Logistics terminal pick-up code identification method, device and equipment and storage medium

Abstract

The present application relates to the technical field of intelligent logistics, and more particularly to a logistics end pick-up code recognition method, device and equipment and storage medium, the method first acquires a mobile device image, pre-processes the mobile device image to obtain a pre-processed image, then performs keyword positioning processing on the pre-processed image to obtain a plurality of initial text regions, respectively performs spatial correlation matching filtering processing on each initial text region based on the pre-processed image to obtain a plurality of effective text regions, then performs quality filtering processing on all effective text regions to obtain a to-be-judged code set, and finally inputs the to-be-judged code set into a pre-trained pick-up code recognition model for pick-up code recognition processing to obtain an effective pick-up code, aiming to solve the problems of traditional OCR recognition of mixed text difficult to distinguish various codes, high cost of manual screening and low recognition accuracy, realize automatic extraction of pick-up code, reduce maintenance cost, and improve end pick-up recognition efficiency and accuracy.

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 identifying pickup codes at the end of logistics. Background Technology

[0002] The pickup code is the core identity credential connecting the logistics back-end system with consumer pickup verification. Its recognition and processing efficiency directly determines the user service experience of express delivery stations and last-mile delivery points. Currently, the industry generally relies on manual verification to complete pickup operations. Staff need to check the pickup information displayed on each consumer's mobile phone screen one by one. During peak periods of express delivery, the manual processing speed has a significant bottleneck and is difficult to cope with the concentrated flow of customers picking up packages at the store. The channels for pushing online pickup information are relatively scattered. Various e-commerce applications, SMS notifications, and WeChat mini-programs all push notifications containing pickup codes. The screen contains a large amount of irrelevant text content. If traditional optical character recognition technology is used to fully recognize the entire screen image, it will produce a massive amount of meaningless text content. Staff still need to manually identify and extract the real pickup code, which will increase the manual operation cost. General optical character recognition algorithms have not been specifically adapted and optimized for mobile phone screen shooting scenarios. They are affected by factors such as screen reflection, image distortion, and stripe noise, which significantly reduces the accuracy of text recognition. Character recognition models alone can only output simple text character sequences, making it difficult to autonomously distinguish similar combinations of numbers and letters such as pickup codes, tracking numbers, order numbers, and SMS verification codes. The industry's conventional solution is to write fixed regular expression matching rules to complete the filtering. However, with different platform interface updates and the emergence of new encoding formats, the rules need to be manually modified, resulting in high long-term system maintenance costs and weak overall scalability and adaptability. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, device, equipment and storage medium for identifying pickup codes at the end of logistics, which aims to solve the problems of traditional OCR recognition of miscellaneous text, difficulty in distinguishing various codes, high cost of manual screening and low recognition accuracy, so as to realize automatic extraction of pickup codes, reduce maintenance costs and improve the efficiency and accuracy of end-of-line pickup recognition.

[0004] The first aspect of the present invention provides a method for identifying pickup codes at the end of logistics, comprising: acquiring an image of a mobile device; preprocessing the image of the mobile device to obtain a preprocessed image; performing keyword localization processing on the preprocessed image to obtain multiple initial text regions; performing spatial association matching filtering processing on each of the initial text regions based on the preprocessed image to obtain multiple valid text regions; performing quality filtering processing on all the valid text regions to obtain a set of codes to be determined; and inputting the set of codes to be determined into a pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the preprocessing of the mobile device image to obtain a preprocessed image includes: adjusting the exposure parameters of the mobile device image using an automatic exposure compensation algorithm to obtain a light-adjusted image; converting the light-adjusted image into a frequency domain image using a moiré filtering algorithm, locating the frequency interference signal of the frequency domain image, and filtering out noise signals from the frequency interference signal to obtain an effective image; performing pixel reconstruction correction on the effective image using a perspective distortion correction algorithm to obtain a corrected image; and performing size normalization on the corrected image using a normalization algorithm to obtain the preprocessed image.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the step of performing keyword localization processing on the preprocessed image to obtain multiple initial text regions includes: calling a pre-trained lightweight object detection model, the lightweight object detection model including a feature extraction module, a keyword matching module, and a region selection module connected in sequence; performing feature extraction processing on the preprocessed image based on the feature extraction module to obtain a global feature image; performing feature matching retrieval processing on the global feature image based on the keyword matching module according to preset logistics text keywords to obtain keyword feature matching results; and determining the initial text regions based on the keyword feature matching results using the region selection module.

[0007] Optionally, in a third implementation of the first aspect of the present invention, based on the preprocessed image, spatial association matching and filtering processing is performed on each of the initial text regions to obtain multiple effective text regions, including: inputting the preprocessed image into a pre-trained image segmentation model for UI component semantic segmentation processing to obtain UI function area contour coordinates; using a text localization algorithm to perform text line localization processing on each of the initial text regions to obtain multiple candidate text line regions; using a multi-feature spatial association analysis algorithm to perform filtering and scoring processing on the multiple candidate text line regions according to the UI function area contour coordinates to obtain multiple target text regions and a confidence score corresponding to each target text region; obtaining a preset confidence threshold, comparing each confidence score with the confidence threshold, and taking the target text region with a confidence score higher than the confidence threshold as the effective text region.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the step of performing quality screening on all the effective text regions to obtain a set of codes to be determined includes: cropping the preprocessed image based on all the effective text regions to obtain multiple candidate region images; inputting each candidate region image into a pre-trained character recognition model for character extraction to obtain multiple original text sequences and a comprehensive text confidence score corresponding to each original text sequence; obtaining a preset text confidence score threshold; comparing each comprehensive text confidence score with the text confidence score threshold; and using the original text sequences whose comprehensive text confidence scores are higher than or equal to the text confidence score threshold as target text sequences; and integrating all the target text sequences to obtain the set of codes to be determined.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the character recognition model includes a character feature extraction module, a sequence decoding module, and a confidence calculation module connected in sequence; the step of inputting each candidate region image into the pre-trained character recognition model for character extraction processing to obtain multiple original text sequences and a comprehensive text confidence score corresponding to each original text sequence includes: inputting each candidate region image into the character feature extraction module for image character feature extraction processing to obtain multiple text feature images; inputting each text feature image into the sequence decoding module for character sequence decoding processing to obtain multiple original text sequences; and calculating the confidence score of all characters in each original text sequence based on the confidence calculation module to obtain a comprehensive text confidence score corresponding to each original text sequence.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the pickup code recognition model includes a prompt word loading module, an auxiliary reasoning module, and an encoding classification and determination module connected in sequence; the step of inputting the set of codes to be determined into the pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code includes: performing standardized prompt word binding processing on the set of codes to be determined based on the prompt word loading module to obtain a text set to be reasoned; performing semantic reasoning processing on the text set to be reasoned based on the auxiliary reasoning module to obtain a set of encoded semantic features; and performing encoding type differentiation processing on the set of encoded semantic features based on the encoding classification and determination module to obtain the valid pickup code.

[0011] A second aspect of the present invention provides a logistics terminal pickup code recognition device, comprising: a preprocessing module for acquiring a mobile device image and preprocessing the mobile device image to obtain a preprocessed image; a keyword localization module for performing keyword localization processing on the preprocessed image to obtain multiple initial text regions; an association matching module for performing spatial association matching and filtering processing on each of the initial text regions based on the preprocessed image to obtain multiple valid text regions; a quality filtering module for performing quality filtering processing on all the valid text regions to obtain a set of codes to be judged; and a pickup code recognition module for inputting the set of codes to be judged into a pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code.

[0012] A third aspect of the present invention provides a logistics terminal pickup code identification device, the logistics terminal pickup code identification device comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the logistics terminal pickup code identification device to perform each step of the logistics terminal pickup code identification method described above.

[0013] 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 end-of-line pickup code identification method described above.

[0014] In the technical solution of this invention, a mobile device image is first acquired, and the mobile device image is preprocessed to obtain a preprocessed image. Then, keyword localization processing is performed on the preprocessed image to obtain multiple initial text regions. Based on the preprocessed image, spatial association matching and filtering processing is performed on each initial text region to obtain multiple valid text regions. Next, quality filtering processing is performed on all valid text regions to obtain a set of codes to be judged. Finally, the set of codes to be judged is input into a pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code. This invention aims to solve the problems of traditional OCR recognition of miscellaneous text, difficulty in distinguishing various codes, high cost of manual screening, and low recognition accuracy. It achieves automatic extraction of pickup codes, reduces maintenance costs, and improves the efficiency and accuracy of end-point pickup recognition. Attached Figure Description

[0015] Figure 1 This is a first flowchart of the logistics terminal pickup code recognition method provided in an embodiment of the present invention; Figure 2 This is a second flowchart of the logistics terminal pickup code recognition method provided in an embodiment of the present invention; Figure 3 This is a third flowchart of the logistics terminal pickup code recognition method provided in the embodiments of the present invention; Figure 4 This is a fourth flowchart of the logistics terminal pickup code recognition method provided in the embodiments of the present invention; Figure 5 The fifth flowchart of the logistics terminal pickup code recognition method provided in the embodiments of the present invention; Figure 6 The sixth flowchart of the logistics terminal pickup code recognition method provided in the embodiments of the present invention; Figure 7 The seventh flowchart of the logistics terminal pickup code recognition method provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of the structure of the logistics terminal pickup code recognition device provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of the logistics terminal pickup code recognition device provided in an embodiment of the present invention. Detailed Implementation

[0016] This invention provides a method, apparatus, device, and storage medium for identifying pickup codes at the end of a logistics process. 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 includes 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.

[0017] 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 terminal pickup code recognition method in this invention includes: 101. Acquire a mobile device image, and preprocess the mobile device image to obtain a preprocessed image; In this embodiment, the mobile device image is the mobile phone screen image displaying logistics pickup information to the user. Image acquisition can be achieved through the station's built-in camera or the user's mobile phone's front-facing camera. After acquiring the mobile device image, multi-dimensional preprocessing operations are performed to eliminate various image defects caused by the shooting scene. First, an automatic exposure compensation algorithm dynamically adjusts the image exposure parameters to optimize the overall brightness and contrast distribution of the image, avoiding interference from abnormal lighting on subsequent recognition, resulting in a light-adjusted image. Based on this, a moiré filtering algorithm is used to convert the light-adjusted image into a frequency domain image, accurately locating and filtering out frequency interference noise generated by screen shooting, retaining complete and valid image information, and generating a valid image. To address image tilt and distortion caused by the shooting angle, a perspective distortion correction algorithm is used to perform pixel reconstruction correction on the valid image, repairing geometric distortion and obtaining a corrected image with regular imaging. Finally, a standardization algorithm is used to uniformly crop and scale the corrected image, fixing the image specifications to a standard resolution, completing the entire preprocessing process, and ultimately obtaining a preprocessed image with uniform image quality, no distortion, no noise, and uniform size. This preprocessing workflow is specifically optimized for the unique image defects of mobile phone screen shooting scenarios. It can effectively eliminate adverse factors such as light deviation, moiré noise, and perspective distortion, unify image input standards, and significantly improve the accuracy and stability of subsequent text detection and recognition, providing a high-quality image foundation for accurate recognition of pickup codes.

[0018] 102. Perform keyword localization processing on the preprocessed image to obtain multiple initial text regions; In this embodiment, keyword localization processing is a preliminary text region screening operation performed on the preprocessed image based on a pre-trained lightweight object detection model. This model consists of a feature extraction module, a keyword matching module, and a region selection module cascaded together. During processing, the feature extraction module first extracts deep features from the preprocessed image, generating a global feature image that characterizes the image's texture and semantic information. Then, the keyword matching module performs comprehensive feature matching retrieval on the global feature image based on preset logistics text keywords, filtering out image information that matches the logistics keyword features and outputting the corresponding matching results. Finally, the region selection module, combined with the matching results, accurately locates the image position of the target text, completing the initial delineation of the target region. The initial text region obtained through this processing is a local image region containing logistics keywords and highly correlated with pickup information, capable of accurately locating the core area where a pickup code is suspected to exist. This processing method utilizes a lightweight model to achieve targeted keyword retrieval, abandoning the traditional full-image traversal detection mode. It can effectively filter irrelevant backgrounds and redundant text interference in images, significantly reducing the computational scope of subsequent text detection and recognition, reducing algorithm computational power consumption, and improving the targeting and accuracy of logistics text region localization, laying a reliable regional foundation for subsequent multi-dimensional screening and accurate recognition work.

[0019] 103. Based on the preprocessed image, perform spatial correlation matching and filtering on each of the initial text regions to obtain multiple valid text regions; In this embodiment, spatial association matching and filtering is the core processing flow for refining and optimizing the initial text region. This flow relies on the collaborative efforts of an image segmentation model and a text localization algorithm to achieve multi-dimensional feature matching and filtering. First, the pre-processed image is input into a pre-trained image segmentation model to complete the semantic segmentation of UI components within the image, obtaining the contour coordinate information of various UI functional areas. Simultaneously, a text localization algorithm is used to perform comprehensive text line detection on the initial text region, extracting all regions with text features and generating several candidate text line regions. Based on this, a multi-feature spatial association analysis algorithm is used, combined with the contour coordinates of the UI functional areas, to perform spatial matching and quantitative scoring on all candidate text line regions, generating a confidence score for each region based on the degree of spatial association. A secondary filtering is performed by introducing a preset confidence threshold, retaining target text regions that meet the score criteria, ultimately forming valid text regions. Valid text regions are high-quality text regions that, after spatial association matching and confidence filtering, fit the logistics information display area and have the potential to contain a pickup code. This processing method combines image semantic segmentation with text spatial feature association, abandoning the limitations of single text detection, effectively eliminating scattered and irrelevant interference text areas in the screen interface, accurately focusing on the core logistics information area, greatly improving the accuracy of text area screening, reducing the subsequent recognition workload of invalid areas, and avoiding interference caused by irrelevant text to the recognition of pickup codes.

[0020] 104. Perform quality screening on all the valid text regions to obtain the set of codes to be judged; In this embodiment, the quality screening process takes the selected effective text regions as the processing objects. First, based on the coordinate information of the effective text regions, a local cropping operation is performed on the preprocessed image to separate candidate region images without redundant background. Then, all candidate region images are sequentially fed into a pre-trained character recognition model for character extraction. The model's built-in feature extraction and sequence decoding logic parses the characters in the image, and simultaneously outputs the original text sequence and comprehensive text confidence score corresponding to each segment of content. Subsequently, a pre-set text confidence score threshold is retrieved, and the comprehensive text confidence score corresponding to each original text sequence is compared with the threshold one by one. Only texts whose comprehensive text confidence score meets the threshold standard are retained as target text sequences. Finally, all compliant target text sequences are aggregated to complete the structured integration, thereby generating a set of codes to be judged. The set of codes to be judged is a coded dataset composed of the selected texts that meet the confidence score after image cropping, character recognition, and confidence score filtering. The entire processing flow first removes irrelevant backgrounds by cropping images to reduce recognition noise, then relies on a lightweight text recognition model to complete character parsing, and sets up a confidence threshold mechanism to filter out low-quality text caused by ambiguity and misidentification. This can significantly reduce the amount of invalid data input in the subsequent encoding and discrimination stages, reduce the inference computing power occupied by invalid text, and at the same time reduce recognition errors caused by image imaging defects from the source, ensuring that the data flowing into the encoding and discrimination stages has sufficient credibility, and improving the running efficiency and discrimination accuracy of the backend pickup code classification and recognition.

[0021] 105. Input the set of codes to be determined into the pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code.

[0022] In this embodiment, the pickup code recognition model adopts a hierarchical architecture with a cascading prompt word loading module, an auxiliary reasoning module, and a code classification and judgment module. The basic model is built on a lightweight large language model with 4-bit quantization, adapting to the local deployment and high-speed inference requirements of edge devices. The overall training and adaptation process of the model revolves around the specific scenario of logistics pickup code recognition. First, a standardized structured prompt word system is established and solidified, clarifying the model's functional positioning as a dedicated logistics recognition expert. The model input is defined as the set of codes to be judged obtained through preliminary screening. The core task of the model is to identify genuine and valid pickup codes and eliminate various interference numbers. The model output data format is unified, and the four- to eight-character alphanumeric combination feature constraints specific to pickup codes are entered to provide a basic rule basis for model judgment. After the rule system is completed, the inference environment of the locally deployed quantized large language model is initialized and parameters are configured. The structured prompt words and preset logistics text sample data are synchronously imported into the model. During the training phase, a few-shot auxiliary learning mechanism is introduced. By inputting multiple sets of pre-labeled positive pickup codes and negative interference text samples such as express tracking numbers, order numbers, and verification codes, the model can fully learn the semantic and format feature differences between valid pickup codes and various similar interference texts. This continuously strengthens the model's ability to classify and distinguish logistics codes in a refined manner. After completing scenario adaptation and fine-tuning, the model's inference weight parameters are solidified.

[0023] In the pickup code recognition processing stage, the prompt word loading module first completes the standardized prompt word binding and fusion of the code set to be judged, generating a text set to be reasoned with fusion rule constraints. The auxiliary reasoning module, relying on the semantic features learned during training and sample learning experience, performs semantic parsing and feature matching on each text set to be reasoned, extracting the deep semantic features of all codes and aggregating them to form a set of code semantic features. Finally, the code classification and judgment module completes the code type differentiation and result verification and filtering, accurately identifying the unique valid pickup code and generating the corresponding confidence parameters and information source channel. At the same time, it executes the intelligent result judgment logic. If no valid pickup code is recognized, the terminal feedback mechanism is automatically triggered, outputting voice and vibration secondary focus commands. If a valid pickup code is recognized, semantic fine-tuning and filtering are completed, and the valid pickup code is output. This training method abandons the traditional fixed regular coding constraint mode, relying on few-sample learning to achieve autonomous feature iteration and adaptation of the model. It does not require frequent manual updates of rule parameters, effectively reducing system maintenance costs. The hierarchical and modular reasoning process combined with the semantic verification mechanism significantly improves the accuracy and robustness of pickup code recognition in complex scenarios.

[0024] In this embodiment, after valid pickup code recognition is completed, a structured output and business closed-loop process will be executed. First, the valid pickup code is structurally encapsulated, and key parameters such as pickup code content, information source, and recognition confidence level are uniformly collected to generate a business data format that conforms to the station's backend standards. Then, the encapsulated structured data is encrypted using an HTTPS encryption transmission algorithm and uploaded via a dedicated encrypted data channel to ensure data security and information integrity throughout the transmission process. After receiving the encrypted data packet, the station's backend management system decrypts and parses the content sequentially, calls the built-in package matching algorithm, and uses the pickup code as a unique retrieval identifier to traverse the inventory database to achieve accurate binding and matching of the corresponding package information and pickup code. After matching is completed, the terminal guidance algorithm is activated to synchronously link the station's supporting hardware devices, triggering LED highlighting prompts and voice broadcast programs to complete automated pickup guidance for pickup users. The system synchronously runs a local log storage algorithm, integrating the original images, candidate texts at each stage, final recognition results, operation timestamps, terminal device identifiers, and other full-link data generated during this recognition process. The data is written to the local cache according to a fixed storage cycle of seven days, building a complete and traceable recognition log system to support subsequent verification of abnormal recognition issues and traceability of business data. The entire recognition process is further enhanced by a federated learning-driven model continuous optimization branch. Gradient acquisition algorithms are deployed at each station terminal, extracting only the gradient update parameters of the local lightweight object detection model, lightweight text recognition model, and large language model LoRA adapter according to a preset cycle. No user's original images or plaintext data are collected or uploaded throughout the process; only the gradient parameters are aggregated and uploaded to the central server. The central server receives gradient data from multiple terminals using a gradient aggregation algorithm, performs weighted fusion, mean aggregation, and parameter integration on all gradients, and generates globally unified gradient update parameters. Then, the global model update algorithm is used to iteratively optimize the global weights of the three types of models. After the model update is completed, a lightweight incremental update package containing only the difference parameters is generated on a monthly cycle and uniformly distributed to each station terminal by the server. The terminal automatically loads the incremental package to complete the local model iteration, allowing the entire recognition system to autonomously adapt to various new application software interface styles and new pickup code encoding formats, forming a complete closed loop from image acquisition, intelligent recognition, business linkage to autonomous model iteration.

[0025] Please see Figure 2 In this embodiment of the invention, the preprocessing of the mobile device image to obtain a preprocessed image includes: 201. An automatic exposure compensation algorithm is used to adjust the exposure parameters of the mobile device image to obtain a light-adjusted image; 202. The light adjustment image is converted into a frequency domain image using a moiré filtering algorithm, the frequency interference signal of the frequency domain image is located, and the noise signal of the frequency interference signal is filtered out to obtain an effective image; 203. The effective image is reconstructed and corrected using a perspective distortion correction algorithm to obtain a corrected image; 204. The corrected image is subjected to size normalization processing using a normalization algorithm to obtain the preprocessed image.

[0026] In this embodiment, the mobile device image is the mobile phone screen image displaying logistics pickup information to the user. Image acquisition can be achieved through the station's built-in camera or the user's mobile phone's front-facing camera. After image acquisition, multiple preprocessing operations are performed sequentially to eliminate imaging defects caused by the shooting scene. The automatic exposure compensation algorithm can be implemented using an adaptive brightness mean equalization algorithm. During the exposure parameter adjustment process, the pixel distribution characteristics of the overall brightness and contrast of the original image are collected in real time. Based on the global pixel grayscale statistics, the image exposure gain and brightness parameters are dynamically corrected to balance the overall light and shadow performance of the image, unify the lighting standards of the device's image acquisition, and thoroughly improve the problems of overly bright, overly dark, and uneven local lighting, resulting in a uniformly high-quality, light-adjusted image. The moiré filtering algorithm can be implemented using an interference suppression algorithm based on frequency domain notch filtering. The frequency domain transformation process converts the spatial domain light adjustment image into a frequency domain image through Fourier transform, completing the mapping conversion from image spatial information to frequency information. Among them, the frequency interference signal specifically refers to the fixed frequency band periodic noise signal generated by the pixel arrangement and refresh flicker shooting of the mobile phone screen, which is also the core interference source forming screen moiré. The noise filtering process accurately locates the corresponding interference frequency band in the frequency domain, and uses a notch filter to target and suppress and filter the target frequency band, completely preserving the effective frequency domain feature information of the image, removing screen shooting-specific noise, and obtaining a clean and interference-free effective image. The perspective distortion correction algorithm can be implemented using a geometric distortion correction algorithm based on Hough line detection. The pixel reconstruction correction process relies on the edge detection operator to traverse and scan the effective image, accurately extract the four edge line features of the mobile phone screen, and construct a pixel distortion mapping matrix based on the inherent rectangular geometric constraints of the screen. According to the mapping relationship, the positions of tilted, stretched, and deformed distorted pixels are reconstructed and corrected to repair the geometric distortion problem caused by the shooting angle deviation, resulting in a corrected image with a regular shape and normal proportions. The standardization algorithm can be implemented using bilinear interpolation scaling and center cropping. The size standardization process adaptively crops and scales the corrected image proportionally, uniformly fixing all images to a fixed resolution of 640×480. This eliminates image size differences caused by different shooting distances and devices, completing the entire preprocessing workflow and outputting a standardized preprocessed image. The entire preprocessing workflow is layered and specifically optimized for various imaging defects specific to mobile phone screen shooting scenarios. Through multi-algorithm collaboration, it completes lighting correction, noise filtering, distortion repair, and size normalization, effectively avoiding the negative impact of complex shooting environments on image quality. It unifies the image input standard for subsequent algorithm models, significantly improving the accuracy and stability of text detection, feature matching, and character recognition, reducing recognition errors caused by invalid noise and image distortion, and providing a high-quality, highly consistent image data foundation for subsequent pickup code positioning and recognition.

[0027] Please see Figure 3In this embodiment of the invention, the keyword localization processing of the preprocessed image to obtain multiple initial text regions includes: 301. Call the pre-trained lightweight object detection model, which includes a feature extraction module, a keyword matching module, and a region bounding box module connected in sequence; 302. Based on the feature extraction module, perform feature extraction processing on the preprocessed image to obtain a global feature image; 303. Based on the keyword matching module, perform feature matching retrieval processing on the global feature image according to the preset logistics text keywords to obtain the keyword feature matching result; 304. Based on the region selection module, determine the initial text region according to the keyword feature matching result.

[0028] In this embodiment, the lightweight object detection model uses the YOLOv2-nano lightweight detection network. Its overall architecture consists of a cascaded feature extraction module, a keyword matching module, and a region bounding box module, resulting in a streamlined structure with a small number of parameters, suitable for real-time detection and inference requirements of edge devices. During model training, a training dataset is constructed based on a large number of mobile logistics interface screen image samples. High-frequency text keywords in logistics scenarios are used as supervisory labels. Through iterative training, the network weight parameters are continuously updated and converged, continuously learning the texture features, edge features, and semantic association features of logistics text. Finally, dedicated pre-trained weights adapted for mobile screen logistics text localization are solidified, possessing a unique perception capability for text targets in logistics scenarios. In the keyword localization process, the feature extraction module first performs multi-layer convolutional feature extraction and feature fusion on the input preprocessed image, progressively mining shallow texture, edge details, and deep semantic features of the image. This aggregates and generates a global feature image covering the entire image, providing complete feature basis for subsequent keyword matching. The preset logistics text keywords are set based on the general scenario attributes and industry-standard naming rules of last-mile delivery services. They mainly include core business terms that can refer to delivery information, such as pickup code, collection code, and self-pickup code, covering the logistics display text of most e-commerce platforms, mini-programs, and SMS notifications. During feature matching and retrieval, the keyword matching module relies on prior knowledge of keyword features learned during model training to perform a region-by-region feature traversal and similarity comparison of the global feature image. It accurately matches target feature regions in the image that match the preset logistics keyword features, outputting keyword feature matching results that characterize the target location and feature similarity. After receiving the keyword feature matching results, the region selection module, combined with the model's preset detection threshold and boundary regression mechanism, performs coordinate correction and bounding box fitting on target regions that meet the matching criteria, locking down the local image range where the logistics keywords are located. Finally, multiple initial text regions are delineated, which are text regions highly correlated with delivery information. This hierarchical processing mode relies on a lightweight dedicated detection model to achieve targeted logistics text retrieval, getting rid of the redundant computational drawbacks of traditional full-image traversal detection. It can effectively filter out background elements and irrelevant text interference, significantly compress the computational range of subsequent text detection and recognition, reduce the consumption of terminal computing resources, and improve the accuracy and stability of logistics keyword positioning through targeted scenario training, providing an accurate and reliable pre-regional foundation for subsequent spatial association filtering and text recognition.

[0029] Please see Figure 4 In this embodiment of the invention, based on the preprocessed image, spatial correlation matching and filtering are performed on each of the initial text regions to obtain multiple valid text regions, including: 401. Input the preprocessed image into a pre-trained image segmentation model to perform semantic segmentation of UI components, and obtain the contour coordinates of the UI functional area; 402. A text positioning algorithm is used to perform text line positioning processing on each of the initial text regions to obtain multiple candidate text line regions; 403. Using a multi-feature spatial association analysis algorithm, multiple candidate text line regions are screened and scored based on the outline coordinates of the UI function area to obtain multiple target text regions and a confidence score corresponding to each target text region. 404. Obtain a preset confidence threshold, compare each confidence score with the confidence threshold, and take the target text region with the confidence score higher than the confidence threshold as the effective text region.

[0030] In this embodiment, the image segmentation model adopts a U-Net symmetrical encoding and decoding architecture. The encoding layer performs progressive compression and extraction of shallow texture and deep semantic features, while the decoding layer achieves feature upsampling and detail restoration. A skip connection structure compensates for edge details lost during downsampling, accurately adapting to the segmentation scenario of small UI components on mobile screens. The model training process uses a large number of diverse mobile logistics interface images as training samples, labeling various information cards, operation buttons, and other UI components with pixel-level semantic tags. Through backpropagation, network parameters are iteratively optimized, continuously converging pixel classification errors and contour fitting errors, learning the morphological characteristics and distribution patterns of screen UI components in logistics scenarios, and finally obtaining pre-trained model weights adapted for mobile screen interface segmentation. The UI component semantic segmentation process involves inputting the pre-processed image into the trained image segmentation model. Through progressive feature inference and pixel-level classification, the model accurately identifies the distribution range of various functional UI components within the image, completing the semantic division of interface elements. Finally, it outputs the contour coordinates of the UI functional area, representing the boundary position and morphological range of each UI component. These coordinates are pixel position parameters that accurately define the effective display areas of information cards, function buttons, etc. The text localization algorithm employs the DBNet++ text detection network, which is OCR-free. The text line localization process relies on this network to perform a global feature traversal of the initial text region. An adaptive threshold distinguishes text pixels from background pixels, extracting all valid text lines with text texture and edge features. This fully captures all text distribution information within the initial text region, generating candidate text line regions that are all local image regions within the initial text region that possess text features and are suspected of containing character information. The multi-feature spatial association analysis algorithm uses a spatial intersection-union-comparison (IUCN) combined with a neighborhood distance weighted analysis strategy. The screening and scoring process uses the UI function area outline coordinates as the effective spatial reference, calculating the spatial overlap and neighborhood association distance between each candidate text line region and the UI function area. Combined with prior features of the logistics text's attachment to information cards and the distribution of function buttons, each candidate text line region is quantitatively scored. Finally, multiple target text regions fitting the logistics display scenario are obtained, along with corresponding confidence scores representing the effectiveness of each region. Higher scores indicate a greater probability that the corresponding text region belongs to the valid information area of ​​the pickup code. The confidence threshold is set comprehensively based on the model detection accuracy and interference text removal requirements of a large number of real-world test scenarios, balancing the retention rate of effective text with the filtering effect of interference text. In general engineering applications, it can be set to 0.6. After threshold comparison and screening, the target text regions with retention confidence scores higher than the threshold are the effective text regions. These regions are high-quality text candidate regions that have been filtered by semantic segmentation constraints and spatial features, eliminating scattered interference text and invalid background text on the screen, and highly fitting the logistics pickup code display scenario.This joint screening mechanism integrates deep semantic segmentation and high-precision text detection technologies. It corrects the blindness of pure text detection through UI space constraints, and completes accurate noise removal by relying on quantitative scoring and threshold screening. It completely solves the problem that traditional text detection is prone to misdetecting redundant text and decorative characters on the interface. It effectively compresses the computational range of subsequent character recognition, reduces the consumption of invalid recognition computing power, and avoids the impact of various interface interference information on the recognition of pickup codes at the regional level. It significantly improves the accuracy and scenario adaptability of logistics text screening, and provides pure and reliable pre-regional data for subsequent high-quality text recognition and encoding judgment.

[0031] Please see Figure 5 In this embodiment of the invention, the step of performing quality screening on all the valid text regions to obtain a set of codes to be judged includes: 501. Based on all the effective text regions, the preprocessed image is cropped to obtain multiple candidate region images; 502. Input each candidate region image into a pre-trained character recognition model for character extraction processing to obtain multiple original text sequences and a comprehensive text confidence score corresponding to each original text sequence; 503. Obtain a preset text confidence threshold, compare each of the comprehensive text confidence scores with the text confidence threshold, and take the original text sequence with the comprehensive text confidence score higher than or equal to the text confidence threshold as the target text sequence; 504. Integrate all the target text sequences to obtain the set of codes to be determined.

[0032] In this embodiment, the cropping process relies on the pixel coordinate information of the effective text region to perform local precise cropping of the preprocessed image. Based on the boundary coordinates of the high-quality text region obtained in the previous screening, the complete preprocessed image is cropped at fixed points and locally cropped, and invalid background and irrelevant interface elements outside the region are removed. Only the local image containing the suspected pickup information is retained, and finally, a candidate region image with a simple structure and very little interference information is obtained. This type of image is a local image material that focuses on the text target and removes redundant background, which can provide a clean input carrier for subsequent text recognition.

[0033] In this embodiment, the text recognition model used is a lightweight PP-OCRv5-mobile network on the edge. The overall architecture consists of a text feature extraction module, a sequence decoding module, and a confidence calculation module connected in series. The model training process relies on a large number of mobile phone screen logistics information image samples to complete scene adaptation training, covering pickup code text samples with different lighting conditions, different interface styles, and different resolutions. By continuously iterating and optimizing the convolution weights and decoding parameters, the model learns the texture features, stroke features, and arrangement rules of alphanumeric combination text in logistics scenarios, and completes model convergence and parameter solidification, enabling it to have high-precision recognition capabilities for short coded logistics text. In the character extraction process, each candidate region image is first fed into the text feature extraction module. Through multi-layer convolution operations, the shallow edge features and deep semantic features of the image text are aggregated and extracted to generate a text feature image with regular dimensions and significant features. Then, the sequence decoding module performs temporal feature parsing and character inference on the feature image, and outputs the character content recognized in the image one by one to form multiple complete original text sequences. Finally, the confidence calculation module calculates the confidence level of single character recognition, and combines the single character confidence level with weighted fusion to obtain the comprehensive text confidence level of the entire text. The original text sequence is all the text content parsed by the model from the candidate region image, and the comprehensive text confidence level is used to quantify the accuracy of the corresponding text recognition result.

[0034] In this embodiment, the text confidence threshold is comprehensively calibrated based on the error tolerance requirements of logistics scenario recognition, the characteristics of fuzzy text recognition, and a large amount of offline test data. While balancing the effective text retention rate and the erroneous text filtering rate, it can be fixed at 0.6 in engineering applications, effectively filtering low-confidence recognition results caused by image blur, image distortion, and feature incompleteness. All original text sequences with a comprehensive text confidence score not lower than the preset threshold are determined as target text sequences. Target text sequences are valid text content that has passed confidence verification, meets recognition accuracy standards, and possesses potential value for pickup codes. By structurally summarizing and integrating all filtered target text sequences, a set of codes to be judged can be generated. This set is a purified pickup code dataset after removing misidentified content, providing high-quality input material for subsequent large-scale model semantic discrimination. This entire processing flow, through a hierarchical processing logic of first trimming and denoising, then accurate recognition, and finally confidence filtering, significantly reduces the interference of background noise on character recognition, effectively filters out invalid text data generated by fuzzy recognition and misrecognition, significantly reduces the data processing pressure and computing power consumption of subsequent semantic reasoning stages, improves the overall accuracy and stability of the pickup code recognition system from the data source, and ensures the reliability and accuracy of subsequent encoding and discrimination results.

[0035] Please see Figure 6In this embodiment of the invention, the character recognition model includes a character feature extraction module, a sequence decoding module, and a confidence calculation module connected in sequence; the step of inputting each candidate region image into the pre-trained character recognition model for character extraction processing to obtain multiple original text sequences and a comprehensive text confidence score corresponding to each original text sequence includes: 601. Input each candidate region image into the text feature extraction module for image text feature extraction processing to obtain multiple text feature images; 602. Input each of the text feature images into the sequence decoding module for character sequence decoding processing to obtain multiple original text sequences; 603. Based on the confidence calculation module, calculate the confidence of all characters in each of the original text sequences to obtain the comprehensive text confidence corresponding to each of the original text sequences.

[0036] In this embodiment, the text recognition model adopts a three-stage serial architecture, consisting of a text feature extraction module, a sequence decoding module, and a confidence calculation module connected sequentially. It leverages a lightweight PP-OCRv5-mobile network structure to achieve efficient character recognition on the device side. During image text feature extraction, the module performs multi-layer convolutional feature sampling and fusion operations on the input candidate region image, extracting shallow detail features such as text edge contours, stroke textures, and grayscale differences layer by layer. Simultaneously, it mines deep semantic features such as text layout patterns and character structure relationships, suppressing redundant background features. Finally, it outputs a text feature image with uniform dimensions and highly focused text features, providing accurate and effective feature input for subsequent character decoding. The character sequence decoding process relies on a temporal decoding mechanism to perform feature inference and character fitting on the text feature image. Combining the prior alphanumeric encoding features learned during model training, it analyzes and classifies local feature units in the feature map segment by segment, completing the temporal recombination of characters according to the text layout order, continuously outputting complete and continuous character content, and finally generating multiple original text sequences corresponding to each candidate region image. The confidence calculation process first performs independent confidence calculations on each recognized character within the original text sequence. Based on character feature matching degree, feature completeness, and the degree of interference from similar features, a single-character confidence value is output. Then, a weighted fusion algorithm integrates the confidence values ​​of all single characters in the entire text sequence, eliminating abnormal deviation values, and comprehensively calculating a comprehensive text confidence value that can represent the overall accuracy of single-text recognition. This modular, step-by-step processing method breaks down image feature extraction, character sequence restoration, and recognition confidence quantification. It can accurately adapt to the short code recognition scenario of logistics pickup codes on mobile phone screens. The refined feature extraction method effectively resists feature loss problems caused by image blur and slight distortion. The temporal decoding mechanism ensures that the character output order is consistent with the actual text layout. The multi-dimensional confidence fusion calculation method can objectively quantify the quality of the recognition results, providing accurate data for subsequent text quality screening. From the model recognition level, it ensures the accuracy and stability of logistics text recognition, effectively avoiding recognition deviations caused by single-character misrecognition and local feature interference, and significantly improving the overall accuracy of subsequent pickup code screening and judgment.

[0037] Please see Figure 7 In this embodiment of the invention, the pickup code recognition model includes a prompt word loading module, an auxiliary reasoning module, and an encoding classification and determination module connected in sequence; the step of inputting the set of codes to be determined into the pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code includes: 701. Based on the prompt word loading module, the set of codes to be determined is subjected to standardized prompt word binding processing to obtain the set of texts to be reasoned; 702. Based on the auxiliary reasoning module, perform semantic reasoning processing on the text set to be reasoned to obtain a set of encoded semantic features; 703. Based on the encoding classification and determination module, the encoding semantic feature set is processed to distinguish the encoding type, and the valid pickup code is obtained.

[0038] In this embodiment, the pickup code recognition model consists of a cascaded prompt word loading module, an auxiliary reasoning module, and an encoding classification and judgment module. The basic network adopts a lightweight large language model with 4-bit quantization compression, which can adapt to the local deployment and low-latency inference requirements of terminal devices. The standardized prompt word binding process involves pre-building a standardized prompt word system adapted to the logistics recognition scenario, clarifying the model's specific working position in logistics pickup code recognition, defining the model input format and core discrimination tasks, solidifying the inherent feature constraints and unified output specifications of four to eight-digit alphanumeric pickup codes, and fusing and binding the formed structured prompt words with the input set of codes to be judged. This allows the discrete candidate code data to carry standardized business constraints and discrimination rules. The final generated text set to be reasoned is a reasonable text dataset that integrates manual prior rules, scenario constraints, and candidate code data, providing rule support for subsequent semantic discrimination. Semantic reasoning processing relies on a few-shot learning mechanism. The model preloads multiple sets of fully labeled positive examples of pickup codes and negative examples of interfering text such as express tracking numbers, order numbers, and verification codes. Through sample comparison, it learns the format features and semantic differences of different logistics codes. The auxiliary reasoning module performs deep semantic parsing, feature matching, and content identification for each candidate code in the text set to be reasoned based on the learned feature rules. It fully explores the character structure features, length features, and business semantic features of each set of codes, and integrates the multi-dimensional feature information of all codes to form a set of code semantic features. This set is a structured feature dataset that covers the feature similarity, compliance, and scenario matching of all candidate codes. The encoding type differentiation process relies on the semantic feature set of the encoding to complete fine-grained classification and verification. Combining the inherent character constraints of the pickup code with the feature differences learned from samples, it performs type identification and compliance verification on all candidate codes, accurately distinguishing valid pickup codes from interference codes of various similar formats. At the same time, it completes semantic fine-tuning and credibility judgment of the recognition results. The system executes the result judgment logic synchronously. When no compliant pickup code is found, it automatically triggers terminal voice and vibration feedback commands to guide the user to refocus and capture images. If a compliant code is successfully selected, the standardized recognition result is solidified, and finally, an accurate and valid pickup code is output. This modular reasoning process deeply integrates rule constraints, small-sample semantic learning, and intelligent result verification, breaking away from the rigid limitations of traditional regular expression matching. It can autonomously adapt to various new encoding formats and interface display styles without the need for manual iterative updates to matching rules, significantly reducing system operation and maintenance costs. At the same time, through multi-layer feature reasoning and accurate classification verification, it greatly improves the accuracy and robustness of pickup code recognition in complex and mixed text scenarios. The perfect terminal feedback mechanism also effectively solves the problem of missed recognition, ensuring the stable closed-loop operation of the overall recognition business.

[0039] The above describes the method for identifying the pickup code at the end of the logistics chain in the embodiments of the present invention. The following describes the device for identifying the pickup code at the end of the logistics chain in the embodiments of the present invention. Please refer to [link / reference]. Figure 8 One embodiment of the logistics terminal pickup code identification device of the present invention includes: Preprocessing module 801: used to acquire mobile device images, preprocess the mobile device images, and obtain preprocessed images; Keyword localization module 802: used to perform keyword localization processing on the preprocessed image to obtain multiple initial text regions; Association matching module 803: is used to perform spatial association matching filtering on each of the initial text regions based on the preprocessed image to obtain multiple valid text regions; Quality screening module 804: used to perform quality screening on all the valid text regions to obtain a set of codes to be judged; Pickup code recognition module 805: used to input the set of codes to be judged into the pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code.

[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 terminal pickup code identification device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The logistics terminal pickup code identification device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0042] Figure 9 This is a schematic diagram of the structure of a logistics last-mile pickup code identification device 900 provided in an embodiment of the present invention. The logistics last-mile pickup code identification 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) 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 last-mile pickup code identification 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 last-mile pickup code identification device 900 to implement the steps of the logistics last-mile pickup code identification method provided in the above-described method embodiments.

[0043] The last-mile delivery code 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 terminal pickup code identification device does not constitute a limitation on the logistics terminal pickup code identification 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 end-of-line pickup code identification 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 identifying pickup codes at the last mile of logistics, characterized in that, include: Acquire a mobile device image, and preprocess the mobile device image to obtain a preprocessed image; The preprocessed image is subjected to keyword localization processing to obtain multiple initial text regions; Based on the preprocessed image, spatial correlation matching and filtering are performed on each of the initial text regions to obtain multiple valid text regions; All the valid text regions are subjected to quality screening to obtain a set of codes to be judged. The set of codes to be determined is input into a pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code.

2. The method for identifying pickup codes at the end of logistics terminals according to claim 1, characterized in that, The preprocessing of the mobile device image to obtain a preprocessed image includes: An automatic exposure compensation algorithm is used to adjust the exposure parameters of the mobile device image to obtain a light-adjusted image. The light-adjusted image is converted into a frequency domain image using a moiré filtering algorithm. The frequency interference signal in the frequency domain image is located, and noise signal filtering is performed on the frequency interference signal to obtain an effective image. A perspective distortion correction algorithm is used to perform pixel reconstruction correction on the effective image to obtain a corrected image. The corrected image is sized and normalized using a normalization algorithm to obtain the preprocessed image.

3. The method for identifying pickup codes at the end of logistics terminals according to claim 1, characterized in that, The preprocessed image is subjected to keyword localization processing to obtain multiple initial text regions, including: A pre-trained lightweight object detection model is invoked, which includes a feature extraction module, a keyword matching module, and a region bounding box module connected in sequence. Based on the feature extraction module, feature extraction processing is performed on the preprocessed image to obtain a global feature image; Based on the keyword matching module, feature matching retrieval processing is performed on the global feature image according to preset logistics text keywords to obtain keyword feature matching results; Based on the region selection module, the initial text region is determined according to the keyword feature matching result.

4. The method for identifying pickup codes at the end of logistics terminals according to claim 1, characterized in that, Based on the preprocessed image, spatial correlation matching and filtering are performed on each of the initial text regions to obtain multiple valid text regions, including: The preprocessed image is input into a pre-trained image segmentation model to perform semantic segmentation of UI components, thereby obtaining the contour coordinates of the UI functional area. A text positioning algorithm is used to perform text line positioning processing on each of the initial text regions to obtain multiple candidate text line regions; A multi-feature spatial correlation analysis algorithm is used to filter and score multiple candidate text line regions based on the outline coordinates of the UI function area, thereby obtaining multiple target text regions and a confidence score corresponding to each target text region. A preset confidence threshold is obtained, and each confidence score is compared with the confidence threshold. The target text region with a confidence score higher than the confidence threshold is taken as the valid text region.

5. The method for identifying pickup codes at the end of logistics terminals according to claim 1, characterized in that, The quality screening process for all the valid text regions yields a set of codes to be judged, including: The preprocessed image is cropped based on all the valid text regions to obtain multiple candidate region images; Each candidate region image is input into a pre-trained character recognition model for character extraction processing to obtain multiple original text sequences and a comprehensive text confidence score corresponding to each original text sequence. Obtain a preset text confidence threshold, compare each comprehensive text confidence score with the text confidence threshold, and take the original text sequence with a comprehensive text confidence score higher than or equal to the text confidence threshold as the target text sequence; Integrate all the target text sequences to obtain the set of codes to be determined.

6. The method for identifying pickup codes at the end of logistics terminals according to claim 5, characterized in that, The text recognition model includes a text feature extraction module, a sequence decoding module, and a confidence calculation module connected in sequence; the step of inputting each candidate region image into the pre-trained text recognition model for character extraction processing to obtain multiple original text sequences and a comprehensive text confidence score corresponding to each original text sequence includes: Each candidate region image is input into the text feature extraction module for image text feature extraction processing to obtain multiple text feature images; Each of the text feature images is input into the sequence decoding module for character sequence decoding processing to obtain multiple original text sequences; The confidence score is calculated based on the confidence score of each character in each original text sequence by the confidence score calculation module, and the comprehensive text confidence score corresponding to each original text sequence is obtained.

7. The method for identifying pickup codes at the end of logistics terminals according to claim 1, characterized in that, The pickup code recognition model includes a prompt word loading module, an auxiliary reasoning module, and a code classification and determination module connected in sequence. The step of inputting the set of codes to be determined into a pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code includes: Based on the prompt word loading module, the set of codes to be judged is standardized and bound with prompt words to obtain the set of texts to be reasoned. Based on the auxiliary reasoning module, semantic reasoning processing is performed on the text set to be reasoned to obtain a set of encoded semantic features; The coding classification and determination module performs coding type differentiation processing on the coding semantic feature set to obtain the valid pickup code.

8. A logistics terminal pickup code identification device, characterized in that, include: Preprocessing module: used to acquire mobile device images, preprocess the mobile device images, and obtain preprocessed images; Keyword localization module: used to perform keyword localization processing on the preprocessed image to obtain multiple initial text regions; The association matching module is used to perform spatial association matching and filtering on each of the initial text regions based on the preprocessed image to obtain multiple valid text regions. Quality screening module: used to perform quality screening on all the valid text regions to obtain a set of codes to be judged; Pickup code recognition module: used to input the set of codes to be judged into the pre-trained pickup code recognition model for pickup code recognition processing to obtain a valid pickup code.

9. A logistics last-mile pickup code identification device, characterized in that, The logistics terminal pickup code identification 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 terminal pickup code identification device to perform the various steps of the logistics terminal pickup code identification method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements each step of the logistics terminal pickup code identification method as described in any one of claims 1-7.

Images