address the challenges of optical character recognition (OCR) in handling cursive signatures and the susceptibility of visual language models to illusions in related technologies, this invention provides an automated handwritten signature recognition and verification system for case files. The system first automatically extracts a list of key individuals from the entire case file to construct a closed-set candidate constraint. Then, it fuses layout semantics and handwriting details through dual-view mapping (large and small images). Leveraging a constrained visual language model, signature recognition is transformed from an open-ended generation problem into a selection problem within a finite candidate set. Furthermore, it combines multi-source confidence fusion and a dual-threshold mechanism to force unknown options to be output for low-confidence samples. Finally, a case-level batch verification engine generates a structured compliance report, significantly improving the accuracy and reliability of signature attribution recognition and enabling end-to-end automated processing and compliance verification of thousands of pages of case files.

[0006] The technical solution of this invention is implemented as follows: an automated recognition and verification system for handwritten signatures in case files, comprising: The handwritten signature capture module is used to detect and identify the handwritten area on each page of the target case file, locate and crop out the handwritten signature area; The key personnel list generation module is used to perform optical character recognition and text analysis on the full text of the target case file to generate a list of candidates bound to the target case file. The context-enhanced constrained signature recognition module is used to construct a large image containing layout context information, a small image containing handwriting details, and a prompt text containing the candidate list and unknown options for each handwritten signature area. The large image, small image, and prompt text are input into the visual language model to output the recognition result. The multi-source confidence management and unknown rejection module is used to calculate the multi-source fusion confidence of the identification result and determine the final output based on the relationship between the multi-source fusion confidence and a preset threshold. The case file-level batch verification engine module is used to call the above modules to perform batch processing on the basis of the entire case file, summarize the recognition results and generate a structured verification report.

[0007] This invention constructs an end-to-end closed-loop processing chain by setting up a handwritten signature capture module, a key personnel list generation module, a context-enhanced constrained signature recognition module, a multi-source confidence management and unknown rejection module, and a case file-level batch verification engine module. This chain consists of signature area localization, candidate list generation, multi-modal constrained recognition integrating layout context and handwriting details, confidence quantification and hierarchical output, and whole-file-level parallel scheduling and structured verification report generation. It can transform signature recognition from a character transcribing or matching problem in an open space into a finite candidate set selection problem constrained by the case file's own information. It significantly improves the accuracy and reliability of cursive signature attribution recognition without the need for external sample library support. Through a clear rejection mechanism, it avoids misleading output from low-confidence samples. Ultimately, it achieves minute-level automated batch processing and auditable, traceable compliance verification of thousands of pages of case files, providing an efficient and reliable intelligent solution for case file document recognition in serious scenarios such as law and finance.

[0008] As a further improvement to the above solution, the handwritten signature capture module is configured to perform the following steps: Perform handwritten character region detection on the input page and extract candidate regions; The candidate regions are classified into general handwritten content and signature-type handwritten content. The classification is based on a comprehensive judgment that considers the semantic information of the surrounding text and the positional relationship of the candidate regions. Perform semantic constraint-based boundary fine-tuning on the handwritten region identified as a signature, correct the boundary of the detection box, and output the cropped handwritten signature region.

[0009] By combining the semantic information of the surrounding text and the positional relationship of the page to determine the type, and by fine-tuning the boundary of the detection box based on semantic constraints, it is possible to effectively distinguish between signature-type handwritten content and general handwritten content, achieve accurate positioning and precise cropping of the signature area, and reduce false detections and missed detections.

[0010] As a further improvement to the above solution, the key personnel list generation module is configured to perform the following steps: Optical character recognition is performed on the full text of the target case file, and the recognition results are filtered by confidence level to retain the text content that meets the preset confidence level conditions. Named entity recognition is performed on text content that meets the pre-set reliability conditions to extract personal name entities; By combining the contextual semantic information near the person's name entity, the role of the person's name entity is inferred; Based on the frequency of occurrence of the named entities in the target case file, the range of page numbers in which they appear, the document type in which they appear, and their contextual association with the signature field, the candidate entity list is filtered and generated.

[0011] By performing confidence screening, named entity recognition, contextual semantic role inference, and comprehensive multi-dimensional feature screening on the optical character recognition results, it is possible to automatically and accurately extract candidate information and their role information related to the signature recognition task from unstructured case file texts, providing a reliable and complete closed-set constraint space for subsequent signature attribution recognition.

[0012] As a further improvement to the above scheme, the context-enhanced constrained signature recognition module is configured to perform the following steps: Construct a large image: Obtain a complete image of the page where the handwritten signature area is located, and indicate the position of the handwritten signature area to be identified on the complete image; Constructing a small image: A partial image of the signature is obtained by cropping from the page image based on the detection box of the handwritten signature area; Constructing prompt text: Generate prompt text, which includes the name and role information of each candidate in the candidate list, the unknown option, and the task instruction. The task instruction is used to instruct the visual language model to select one of the candidates and the unknown option as the final attribution conclusion. Perform multimodal reasoning: The large image, small image, and prompt text are passed together as multimodal inputs to the visual language model for reasoning. The visual language model first outputs the reasoning process and then outputs the final attribution conclusion, which is a specific person in the candidate list or the unknown option.

[0013] By simultaneously constructing a large image containing the layout context, a small image for observing handwriting details, and a prompt text containing a list of candidate entities and unknown options, and inputting them together into a visual language model to perform multimodal reasoning, the signature recognition process integrates layout semantic perception, handwriting detail observation, and candidate constraint driving. This effectively suppresses the illusion problem caused by open generation and improves the accuracy, interpretability, and reliability of signature attribution recognition.

[0014] As a further improvement to the above scheme, the training process of the visual language model includes: In the inference chain distillation supervised fine-tuning stage, the teacher model samples the training samples multiple times, selects the inference chains that meet the quality conditions, and uses the selected inference chains to supervise and fine-tune the student model, distilling to obtain the structured character reasoning ability. In the iterative direct preference optimization stage, preference pairs are constructed based on the preference order consisting of correct identification results, reasonable rejection results, and incorrect identification results. Multiple rounds of direct preference optimization are then performed on the supervised fine-tuned student model to calibrate the trade-off strategy between identification and rejection and correct systemic behavioral defects.

[0015] Through two-stage training—supervised fine-tuning via inference chain distillation and iterative direct preference optimization—the visual language model acquires structured glyph reasoning capabilities and establishes a clear preference order among correct recognition, reasonable rejection, and incorrect recognition. This effectively calibrates the model's trade-off strategy between recognition and rejection, significantly reducing the risk of misjudgment in cursive signature scenarios.

[0016] As a further improvement to the above scheme, the multi-source fusion confidence score is calculated using the following formula: final_confidence = w1 × C_self_unknown + w2 × C_reason_consistency+ w3 × C_sampling_consistency + w4 × C_semantic_prior; Wherein, C_self_unknown represents the confidence level of the model in determining that the current signature should output an unknown option, C_reason_consistency represents the confidence level of the consistency between the reasoning process and the final output, C_sampling_consistency represents the confidence level of the consistency between multiple sampling results, and C_semantic_prior represents the confidence level of the semantic matching between the case file and the prior material; w1, w2, w3, and w4 are the corresponding weights; The preset thresholds include a high threshold and a low threshold, and the multi-source confidence management and unknown rejection module determines the final output according to the following rules: When the confidence level of the multi-source fusion is greater than or equal to the high threshold, a specific person in the candidate list is directly output. When the multi-source fusion confidence level is greater than or equal to the low threshold and less than the high threshold, output a specific person in the candidate list and mark it as requiring manual review. When the confidence level of the multi-source fusion is less than the low threshold, the unknown option is forcibly output.

[0017] By calculating the comprehensive confidence level through weighted fusion of multi-source signals and combining it with dual thresholds to divide the output into three levels: direct output, manual verification, and forced unknown, a complete reliability assurance mechanism from confidence quantification to hierarchical output is constructed. This ensures efficient output of high-confidence results and effectively filters out unreliable identification results through verification and rejection, thereby improving security and practicality.

[0018] As a further improvement to the above solution, the case file-level batch verification engine module is configured to perform the following steps: The case file is split into multiple page-level tasks, and signature detection and recognition are performed in parallel by multiple computing nodes. The candidate list is shared across the entire volume, and intermediate text recognition results, page images, and signature detection boxes are cached. The recognition results of each signature are stored in the database in a structured form. The recognition results include at least the volume identifier, page identifier, bounding box coordinates, the person to whom the signature belongs or the unknown option, role, fusion confidence, and model reasoning. Load configurable business rule templates and perform compliance determination on the identification results; Generate a verification report covering the entire volume, which includes a list of missing signature pages, a list of suspicious signature pages, a list of manually reviewed signatures, and a structured list of all signatures.

[0019] By using page-level parallelism and volume-level caching, we achieve efficient processing of case files with thousands of pages. After the recognition results are structured and stored in the database, they are combined with configurable business rule templates for automatic compliance judgment. Finally, a detailed verification report covering the entire file is generated, which effectively improves the automation, processing efficiency and traceability of case file signature recognition.

[0020] As a further improvement to the above scheme, the iterative direct preference optimization stage includes: In the first sub-stage, multiple iterations are conducted with accuracy as the guide. In each round, the training samples are sampled using the current checkpoint as the strategy model. Preference pairs are constructed for optimization based on the preference order that correct identification results are better than reasonable rejection results, which are better than incorrect identification results. The second sub-stage is guided by behavior modification, and constructs preference pairs to optimize three types of failure modes: inconsistency between reasoning and output, out-of-bounds identification, and repetitive loops.

[0021] By establishing a priority gradient for recognition behavior through multi-round iterative preference optimization guided by accuracy, and then conducting targeted optimization for three types of systematic failure modes guided by behavior correction, the two-stage collaboration achieves a systematic improvement in the model's recognition accuracy and output standardization.

[0022] As a further improvement to the above solution, the verification report can be output in JSON, Excel, PDF or web page format.

[0023] By supporting multiple report output formats, the system can adapt to the needs of viewing, archiving, and secondary processing in different business scenarios, thereby improving its ease of use and versatility.

[0024] As a further improvement to the above solution, the case file-level batch verification engine module retryes page-level tasks that fail and records exception information when executing page-level tasks.

[0025] By retrying page-level tasks that fail and logging exception information, the fault tolerance and success rate of the batch processing flow are improved, while providing effective log evidence for system operation and maintenance and problem tracing.

[0026] Compared with traditional processing methods, the present invention has the following advantages: (1) This invention constructs a candidate list by automatically extracting printed names and role information from case files, transforming signature recognition from open text generation to discrimination selection within a limited candidate set, fundamentally reducing the decision space of the model. At the same time, it uses a large image containing the page context and a small image that retains handwriting details to form a dual-view input, enabling the visual language model to make a comprehensive judgment by jointly utilizing the semantic information of the page and the glyph features. Compared with general OCR or visual language model solutions, the accuracy of signature attribution recognition is greatly improved. As the upstream link of the case file processing process, the improvement of recognition accuracy can effectively support the accurate implementation of downstream case file review, compliance verification and other automated analysis tasks, and avoid the failure of downstream analysis due to the transmission of upstream recognition errors.

[0027] (2) This invention splits the entire case file into multiple independent page-level tasks and executes signature detection and recognition in parallel by multiple computing nodes. At the same time, the candidate list and OCR intermediate results are shared and cached throughout the entire file to avoid duplicate calculations. This allows a thousand-page case file to complete full signature recognition and compliance verification within minutes. Compared with the traditional processing mode of manual page-by-page review in the prior art, which takes 1 to 3 working days, this invention achieves an efficiency improvement of tens or even hundreds of times.

[0028] (3) The present invention stores the recognition results of each signature in a structured form. Each recognition record includes volume identifier, page identifier, bounding box coordinates, belonging person or unknown option, role, fusion confidence and model reasoning reason, providing a complete chain of evidence for each recognition conclusion, which is convenient for post-event traceability and compliance review. Compared with the current situation where manual review is difficult to form structured records, the structured output of the present invention can itself serve as the basic data asset for compliance management.

[0029] (4) This invention obtains a calibrated comprehensive confidence level by weighted fusion of multiple source signals such as token-level probability of the output sequence of the visual language model, consistency ratio of multiple sampling, consistency verification result of reasoning process and final output, role prior matching degree and cross-page handwriting similarity. Based on dual thresholds, the output is divided into three levels: direct output, manual review and forced rejection. This effectively solves the "illusion" problem of existing models outputting erroneous results with high confidence on low confidence samples. It ensures that the system forces the output of unknown options instead of making misleading guesses when it cannot reliably identify the data, and fully adapts to the strict requirements of output reliability in legal, financial and other scenarios. Attached Figure Description

[0030] Figure 1 This is a structural block diagram of an automated handwritten signature recognition and verification system for case files provided in an embodiment of the present invention; Figure 2 An example diagram illustrating the inference flow of the context-enhanced constrained signature recognition module provided in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example like Figure 1 As shown, this embodiment provides an automated recognition and verification system for case file document signatures, including a handwritten signature capture module, a key personnel list generation module, a context-enhanced constrained signature recognition module, a multi-source confidence management and unknown rejection module, and a case file-level batch verification engine module.

[0033] In this embodiment, the target case file document is input into the handwritten signature capture module and the key personnel list generation module, and the two modules start working in parallel.

[0034] In the first branch of parallel processing, the handwritten signature capture module detects handwritten regions on the input page and extracts candidate regions. It then performs type discrimination on the candidate regions to distinguish between general handwritten content and signature-type handwritten content. The type discrimination is combined with the semantic information of the surrounding text and the positional relationship of the page to make a comprehensive judgment. For the handwritten regions that are determined to be signatures, it performs boundary fine-tuning processing based on semantic constraints, corrects the boundaries of the detection box, and outputs the cropped handwritten signature region and its associated information, including page number, coordinate position, adjacent text context, and signature field type.

[0035] On another parallel path, the key personnel list generation module performs optical character recognition (OCR) on the full text of the target case file and filters the recognition results based on confidence levels, retaining text content that meets preset confidence conditions. It then performs named entity recognition on the text content that meets the preset confidence conditions to extract personal name entities. Combining the contextual semantic information surrounding the personal name entities, it infers the roles of these entities. Based on the frequency of occurrence of the personal name entities in the target case file, the range of page numbers they appear on, the document type they belong to, and their contextual association with the signature field, it filters and generates a candidate list. This candidate list includes at least the names of the personnel and their role information, serving as a candidate set for subsequent signature recognition.

[0036] After the parallel processing described above is completed, the handwritten signature region output by the handwritten signature capture module and the candidate list output by the key personnel list generation module are used as input to the context-enhanced constrained signature recognition module. The context-enhanced constrained signature recognition module performs the following processing steps for each handwritten signature region: First, constructing a large image: acquiring a complete image of the page containing the handwritten signature region and marking the position of the target handwritten signature region to be recognized with a highlighted border on the complete image, enabling the visual language model to perceive the signature's position on the page, such as whether it falls in the "representative lawyer" or "client" column. Second, constructing a small image: cropping a partial image of the signature from the page image based on the detection bounding box of the handwritten signature region, allowing the model to observe handwriting details. Third, constructing prompt text: generating prompt text containing the names and roles of each candidate in the candidate list, the unknown option "Unknown," and task instructions. These task instructions instruct the visual language model to select one from the closed set formed by the candidate list and the unknown option "Unknown" as the final attribution conclusion, and output the reasoning process and the final conclusion. Subsequently, the module takes the large image, small image, and prompt text as multimodal inputs and passes them to the visual language model to perform inference. The visual language model first outputs the inference process and then outputs the final attribution conclusion, which is a specific person in the candidate list or an unknown option.

[0037] In this embodiment, the visual language model used is a domain-adaptive model that has undergone enhanced fine-tuning. Its training process includes a distillation-based supervised fine-tuning stage for inference chains and an iterative direct preference optimization stage. In the distillation-based supervised fine-tuning stage, the teacher model Kimi-K2.5 performs 8 rollout samplings on each sample in the training set at a temperature τ=0.7. If all rollouts fail, the number of rollouts is increased to 64. The sampling results are filtered based on both answer correctness and inference chain quality criteria, eliminating inference chains with "assumption-first" assumptions and excessive self-affirmation, retaining inference chains with self-correcting behavior and upsampling them to approximately 10%. The student model Qwen3-VL-8B-Thinking is then fine-tuned using LoRA on the 1500 high-quality inference chains selected. In this embodiment, rank=32, training for 1 epoch, and a learning rate of 1×10⁻⁶. -5 The warmup ratio is 0.1. Thus, the student model acquires the structured character reasoning ability of the teacher model, including radical decomposition, stroke comparison, and layout cross-referencing. In the iterative direct preference optimization stage, based on the supervised fine-tuned student model, preference pairs are constructed and multiple rounds of direct preference optimization are performed based on the preference order consisting of correct recognition results, reasonable rejection results, and incorrect recognition results. This stage further includes a first sub-stage and a second sub-stage. The first sub-stage is accuracy-oriented and consists of 6 iterations: each round uses the current checkpoint as the policy model, samples N=8 rollouts for each training sample at a temperature τ=0.95, constructs three types of preference pairs according to the preference order "correct recognition results are better than reasonable rejection results are better than incorrect recognition results," and samples them in a 1:1:1 ratio. Approximately 66k preference pairs are constructed in 6 rounds. Each round uses DPO loss for optimization, with a learning rate of 2×10⁻⁶. -5 The batch size is 32. The second sub-phase is behavior correction-oriented and consists of one round: Approximately 2000 targeted preference pairs are constructed for the three types of systematic failure modes left over from the first sub-phase—approximately 600 pairs are constructed for the inconsistency between reasoning and output mode, favoring samples where "the final output is consistent with the reasoning conclusion"; approximately 600 pairs are constructed for the out-of-bounds recognition mode; and approximately 800 pairs are constructed for the repetitive loop mode, favoring non-repetitive outputs that terminate normally within the token budget. A single round of DPO is performed on these approximately 2000 preference pairs, with a learning rate of 1×10⁻⁶. -5With a batch size of 4, the final model was obtained. In the reinforcement and fine-tuning phase, the reward function integrates four types of signals: a reward for correct attribution (highest weight), guiding the model to output the correct attribution conclusion; a reward for reasonable rejection of unknowns, where the model outputs the unknown option (Unknown) to receive a positive reward for samples where the unknown option should be output, thus encouraging the model to reject the unknown when evidence is insufficient; a confidence calibration reward, guiding the model to give accurate outputs at high confidence levels; and a reasoning interpretability reward, encouraging the model to output interpretable reasoning processes. Through this reward design, the model gradually learns in post-training when to output unknown options, how to use layout information to constrain identities, and how to match cursive signatures with the character features of candidate names.

[0038] The recognition results output by the context-enhanced constrained signature recognition module then enter the multi-source confidence management and unknown rejection module. This module first calculates the multi-source fusion confidence of the recognition results using the following formula: final_confidence = w1 × C_self_unknown + w2 × C_reason_consistency+ w3 × C_sampling_consistency + w4 × C_semantic_prior; Wherein, C_self_unknown represents the confidence level of the model's judgment that the current signature should output an unknown option; C_reason_consistency represents the confidence level of consistency between the reasoning process and the final output; C_sampling_consistency represents the confidence level of consistency between multiple sampling results; and C_semantic_prior represents the confidence level of the semantic matching between the case file and the prior information of the materials. w1, w2, w3, and w4 are the corresponding weights, which can be adjusted according to different document types, business scenarios, test set performance, and misidentification risk requirements. Through the above weighted fusion calculation, the final fusion confidence level is obtained, with a value ranging from 0 to 1.

[0039] After the fused confidence is obtained, the module determines the final output based on a preset threshold. The preset thresholds include a high threshold T_high and a low threshold T_low, where T_high > T_low. In this embodiment, T_high is set to 0.85, and T_low is set to 0.55, which can be adjusted according to the service scenario in the actual application process. When final_confidence ≥ T_high, the system directly outputs a specific person in the candidate person list; when T_low ≤ final_confidence < T_high, the system outputs a specific person in the candidate person list with a mark requiring manual review; when final_confidence < T_low, the system forcibly outputs the unknown option Unknown. In the processing logic of this module, the unknown option Unknown is treated as a first-class output value of the system, and is treated equally with the specific persons in the candidate person list in all subsequent processing flows, without being converted into any guess result.

[0040] The final output results of each page are collected to the dossier-level batch verification engine module. The dossier-level batch verification engine module performs batch processing by taking the entire dossier as the processing unit. First, the dossier is split by page and then enqueued, and multiple computing nodes execute signature detection and recognition in parallel, retry the page-level tasks that fail to execute, and record exception information. Within the scope of the entire dossier, the candidate person list is shared by all computing nodes to avoid repeated extraction page by page; at the same time, intermediate results such as text recognition intermediate results, page images and signature detection frames are cached by dossier for reuse in subsequent stages. After each computing node completes recognition, the system stores the recognition result of each signature in a structured form into the database, and each record at least includes the dossier identifier, page identifier, bounding box coordinates, attributed person or unknown option, role, fused confidence, model inference reason, sampling seed, model version and processing status. After all pages of the entire dossier are processed, the system loads a configurable service rule template to perform compliance judgment on the recognition results. Finally, the system generates a verification report covering the entire dossier, which includes a list of pages with missing signatures, a list of pages with suspicious signatures, a manual review list and full structured details of all signatures, and can be output in the form of JSON, Excel, PDF or web page for service system calling and manual review.

[0041] As Figure 2As shown, the upstream processing of the inference process includes a cropped handwritten signature image output by the handwritten signature capture module. Input data includes the handwritten signature image and a structured candidate list output by the key personnel list generation module. Each candidate in the candidate list includes their name and role information; for example, "Yan Wenliang" is the plaintiff's litigation agent, "Yan Wenhao" is a judge, and "Zhang Yunlin" is a police officer. During the inference process, the visual language model compares the signature to be identified with each candidate in the candidate list one by one based on the page context information and the handwriting features of the cropped signature image, ultimately outputting "Yan Wenliang" as the recognition result.

[0042] In this embodiment, an exemplary form of the prompt text constructed by the context-enhanced constrained signature recognition module is described below. Assuming the candidate list output by the key personnel list generation module is: Zhang San (suspect, suspected of theft, male, 32 years old), Li Si (victim, male, 45 years old), Wang Wu (defense lawyer, Guangdong XX Law Firm), Zhao Liu (court clerk, XX District People's Court), then the constructed complete prompt text is: [Task Description] You are a handwritten signature recognition expert. Based on the following input information, identify the signatory of the target signature area (marked in red).

[0043] [List of Candidates] 1. Zhang San – Suspect, “Zhang” (bow and long Zhang, left-right structure), “San” (three horizontal strokes); 2. Li Si – Victim, “Li” (wood and child Li, top-bottom structure), “Si” (four frames containing “er”); 3. Wang Wu – Defense Lawyer, “Wang” (three horizontal strokes and one vertical stroke), “Wu” (horizontal stroke followed by vertical stroke); 4. Zhao Liu – Clerk, “Zhao” (walking radical + “X”), “Liu” (dot and horizontal stroke + “eight”).

[0044] [Optional] If the target signature cannot match any of the people in the above list, or if the visual evidence is insufficient to make a reliable judgment, please output an unknown option.

[0045]

Output Requirements

[0046] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. An automated recognition and verification system for handwritten signatures in case files, characterized in that, include: The handwritten signature capture module is used to detect and identify the handwritten area on each page of the target case file, locate and crop out the handwritten signature area; The key personnel list generation module is used to perform optical character recognition and text analysis on the full text of the target case file to generate a list of candidates bound to the target case file. The context-enhanced constrained signature recognition module is used to construct a large image containing layout context information, a small image containing handwriting details, and a prompt text containing the candidate list and unknown options for each handwritten signature area. The large image, small image, and prompt text are input into the visual language model to output the recognition result. The multi-source confidence management and unknown rejection module is used to calculate the multi-source fusion confidence of the identification result and determine the final output based on the relationship between the multi-source fusion confidence and a preset threshold. The case file-level batch verification engine module is used to call the above modules to perform batch processing on the basis of the entire case file, summarize the recognition results and generate a structured verification report.

2. The automated recognition and verification system for handwritten signatures in case files as described in claim 1, characterized in that, The handwritten signature capture module is configured to perform the following steps: Perform handwritten character region detection on the input page and extract candidate regions; The candidate regions are classified to distinguish between general handwritten content and signature-type handwritten content; The type discrimination is based on a comprehensive judgment combining the semantic information of the text surrounding the candidate region and the positional relationship on the page. Perform semantic constraint-based boundary fine-tuning on the handwritten region identified as a signature, correct the boundary of the detection box, and output the cropped handwritten signature region.

3. The automated recognition and verification system for handwritten signatures in case files as described in claim 1, characterized in that, The key personnel list generation module is configured to perform the following steps: Optical character recognition is performed on the full text of the target case file, and the recognition results are filtered by confidence level to retain the text content that meets the preset confidence level conditions. Named entity recognition is performed on text content that meets the pre-set reliability conditions to extract personal name entities; By combining the contextual semantic information near the person's name entity, the role of the person's name entity is inferred; Based on the frequency of occurrence of the named entities in the target case file, the range of page numbers in which they appear, the document type in which they appear, and their contextual association with the signature field, the candidate entity list is filtered and generated.

4. The automated recognition and verification system for handwritten signatures in case files as described in claim 1, characterized in that, The context-enhanced constrained signature recognition module is configured to perform the following steps: Construct a large image: Obtain a complete image of the page where the handwritten signature area is located, and indicate the position of the handwritten signature area to be identified on the complete image; Constructing a small image: A partial image of the signature is obtained by cropping from the page image based on the detection box of the handwritten signature area; Constructing prompt text: Generate prompt text, which includes the name and role information of each candidate in the candidate list, the unknown option, and the task instruction. The task instruction is used to instruct the visual language model to select one of the candidates and the unknown option as the final attribution conclusion. Perform multimodal reasoning: The large image, small image, and prompt text are passed together as multimodal inputs to the visual language model for reasoning. The visual language model first outputs the reasoning process and then outputs the final attribution conclusion, which is a specific person in the candidate list or the unknown option.

5. The automated recognition and verification system for handwritten signatures in case files as described in claim 1, characterized in that, The training process of the visual language model includes: In the inference chain distillation supervised fine-tuning stage, the teacher model samples the training samples multiple times, selects the inference chains that meet the quality conditions, and uses the selected inference chains to supervise and fine-tune the student model, distilling to obtain the structured character reasoning ability. In the iterative direct preference optimization stage, preference pairs are constructed based on the preference order consisting of correct identification results, reasonable rejection results, and incorrect identification results. Multiple rounds of direct preference optimization are then performed on the supervised fine-tuned student model to calibrate the trade-off strategy between identification and rejection and correct systemic behavioral defects.

6. The automated recognition and verification system for handwritten signatures in case files according to claim 1, characterized in that, The multi-source fusion confidence score is calculated using the following formula: final_confidence = w1 × C_self_unknown + w2 × C_reason_consistency + w3 × C_sampling_consistency + w4 × C_semantic_prior; Wherein, C_self_unknown represents the confidence level of the model in determining that the current signature should output an unknown option, C_reason_consistency represents the confidence level of the consistency between the reasoning process and the final output, C_sampling_consistency represents the confidence level of the consistency between multiple sampling results, and C_semantic_prior represents the confidence level of the semantic matching between the case file and the prior material; w1, w2, w3, and w4 are the corresponding weights; The preset thresholds include a high threshold and a low threshold, and the multi-source confidence management and unknown rejection module determines the final output according to the following rules: When the confidence level of the multi-source fusion is greater than or equal to the high threshold, a specific person in the candidate list is directly output. When the multi-source fusion confidence level is greater than or equal to the low threshold and less than the high threshold, output a specific person in the candidate list and mark it as requiring manual review. When the confidence level of the multi-source fusion is less than the low threshold, the unknown option is forcibly output.

7. The automated recognition and verification system for handwritten signatures in case files according to claim 1, characterized in that, The case file-level batch verification engine module is configured to perform the following steps: The case file is split into multiple page-level tasks, and signature detection and recognition are performed in parallel by multiple computing nodes. The candidate list is shared across the entire volume, and intermediate text recognition results, page images, and signature detection boxes are cached. The recognition results of each signature are stored in the database in a structured form. The recognition results include at least the volume identifier, page identifier, bounding box coordinates, the person to whom the signature belongs or the unknown option, role, fusion confidence, and model reasoning. Load configurable business rule templates and perform compliance determination on the identification results; Generate a verification report covering the entire volume. The verification report includes a list of pages with missing signatures, a list of pages with suspicious signatures, a list of pages that have been manually reviewed, and a structured list of all signatures.

8. The automated recognition and verification system for handwritten signatures in case files as described in claim 5, characterized in that, The iterative direct preference optimization phase includes: In the first sub-stage, multiple iterations are conducted with accuracy as the guide. In each round, the training samples are sampled using the current checkpoint as the strategy model. Preference pairs are constructed for optimization based on the preference order that correct identification results are better than reasonable rejection results, which are better than incorrect identification results. The second sub-stage is guided by behavior modification, and constructs preference pairs to optimize three types of failure modes: inconsistency between reasoning and output, out-of-bounds identification, and repetitive loops.

9. The automated recognition and verification system for handwritten signatures in case files as described in claim 7, characterized in that, The verification report can be output in JSON, Excel, PDF or web page format.

10. The automated recognition and verification system for handwritten signatures in case files according to claim 7, characterized in that, When executing page-level tasks, the case file-level batch verification engine module retryes page-level tasks that fail and records exception information.

Images