Logistics document image recognition method and device, equipment and storage medium

Abstract

This invention relates to the field of intelligent logistics technology, and in particular to a method, apparatus, device, and storage medium for logistics document image recognition. The method first acquires an image of the logistics document to be recognized, a set of labeled image samples, a pre-constructed initial image recognition model, and preset semantic verification rules. Based on the labeled image sample set, the initial image recognition model undergoes multiple rounds of iterative training. Based on the obtained target image recognition model, noise consistency analysis is performed on the logistics document image to be recognized to obtain noise detection results. Semantic feature extraction processing is performed on the logistics document image to be recognized, and semantic verification processing is performed on the obtained semantic feature set based on the semantic verification rules to obtain semantic verification results. Based on the noise detection results and semantic verification results, image tampering analysis processing is performed on the logistics document image to be recognized to obtain image tampering detection results. The aim is to improve the accuracy and efficiency of logistics document image tampering detection and meet the needs of modern large-scale risk control.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics technology, and in particular to a method, apparatus, equipment and storage medium for image recognition of logistics documents. Background Technology

[0002] Logistics documents serve as crucial evidence for logistics business settlement, responsibility allocation, and financial reconciliation; their authenticity directly impacts operational security and fund management. With the accelerated digitalization of the industry, electronic waybills, invoices, and receipts are widely used. While this significantly improves document flow efficiency, it also exacerbates the risk of data tampering. Currently, image editing and modification tools are easy to use and readily available, allowing criminals to easily alter key information and splice parts of document images. By falsifying amounts, dates, and receipt information, they can commit illegal activities, causing direct economic losses to businesses.

[0003] Traditional document verification relies primarily on manual, one-by-one review, which is not only time-consuming and labor-intensive, but also inefficient. When faced with massive volumes of documents, reviewers are prone to visual fatigue, leading to missed detections and misjudgments, making it difficult to guarantee accuracy and failing to meet the demands of large-scale, high-frequency risk control verification. Currently, most existing verification methods in the industry rely solely on basic OCR technology for text extraction and simple content comparison, only capable of verifying surface-level text information and lacking effective detection capabilities for tampering with the image itself. Existing solutions cannot accurately identify more subtle tampering behaviors such as pixel modification and region splicing, resulting in significant vulnerabilities in risk control and failing to meet the stringent requirements of modern logistics operations for document authenticity verification. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, apparatus, device and storage medium for logistics document image recognition, which aims to improve the accuracy and efficiency of logistics document image tampering detection.

[0005] The first aspect of this invention provides a method for recognizing logistics documents, comprising: acquiring an image of a logistics document to be recognized, a set of labeled image samples, a pre-constructed initial image recognition model, and preset semantic verification rules; performing multiple rounds of iterative training on the initial image recognition model based on the labeled image sample set to obtain a target image recognition model; performing noise consistency analysis on the image of the logistics document to be recognized based on the target image recognition model to obtain a noise detection result; performing semantic feature extraction processing on the image of the logistics document to be recognized to obtain a semantic feature set; performing semantic verification processing on the semantic feature set based on the semantic verification rules to obtain a semantic verification result; and performing image tampering analysis processing on the image of the logistics document to be recognized based on the noise detection result and the semantic verification result to obtain an image tampering detection result.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the labeled image sample set includes a normal document subset, a labeled modified sample subset, and an unlabeled modified sample subset; the step of performing multiple rounds of iterative training on the initial image recognition model based on the labeled image sample set to obtain a target image recognition model includes: inputting the normal document subset into the initial image recognition model for basic training using normal noise features to obtain a normal noise baseline model; inputting the labeled modified sample subset into the normal noise baseline model for fine-tuning the detection head to obtain a fine-tuned image recognition model; inputting the unlabeled modified sample subset into the fine-tuned image recognition model for semi-supervised optimization training to obtain a semi-supervised optimized model; and performing multiple rounds of iterative training on the semi-supervised optimized model based on the labeled image sample set to obtain the target image recognition model.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the step of performing multiple rounds of iterative training on the semi-supervised optimization model based on the labeled image sample set to obtain the target image recognition model includes: obtaining actual noise detection results corresponding to the labeled image sample set; inputting the labeled image sample set into the semi-supervised optimization model for noise consistency analysis to obtain predicted noise detection results; calculating model loss based on the actual noise detection results and the predicted noise detection results; iteratively updating the weight parameters of the semi-supervised optimization model using a backpropagation algorithm based on the model loss; and outputting the target image recognition model when a preset iteration stopping condition is met.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the target image recognition model includes an image preprocessing module, a scale division module, a noise extraction module, and a noise analysis module connected in sequence; the step of performing noise consistency analysis on the logistics document image to be identified based on the target image recognition model to obtain a noise detection result includes: performing image preprocessing on the logistics document image to be identified based on the image preprocessing module to obtain a preprocessed image; performing multi-scale region division processing on the preprocessed image based on the scale division module to obtain multiple image region blocks; performing noise extraction processing on each of the image region blocks based on the noise extraction module to obtain a noise feature map corresponding to each of the image region blocks; and performing noise analysis on each of the noise feature maps based on the noise analysis module to obtain the noise detection result.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the noise analysis module includes a noise similarity calculation submodule, a fusion submodule, and an analysis submodule connected in sequence; the step of performing noise analysis on each of the noise feature maps based on the noise analysis module to obtain the noise detection result includes: calculating the noise similarity of each of the noise feature maps based on the noise similarity calculation submodule; performing a weighted fusion calculation on multiple noise similarities based on the fusion submodule to obtain a comprehensive score; and performing noise anomaly determination on the comprehensive score based on the analysis submodule to obtain the noise detection result.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the step of extracting semantic features from the logistics document image to be identified to obtain a semantic feature set includes: using an OCR algorithm to perform text recognition on the logistics document image to be identified to obtain multiple semantic fields; using a text positioning algorithm to locate the pixel position of each semantic field to obtain position information corresponding to each semantic field; and using a data structuring algorithm to perform structured integration processing on the multiple semantic fields and the position information corresponding to each semantic field to obtain the semantic feature set.

[0011] Optionally, in a sixth implementation of the first aspect of the present invention, the step of performing image tampering analysis processing on the logistics document image to be identified based on the noise detection result and the semantic verification result to obtain the image tampering detection result includes: using a fusion algorithm to perform feature fusion processing on the noise detection result and the semantic verification result to obtain a fusion feature matrix, and determining the coordinates of the modified area and the modification type based on the fusion feature matrix; calculating a tampering anomaly score based on the modified area coordinates and the modification type; obtaining a preset risk grading rule, determining the risk level of the logistics document image to be identified based on the tampering anomaly score and the risk grading rule; and integrating the modified area coordinates, the modification type, the anomaly score, and the risk level to obtain the image tampering detection result.

[0012] A second aspect of the present invention provides a logistics document image recognition device, comprising: a data acquisition module for acquiring an image of a logistics document to be recognized, a set of labeled image samples, a pre-constructed initial image recognition model, and preset semantic verification rules; a model training module for performing multiple rounds of iterative training on the initial image recognition model based on the set of labeled image samples to obtain a target image recognition model; a noise analysis module for performing noise consistency analysis on the image of the logistics document to be recognized based on the target image recognition model to obtain a noise detection result; a semantic feature extraction module for performing semantic feature extraction processing on the image of the logistics document to be recognized to obtain a semantic feature set; a semantic verification module for performing semantic verification processing on the semantic feature set based on the semantic verification rules to obtain a semantic verification result; and a tampering analysis module for performing image tampering analysis processing on the image of the logistics document to be recognized based on the noise detection result and the semantic verification result to obtain an image tampering detection result.

[0013] A third aspect of the present invention provides a logistics document image recognition device, the logistics document image recognition device comprising: a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the logistics document image recognition device to perform the various steps of the logistics document image recognition method described in any of the preceding claims.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the logistics document image recognition method described in any of the preceding claims.

[0015] In the technical solution of this invention, the following steps are taken: First, an image of the logistics document to be identified, a set of labeled image samples, a pre-constructed initial image recognition model, and preset semantic verification rules are acquired. Then, the initial image recognition model is trained iteratively multiple times based on the labeled image sample set to obtain a target image recognition model. Based on the target image recognition model, noise consistency analysis is performed on the logistics document image to be identified to obtain noise detection results. Semantic feature extraction processing is then performed on the logistics document image to be identified to obtain a semantic feature set. Next, semantic verification processing is performed on the semantic feature set based on the semantic verification rules to obtain semantic verification results. Finally, image tampering analysis processing is performed on the logistics document image to be identified based on the noise detection results and semantic verification results to obtain image tampering detection results. This invention aims to solve the technical problems of low efficiency, high false negative rate, and lack of effective identification of document image tampering traces in existing logistics document verification, which cannot adapt to the risk control needs of massive documents. It improves the accuracy and efficiency of logistics document image tampering detection and meets the risk control needs of modern logistics on a large scale. Attached Figure Description

[0016] Figure 1 This is a first flowchart of a logistics document image recognition method provided in an embodiment of the present invention; Figure 2 This is a second flowchart of the logistics document image recognition method provided in an embodiment of the present invention; Figure 3 This is a third flowchart of the logistics document image recognition method provided in the embodiments of the present invention; Figure 4 This is a fourth flowchart of the logistics document image recognition method provided in the embodiments of the present invention; Figure 5 This is a fifth flowchart of the logistics document image recognition method provided in the embodiments of the present invention; Figure 6 This is a sixth flowchart of the logistics document image recognition method provided in an embodiment of the present invention; Figure 7 The seventh flowchart of the logistics document image recognition method provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of the structure of the logistics document image recognition device provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of the logistics document image recognition device provided in an embodiment of the present invention. Detailed Implementation

[0017] This invention provides a method, apparatus, device, and storage medium for image recognition of logistics documents. In this invention, the terms "first," "second," "third," "fourth," etc. (if applicable) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the logistics document image recognition method in this invention includes: 101. Obtain the image of the logistics document to be identified, the labeled image sample set, the pre-built initial image recognition model, and the preset semantic verification rules; 102. Based on the labeled image sample set, the initial image recognition model is trained iteratively in multiple rounds to obtain the target image recognition model; In this embodiment, the labeled image sample set consists of a normal document subset, a labeled modified sample subset, and an unlabeled modified sample subset. The normal document subset consists of logistics document images that conform to the specifications and is used for learning normal noise features of the model. The labeled modified sample subset consists of document images with manually labeled tampered areas, modification types, and specific locations, and is used for calibrating the model's detection accuracy. The unlabeled modified sample subset consists of tampered document images that have not been manually labeled and is used to improve the model's generalization ability. The initial image recognition model is based on the Noiseprint++ network architecture. The training process begins by inputting a subset of normal documents into the initial model for basic training using normal noise features. This allows the model to learn the noise patterns of real, compliant documents, resulting in a normal noise baseline model. Next, a subset of labeled modified samples is input into this baseline model. By fine-tuning the model's detection head parameters, the model can accurately identify the labeled tampered areas and types, resulting in a fine-tuned image recognition model. Then, a subset of unlabeled modified samples is input into the fine-tuned model. Semi-supervised learning is used to optimize the model parameters, improving its ability to recognize unlabeled tampering behavior, resulting in a semi-supervised optimized model. Finally, the optimized model is iteratively trained multiple times based on the labeled image sample set, continuously adjusting the model weight parameters to ensure the model can accurately capture various tampering traces, ultimately forming a target image recognition model with high recognition accuracy. This training method ensures the model's accurate learning of noise features on normal documents and enhances its generalization ability through semi-supervised training. It effectively solves the problems of low efficiency and high false negative rate in traditional manual verification, while also making up for the shortcoming of single text comparison being unable to identify image tampering. This enables efficient and accurate detection of logistics document tampering, improves the standardization and rigor of document verification, and provides reliable technical support for the compliance management of logistics operations.

[0019] 103. Based on the target image recognition model, perform noise consistency analysis on the logistics document image to be recognized to obtain noise detection results; In this embodiment, the logistics document image to be identified refers to various logistics-related document images that require tamper detection and authenticity verification, covering different types such as waybills and invoices, and adapting to the input requirements of the target image recognition model. The target image recognition model includes an image preprocessing module, a scale division module, a noise extraction module, and a noise analysis module connected in sequence. When performing noise consistency analysis on the logistics document image to be identified based on the target image recognition model, the image preprocessing module first performs preprocessing such as normalization and denoising to eliminate interference caused by differences in lighting and resolution, converting the image into a standard processable format; then, the scale division module divides the preprocessed image into image region blocks of different granularities according to three preset core scales, achieving comprehensive coverage from minor modification traces to large-area spliced ​​regions; next, the noise extraction module extracts noise features from each region block, capturing noise patterns in different regions and generating corresponding noise feature maps; finally, the noise analysis module analyzes the extracted noise... Feature similarity calculation and cross-scale fusion analysis are performed. By combining the weighted fusion results of multi-scale noise similarity, abnormal noise regions are identified, and noise detection results including the location and confidence level of the noise anomaly are generated, completing the entire noise consistency analysis process. This effectively avoids the limitations of single-scale detection. Through the synergistic effect of multi-scale noise analysis and preprocessing, the accuracy and stability of noise extraction are improved, and the influence of irrelevant interference factors on the detection results is reduced. At the same time, with the orderly collaboration of various modules, the accurate identification of tampering traces in logistics document images is achieved, significantly improving the comprehensiveness and reliability of detection. This effectively solves the problems of low efficiency and high false negative rate of traditional manual verification, providing accurate and reliable noise detection basis for subsequent tampering judgment, and ensuring the rigor and efficiency of document authenticity verification.

[0020] 104. Perform semantic feature extraction processing on the image of the logistics document to be identified to obtain a semantic feature set; In this embodiment, the semantic feature set is a feature set formed after the core semantic information in the logistics document image is extracted in a structured manner. It integrates the key semantic fields of the logistics document and their corresponding location information, which can accurately reflect the core content and structural features of the document, and provide standardized and analyzable basic data support for subsequent semantic verification. The semantic feature extraction process for the logistics document image to obtain the semantic feature set must follow a standardized extraction procedure. First, an OCR algorithm is used to perform comprehensive text recognition on the logistics document image, accurately capturing various text information in the image and filtering out core semantic fields such as amount, date, signature, document number, and logistics company identification. Then, a text positioning algorithm is used to accurately locate the pixel position of each identified semantic field, clarifying the specific location information of each semantic field in the logistics document image, ensuring a one-to-one correspondence between field content and location information. Finally, a data structuring algorithm is used to systematically integrate the extracted multiple semantic fields and their corresponding location information, converting unstructured text and location data into a structured format, ultimately forming the semantic feature set. By leveraging the synergistic effects of OCR recognition, text localization, and structured integration, the accuracy and standardization of semantic information extraction are improved. This provides reliable data support for subsequent logical verification of document information based on semantic verification rules, effectively avoiding verification errors caused by incomplete or non-standard semantic information extraction. Furthermore, it enhances the comprehensiveness and accuracy of logistics document tampering detection and ensures the rigor of document verification.

[0021] 105. Perform semantic verification processing on the semantic feature set based on the semantic verification rules to obtain the semantic verification result; In this embodiment, the semantic verification rules are set in combination with the actual business needs of logistics documents, industry standards and the coding standards of logistics companies. The core is to formulate the rules around the accuracy, logic and standardization of key information in logistics documents, so as to ensure that the verification standards fit the actual application scenarios and provide clear judgment basis for semantic verification. Semantic validation is performed on the semantic feature set based on semantic validation rules. The preset semantic validation rules include: the format and logical relationship of key information such as amount, date, and signature in logistics documents should conform to industry standards; the document number should be unique; and the logistics company identification should be consistent with the actual business. These rules are formulated based on statistical analysis of a large number of logistics document samples and the experience of industry experts. The specific processing procedure for semantic validation is as follows: First, extract various core fields from the semantic feature set. Then, in combination with the preset semantic validation rules, verify the consistency of uppercase and lowercase numbers for the amount, confirm whether the date format conforms to the specifications and whether the time range is reasonable, analyze the pixel superposition relationship between the signature area and the printed text area, determine whether there is an abnormal overlay order, and verify whether the coding rules of the document number conform to the specifications of the corresponding logistics company. During the validation process, the field position information in the semantic feature set is synchronously associated, and logical anomalies, format errors, or position anomalies are recorded. Finally, the semantic validation result is formed. The semantic verification results mainly include two aspects: first, clearly marking the fields that conform to the semantic rules and their corresponding positions; and second, recording in detail the semantic anomalies found during the verification process, including the specific content, location, and anomaly type of the anomaly fields. This provides semantic support for subsequent judgments on whether documents have been tampered with, effectively compensating for the limitations of relying solely on noise detection. By accurately verifying the core semantic information of documents, the comprehensiveness and accuracy of judging the authenticity of logistics documents are further improved, avoiding situations where semantic logic anomalies are ignored due to focusing only on image noise. This ensures that the detection of document tampering is more targeted and reliable, while standardizing the verification criteria for semantic information, reducing verification deviations caused by semantic errors, and providing more comprehensive protection for the authenticity verification of logistics documents.

[0022] 106. Based on the noise detection results and the semantic verification results, perform image tampering analysis processing on the logistics document image to be identified to obtain image tampering detection results.

[0023] In this embodiment, the noise detection result is output by the target image recognition model, including noise features at multiple scales, coordinates of suspected tampered areas, and confidence levels, reflecting potential tampering traces in the document image. The semantic verification result extracts core semantic fields from the document and, combined with preset semantic rules, verifies the logic and format compliance of the fields to identify semantic anomalies. During tampering analysis, a fusion algorithm is first used to fuse the noise detection result and the semantic verification result, generating a fusion feature matrix. This matrix accurately locates the specific coordinates of the tampered area, identifies the tampering type, and calculates a tampering anomaly score to quantify the degree of tampering. Combined with preset risk grading rules, the risk level is determined based on the anomaly score. Finally, the coordinates of the tampered area, the tampering type, the anomaly score, and the risk level are integrated to form a complete image tampering detection result. The image tampering detection result specifically includes the pixel coordinates of the tampered area, the tampering type, the tampering anomaly score, and the corresponding risk level, clearly showing whether the document has been tampered with, the specific location and type of tampering, providing direct evidence for document authenticity verification. This analytical approach, which integrates noise detection and semantic verification, effectively overcomes the limitations of single noise detection or single semantic verification. It avoids missed or false detections caused by a single detection dimension, significantly improves the accuracy and comprehensiveness of tamper detection, and makes the detection results more practical. It can directly provide scientific and reliable support for determining the authenticity of logistics documents, helping enterprises achieve efficient and accurate document risk control management and reduce operational risks caused by document tampering.

[0024] Please see Figure 2 In this embodiment of the invention, the labeled image sample set includes a normal document subset, a labeled modified sample subset, and an unlabeled modified sample subset; the step of performing multiple rounds of iterative training on the initial image recognition model based on the labeled image sample set to obtain the target image recognition model includes: 201. Input the subset of normal documents into the initial image recognition model for basic training of normal noise features to obtain a normal noise baseline model; 202. Input the labeled modified sample subset into the normal noise benchmark model for fine-tuning of the detection head to obtain the fine-tuned image recognition model; 203. Input the unlabeled modified sample subset into the fine-tuned image recognition model for semi-supervised optimization training to obtain a semi-supervised optimized model; 204. The semi-supervised optimization model is trained iteratively for multiple rounds based on the labeled image sample set to obtain the target image recognition model.

[0025] In this embodiment, the labeled image sample set includes a normal document subset, a labeled modified sample subset, and an unlabeled modified sample subset. The normal document subset is used for basic training of normal noise features. Based on the Noiseprint++ network architecture, the noise patterns of normal documents are used as training samples, allowing the model to learn the noise features of real, compliant documents, grasp the noise distribution patterns in normal scenarios, and thus form a normal noise baseline model. This step is the process of basic training of normal noise features, aiming to enable the model to clearly define the noise feature baseline of normal documents, providing a comparative basis for subsequent tamper detection. The labeled modified sample subset is used for fine-tuning the detection head of the model. At this point, the model already possesses basic noise recognition capabilities. By fine-tuning the weight parameters of the detection head, the model can accurately identify the labeled modified areas, clarify the modification type and location, and achieve accurate identification of the labeled samples. The unlabeled modified sample subset is used for semi-supervised optimization training. It is input into the fine-tuned model and uses a pseudo-label mechanism. The pseudo-label mechanism refers to using the fine-tuned model to infer and predict the unlabeled modified samples, automatically identifying the tampered regions and modification types in the samples and generating virtual label information. In this way, the unlabeled samples are transformed into effective data that can participate in training. The model can learn the tampering features autonomously without human annotation, further optimize the model parameters, improve the model's adaptability to unlabeled tampering scenarios, make up for the limitations of training with single labeled samples, and enable the model to cope with more complex tampering situations in real applications. The entire training process first establishes a noise baseline using a subset of normal documents, then calibrates the detection accuracy using a subset of labeled modified samples, and finally improves the model's generalization ability using a subset of unlabeled modified samples. This iterative optimization process gradually completes the model's optimization, ultimately resulting in a target image recognition model capable of accurately identifying tampering behavior. This layered training approach ensures accurate identification of known tampering scenarios while enhancing the model's adaptability to unknown tampering situations. It effectively solves the problems of low efficiency and high false negative rate in traditional manual verification, while balancing detection accuracy and generalization ability. This ensures the model can stably and accurately identify document tampering in practical applications, providing reliable technical support for the authenticity verification of logistics documents, significantly reducing the workload of manual verification, improving the efficiency and accuracy of document review, and ensuring the compliant operation of logistics businesses.

[0026] Please see Figure 3 In this embodiment of the invention, the step of performing multiple rounds of iterative training on the semi-supervised optimization model based on the labeled image sample set to obtain the target image recognition model includes: 301. Obtain the actual noise detection results corresponding to the labeled image sample set; 302. Input the labeled image sample set into the semi-supervised optimization model for noise consistency analysis to obtain the predicted noise detection results; 303. Calculate the model loss based on the actual noise detection results and the predicted noise detection results; 304. Based on the model loss, the weight parameters of the semi-supervised optimization model are iteratively updated using the backpropagation algorithm; 305. When the preset iteration stopping condition is met, output the target image recognition model.

[0027] In this embodiment, the actual noise detection results corresponding to the labeled image sample set are obtained. These results are noise anomaly baseline data determined based on the manually labeled coordinates and modification types of the tampered regions in the labeled image sample set, combined with the actual noise distribution patterns in the images. This includes the actual location of the noise anomaly regions and the noise anomaly confidence level. The labeled image sample set is input into a semi-supervised optimization model for noise consistency analysis. The model's output of noise anomaly regions and anomaly confidence levels, obtained through multi-scale noise feature extraction and similarity calculation, constitutes the predicted noise detection results. The model loss is calculated based on the actual and predicted noise detection results, quantifying the deviation between the model's predicted data and the actual noise anomaly data. Then, the weight parameters of the semi-supervised optimization model are iteratively updated using the backpropagation algorithm based on the model loss, continuously reducing the prediction deviation. The semi-supervised optimization model has already completed normal noise feature learning and detector head fine-tuning in the early stages, and has completed semi-supervised training using a pseudo-label mechanism. This mechanism inputs unlabeled modified samples into the fine-tuned model, uses the high-confidence prediction results output by the model as pseudo-labels, mixes them with labeled samples, and then fine-tunes the detector head again, thereby strengthening the model's ability to identify various tampered noise features. During the iterative update process, performance metrics such as detection accuracy and recall on the validation set are calculated. Hyperparameters such as the learning rate and regularization coefficient are adjusted based on changes in these metrics, and the training and optimization process is repeated. The preset stopping conditions are either the model loss value converging to a stable range and no longer decreasing, or the number of iterations reaching a set upper limit. The iteration terminates and the target image recognition model is output when either condition is met. This iterative optimization method uses real noise detection results as a supervised benchmark, combining semi-supervised learning and backpropagation algorithms to achieve fine-tuning of model parameters. This ensures that the model's predictions of noise anomalies closely match the noise distribution characteristics in real tampering scenarios. Furthermore, multi-dimensional performance metrics and hyperparameter tuning improve model stability. Simultaneously, a pseudo-label mechanism expands the training data, reducing manual annotation costs and enhancing the model's generalization ability to different tampering scenarios. The resulting target image recognition model can accurately identify noise anomalies and tampering traces in logistics document images, effectively improving the accuracy and robustness of document tampering detection and providing solid technical support for verifying the authenticity of logistics documents.

[0028] Please see Figure 4In this embodiment of the invention, the target image recognition model includes an image preprocessing module, a scale division module, a noise extraction module, and a noise analysis module connected in sequence; the step of performing noise consistency analysis on the logistics document image to be recognized based on the target image recognition model to obtain noise detection results includes: 401. Based on the image preprocessing module, perform image preprocessing on the logistics document image to be identified to obtain a preprocessed image; 402. Based on the scale division module, the preprocessed image is divided into multiple image regions to obtain multiple image region blocks; 403. Based on the noise extraction module, noise extraction processing is performed on each of the image regions to obtain a noise feature map corresponding to each of the image regions; 404. Based on the noise analysis module, perform noise analysis on each of the noise feature maps to obtain the noise detection results.

[0029] In this embodiment, the target image recognition model consists of an image preprocessing module, a scale segmentation module, a noise extraction module, and a noise analysis module connected sequentially. The entire model is built upon the Noiseprint++ basic model framework. Each module works collaboratively to perform noise consistency analysis on the logistics document image to be recognized and outputs the noise detection results. The image preprocessing module performs normalization and denoising operations on the input logistics document image to be recognized. Through standardization, it eliminates interference factors such as uneven lighting and resolution differences generated during image acquisition. Simultaneously, it converts the image into a standard format suitable for model computation, removing irrelevant interference for subsequent noise feature extraction and ensuring the purity of the extracted noise features. The scale segmentation module divides the preprocessed image into regions according to a preset scale. For example, it performs image segmentation with fine granularity of 4×4 pixel blocks, medium granularity of 16×16 pixel blocks, and coarse granularity of 64×64 pixel blocks, achieving full-range detection coverage from minor tampering traces to large-area spliced ​​tampering areas, forming multiple image region blocks of different scales. The noise extraction module extracts noise features from each segmented image region. Utilizing a multi-level feature extraction structure built on a convolutional neural network, the first convolutional layer captures basic pixel noise features. Subsequent layers progressively increase the number of convolutional kernels and the receptive field, extracting sensor noise pattern features from low to high dimensions. A residual connection structure is introduced to avoid the gradient vanishing problem that occurs during deep network training. Finally, a corresponding noise feature map is generated for each image region. This noise feature map is a numerical representation of the sensor noise distribution, intensity, and texture features within each region, comprehensively recording the inherent patterns of regional noise and providing intuitive data for subsequent noise consistency determination. The noise analysis module analyzes the noise feature maps at each scale, ultimately outputting the coordinates and confidence scores of areas with inconsistent noise, forming a complete noise detection result. External interference is eliminated through preprocessing, and multi-scale segmentation achieves comprehensive tamper detection. Refined noise extraction and cross-scale fusion analysis significantly improve the accuracy of noise anomaly identification. Meanwhile, the residual structure ensures the stability of model training. It can effectively identify noise anomalies caused by various tamperings such as splicing, copying and moving in logistics document images, providing accurate and reliable noise feature basis for subsequent image tampering analysis, and significantly improving the completeness and accuracy of logistics document tamper detection.

[0030] Please see Figure 5 In this embodiment of the invention, the noise analysis module includes a noise similarity calculation submodule, a fusion submodule, and an analysis submodule connected in sequence; the noise analysis performed on each noise feature map based on the noise analysis module to obtain the noise detection result includes: 501. Calculate the noise similarity of each noise feature map based on the noise similarity calculation submodule; 502. Based on the fusion submodule, perform weighted fusion calculation on multiple noise similarities to obtain a comprehensive score; 503. Based on the analysis submodule, noise anomaly determination is performed on the comprehensive score to obtain the noise detection result.

[0031] In this embodiment, the noise analysis module consists of a noise similarity calculation submodule, a fusion submodule, and an analysis submodule connected sequentially. The noise feature map is a numerical representation of the sensor noise distribution, intensity, and texture features in image regions at various scales, fully recording the inherent noise patterns of different regions and providing basic data support for noise analysis. The core function of the noise similarity calculation submodule is to perform targeted processing on each noise feature map, calculating the similarity between noise features in each region one by one, constructing a similarity matrix, and capturing the differences in noise patterns in different regions through matrix analysis, accurately identifying suspicious regions with inconsistent noise distribution. This process is consistent with the core logic of noise consistency detection in the Noiseprint++ basic model framework, laying the foundation for subsequent fusion analysis. The fusion submodule is responsible for integrating the multi-scale noise similarity calculation results. Combining the preset weight allocation rules, it performs weighted fusion calculations on the noise similarity results at the three scales: fine-grained, medium-grained, and coarse-grained. For example, the weight for fine-grained can be set to 0.5, for medium-grained to 0.3, and for coarse-grained to 0.2. This weight allocation highlights the ability of fine-grained to accurately capture minute tampering traces while also taking into account the coverage of large-area splicing tampering by medium and coarse-grained. Finally, it generates a comprehensive score that integrates multi-scale information, effectively avoiding the limitations of single-scale analysis. The analysis submodule receives the comprehensive score output by the fusion submodule and performs noise anomaly judgment on the comprehensive score according to the preset judgment criteria. The preset threshold is a critical score value obtained by statistical analysis of the noise distribution characteristics of a large number of real logistics document samples. This threshold is a normalized value with a value range between 0 and 1. In the logistics document tampering detection scenario, it is usually set in the range of 0.6 to 0.8, which can effectively distinguish between the natural noise fluctuations of normal documents and the noise anomalies caused by tampering operations. Areas with comprehensive scores higher than the preset threshold are marked as suspicious modification areas. At the same time, the coordinate information and confidence level of all suspicious areas are integrated to form a complete noise detection result, realizing the accurate location and quantitative evaluation of noise anomaly areas. The entire noise analysis process relies on the orderly collaboration of three sub-modules, organically combining multi-scale noise feature extraction, similarity calculation, cross-scale fusion, and anomaly detection. This ensures both the comprehensiveness of noise analysis and the accuracy of noise anomaly identification. It can effectively capture the differences in noise patterns caused by various tampering behaviors such as splicing, copying, and moving in logistics document images, providing a reliable noise-level basis for subsequent image tampering analysis. This significantly improves the accuracy and robustness of logistics document tampering detection, while avoiding the problems of missed detections and false detections caused by single-scale or single-dimensional analysis. It further enhances the detection performance of the target image recognition model and provides solid technical support for the verification of the authenticity of logistics documents.

[0032] Please see Figure 6 In this embodiment of the invention, the semantic feature extraction process performed on the logistics document image to be identified to obtain a semantic feature set includes: 601. Use an OCR algorithm to perform text recognition on the image of the logistics document to be recognized, and obtain multiple semantic fields; 602. Use a text positioning algorithm to locate the pixel position of each semantic field to obtain the position information corresponding to each semantic field; 603. The semantic feature set is obtained by using a data structuring algorithm to perform structured integration processing on multiple semantic fields and the location information corresponding to each semantic field.

[0033] In this embodiment, semantic fields refer to key fields in logistics documents that have core business significance and are directly related to the authenticity and compliance of the documents. These include amounts, dates, signatures, document numbers, and logistics company identifiers. These fields are the core objects of semantic verification and tamper detection, directly determining the validity of the documents. The location information corresponding to each semantic field refers to the specific pixel coordinate range of each semantic field in the logistics document image. Typically, the upper left corner of the image is used as the origin, clearly defining the upper left and lower right pixel coordinates of the area where the field is located, clearly defining the spatial distribution of each semantic field in the image, and providing a spatial basis for subsequent semantic rule verification. In the specific extraction process, an OCR algorithm is first used to perform comprehensive text recognition on the logistics document image. Selectable OCR algorithms include Tesseract OCR, which can accurately recognize printed and handwritten text in the document, effectively filtering interference from image noise, filtering out core semantic fields from complex document images, eliminating irrelevant and redundant text, and finally obtaining multiple semantic fields. Subsequently, text localization algorithms are used to locate the pixel position of each semantic field. Commonly used text localization algorithms include EAST text detection algorithm and CTPN algorithm. These algorithms can accurately capture the spatial position of each semantic field in the image. By calibrating the pixel coordinates, the position information corresponding to each semantic field is obtained, ensuring the accurate association between field content and spatial position. Finally, data structuring algorithms are used to structurally integrate multiple semantic fields and their corresponding position information. JSON structuring algorithms, XML parsing algorithms, etc., can be used to convert unstructured text content and position data into standardized structured data formats according to preset specifications, ultimately forming a semantic feature set. The semantic feature set is a structured feature set that integrates all core semantic fields of the logistics document and their corresponding position information. It not only contains the specific text content of each semantic field but also associates it with its spatial position information in the image. It can clearly present the core information and field distribution patterns of the document, providing standardized and analyzable basic data support for subsequent logical verification of document information based on semantic verification rules. The selected algorithms are all adapted to the recognition scenarios of logistics document images, and can cope with the interference caused by differences in lighting and resolution, improving the stability and accuracy of semantic information extraction. At the same time, the structured and integrated semantic feature set greatly reduces the complexity of subsequent semantic verification, ensuring that semantic verification can be carried out efficiently and accurately. This provides a reliable semantic basis for detecting tampering of logistics documents, further improving the rigor and comprehensiveness of document verification, helping to avoid verification errors caused by non-standard or incomplete semantic information extraction, and ensuring the compliance management of logistics business.

[0034] Please see Figure 7 In this embodiment of the invention, the step of performing image tampering analysis processing on the logistics document image to be identified based on the noise detection result and the semantic verification result to obtain the image tampering detection result includes: 701. A fusion algorithm is used to perform feature fusion processing on the noise detection results and the semantic verification results to obtain a fusion feature matrix, and the coordinates of the modified region and the modification type are determined based on the fusion feature matrix; 702. Calculate the tampering anomaly score based on the modified area coordinates and the modification type; 703. Obtain preset risk grading rules, and determine the risk level of the logistics document image to be identified based on the tampering anomaly score and the risk grading rules; 704. Integrate the coordinates of the modified area, the modification type, the anomaly score, and the risk level to obtain the image tampering detection result.

[0035] In this embodiment, the fusion algorithm can employ a weighted fusion algorithm or a principal component analysis fusion algorithm, or a neural network-based fusion algorithm, such as a convolutional neural network fusion algorithm or an attention mechanism fusion algorithm. These algorithms can effectively integrate the core features of noise detection results and semantic verification results, avoiding the limitations of a single feature dimension. Through weight allocation and collaborative analysis of the two types of features, a fusion feature matrix that comprehensively reflects the document tampering situation is generated. Based on this fusion feature matrix, the coordinates of the modified area corresponding to the tampering behavior can be accurately located, and the specific type of tampering can be identified, covering common tampering types such as amount modification, date tampering, signature forgery, and area splicing, thus achieving accurate characterization of tampered information.

[0036] In this embodiment, the tampering anomaly score is a numerical value calculated based on the coordinates of the modified area and the type of modification. It is used to measure the severity of document tampering. The size of the modified area, the importance of its location, and the severity of the modification type all affect the score result. Tampering of core fields such as amount and signature will correspond to higher score weights. The preset risk classification rule is a classification standard preset by combining the actual business needs of logistics document tampering detection, the statistical analysis results of a large number of real tampering samples, and the severity of tampering behavior. Its core is to achieve hierarchical control of document tampering risk through the division of tampering anomaly scores. The specific classification standard can be set as follows: tampering anomaly score greater than or equal to 80 is high risk, tampering anomaly score greater than or equal to 50 or less than 80 is medium risk, and tampering anomaly score less than 50 is low risk. The preset classification rule not only refers to the actual risk level corresponding to different scores in historical tampering samples, but also combines the compliance requirements of the logistics industry for document authenticity control, ensuring that the risk classification is consistent with the actual application scenario.

[0037] In this embodiment, the specific process of determining the risk level of the logistics document image to be identified based on the tampering anomaly score and risk grading rules involves comparing the calculated tampering anomaly score with a preset grading threshold one by one. If the tampering anomaly score reaches or exceeds 80 points, it is judged as high risk; such documents usually have a serious possibility of tampering and require key control. If the tampering anomaly score is between 50 and 79 points, it is judged as medium risk; such documents have a certain possibility of tampering and require further verification. If the tampering anomaly score is below 50 points, it is judged as low risk; such documents have a low possibility of tampering and can be reviewed according to the regular process. After completing the risk level determination, the modified area coordinates, modification type, tampering anomaly score, and determined risk level are systematically integrated to form a complete image tampering detection result. This result can clearly present the core information and risk level of document tampering, providing a clear basis for subsequent document processing.

[0038] By employing a fusion algorithm to achieve collaborative analysis of noise and semantic features, the limitations of single noise detection or semantic verification are effectively overcome, improving the accuracy and comprehensiveness of tamper detection. Pre-set risk grading rules align with actual business needs, enabling tiered control of tamper risks. This facilitates targeted audits by staff, with high-risk documents triggering manual review and low-risk documents undergoing routine audits, significantly improving the efficiency and rigor of document review. Simultaneously, the quantitative method of tamper anomaly scoring makes risk assessment more objective, avoiding subjective biases in manual judgment. The selection of the fusion algorithm also enhances the rationality and stability of feature integration. Ultimately, this ensures that image tamper detection results provide reliable support for verifying the authenticity of logistics documents, helping enterprises avoid operational risks caused by document tampering and ensuring the compliant and orderly operation of logistics businesses.

[0039] The above describes the logistics document image recognition method in the embodiments of the present invention. The following describes the logistics document image recognition device in the embodiments of the present invention. Please refer to [link / reference]. Figure 8 One embodiment of the logistics document image recognition device in this invention includes: Data acquisition module 801: used to acquire images of logistics documents to be identified, labeled image sample sets, pre-built initial image recognition models, and preset semantic verification rules; Model training module 802: used to perform multiple rounds of iterative training on the initial image recognition model based on the labeled image sample set to obtain the target image recognition model; Noise analysis module 803: used to perform noise consistency analysis on the logistics document image to be identified based on the target image recognition model, and obtain noise detection results; Semantic feature extraction module 804: used to perform semantic feature extraction processing on the logistics document image to be identified, and obtain a semantic feature set; Semantic verification module 805: used to perform semantic verification processing on the semantic feature set based on the semantic verification rules, and obtain semantic verification results; Tampering analysis module 806: used to perform image tampering analysis processing on the logistics document image to be identified based on the noise detection result and the semantic verification result, and obtain the image tampering detection result.

[0040] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0041] above Figure 8 The logistics document image recognition device in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The logistics document image recognition device in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0042] Figure 9 This is a schematic diagram of the structure of a logistics document image recognition device 900 provided in an embodiment of the present invention. The logistics document image recognition device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) for storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the logistics document image recognition device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute the series of instruction operations in the storage media 930 on the logistics document image recognition device 900 to implement the steps of the logistics document image recognition method provided in the above-described method embodiments.

[0043] The logistics document image recognition device 900 may also include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 9 The illustrated structure of the logistics document image recognition device does not constitute a limitation on the logistics document image recognition device. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0044] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the logistics document image recognition method.

[0045] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0046] If the integrated unit 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 the present invention, in essence, or the part that contributes to the prior art, or all or 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 described in the various embodiments of the present 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.

[0047] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for image recognition of logistics documents, characterized in that, include: Acquire the image of the logistics document to be identified, the set of labeled image samples, the pre-built initial image recognition model, and the preset semantic verification rules; The initial image recognition model is trained iteratively multiple times based on the labeled image sample set to obtain the target image recognition model. Based on the target image recognition model, noise consistency analysis is performed on the logistics document image to be recognized to obtain noise detection results; The semantic feature extraction process is performed on the image of the logistics document to be identified to obtain a semantic feature set; Based on the semantic verification rules, the semantic feature set is subjected to semantic verification processing to obtain the semantic verification result; Based on the noise detection results and the semantic verification results, the image of the logistics document to be identified is subjected to image tampering analysis processing to obtain the image tampering detection results.

2. The logistics document image recognition method according to claim 1, characterized in that, The labeled image sample set includes a subset of normal documents, a subset of labeled modified documents, and a subset of unlabeled modified documents; the step of performing multiple rounds of iterative training on the initial image recognition model based on the labeled image sample set to obtain the target image recognition model includes: The normal document subset is input into the initial image recognition model for normal noise feature-based training to obtain a normal noise baseline model; The labeled modified sample subset is input into the normal noise baseline model for fine-tuning of the detection head to obtain the fine-tuned image recognition model. The unlabeled modified sample subset is input into the fine-tuned image recognition model for semi-supervised optimization training to obtain a semi-supervised optimized model. The semi-supervised optimization model is trained iteratively through multiple rounds based on the labeled image sample set to obtain the target image recognition model.

3. The logistics document image recognition method according to claim 2, characterized in that, The process of performing multiple rounds of iterative training on the semi-supervised optimization model based on the labeled image sample set to obtain the target image recognition model includes: Obtain the actual noise detection results corresponding to the labeled image sample set; The labeled image sample set is input into the semi-supervised optimization model for noise consistency analysis to obtain the predicted noise detection results. The model loss is calculated based on the actual noise detection results and the predicted noise detection results. Based on the model loss, the weight parameters of the semi-supervised optimization model are iteratively updated using the backpropagation algorithm; When the preset iteration stopping condition is met, the target image recognition model is output.

4. The logistics document image recognition method according to claim 1, characterized in that, The target image recognition model includes an image preprocessing module, a scale division module, a noise extraction module, and a noise analysis module connected in sequence. The step of performing noise consistency analysis on the logistics document image to be identified based on the target image recognition model to obtain noise detection results includes: The image preprocessing module performs image preprocessing on the logistics document image to be identified to obtain a preprocessed image. Based on the scale segmentation module, the preprocessed image is divided into multiple image regions to obtain multiple image region blocks. Based on the noise extraction module, noise extraction processing is performed on each of the image regions to obtain a noise feature map corresponding to each of the image regions. Based on the noise analysis module, noise analysis is performed on each of the noise feature maps to obtain the noise detection results.

5. The logistics document image recognition method according to claim 4, characterized in that, The noise analysis module includes a noise similarity calculation submodule, a fusion submodule, and an analysis submodule connected in sequence. The noise analysis module performs noise analysis on each of the noise feature maps to obtain the noise detection results, including: The noise similarity calculation submodule calculates the noise similarity of each noise feature map. Based on the fusion submodule, a weighted fusion calculation is performed on multiple noise similarities to obtain a comprehensive score; Based on the analysis submodule, noise anomaly determination is performed on the comprehensive score to obtain the noise detection result.

6. The logistics document image recognition method according to claim 1, characterized in that, The semantic feature extraction process performed on the image of the logistics document to be identified yields a semantic feature set, including: The OCR algorithm is used to perform text recognition on the logistics document image to be identified, and multiple semantic fields are obtained; A text localization algorithm is used to locate the pixel position of each semantic field to obtain the position information corresponding to each semantic field; The semantic feature set is obtained by using a data structuring algorithm to structurally integrate multiple semantic fields and the location information corresponding to each semantic field.

7. The logistics document image recognition method according to claim 1, characterized in that, The image tampering analysis processing of the logistics document image to be identified, based on the noise detection result and the semantic verification result, yields the image tampering detection result, including: A fusion algorithm is used to perform feature fusion processing on the noise detection results and the semantic verification results to obtain a fusion feature matrix, and the coordinates of the modified region and the modification type are determined based on the fusion feature matrix; Calculate the tampering anomaly score based on the modified region coordinates and the modification type; Obtain preset risk grading rules, and determine the risk level of the logistics document image to be identified based on the tampering anomaly score and the risk grading rules; The image tampering detection result is obtained by integrating the coordinates of the modified area, the modification type, the anomaly score, and the risk level.

8. A logistics document image recognition device, characterized in that, include: Data acquisition module: used to acquire images of logistics documents to be identified, labeled image sample sets, pre-built initial image recognition models, and preset semantic verification rules; Model training module: used to perform multiple rounds of iterative training on the initial image recognition model based on the labeled image sample set to obtain the target image recognition model; Noise analysis module: used to perform noise consistency analysis on the logistics document image to be identified based on the target image recognition model, and obtain noise detection results; Semantic feature extraction module: used to perform semantic feature extraction processing on the logistics document image to be identified, and obtain a semantic feature set; Semantic verification module: used to perform semantic verification processing on the semantic feature set based on the semantic verification rules, and obtain semantic verification results; Tampering analysis module: used to perform image tampering analysis processing on the logistics document image to be identified based on the noise detection results and the semantic verification results, and obtain image tampering detection results.

9. A logistics document image recognition device, characterized in that, The logistics document image recognition device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the logistics document image recognition device to perform the various steps of the logistics document image recognition method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the logistics document image recognition method as described in any one of claims 1-7.

Images