CN-121982421-A - Visual identification system and method for goods sorting and transportation

CN121982421ACN 121982421 ACN121982421 ACN 121982421ACN-121982421-A

Abstract

The application relates to a visual identification system and a visual identification method for sorting and transporting goods, which relate to the technical field of logistics transportation control, wherein the visual identification system comprises an image acquisition module, a control module and a control module, wherein the image acquisition module acquires original object images of goods in transmission according to different visual angles; the intelligent logistics sorting system comprises an image processing module, a model construction module, a screening and identifying module, an updating and optimizing module, a multi-view image merging module, a multi-view sorting and distinguishing module and a depth convolution neural network, wherein the image processing module is used for cleaning and standardizing an original object image to obtain an object multi-view merging image, extracting logistics distribution positions, the model construction module is used for converging historical object images to be a training set and combining with a target detection algorithm for iterative training to generate a multi-view sorting and distinguishing model, the screening and identifying module is used for judging the object multi-view merging image according to the multi-view sorting and distinguishing model and combining with the logistics distribution positions to determine an object sorting conclusion, the updating and optimizing module is used for filling the object multi-view merging image into an object database according to the object sorting conclusion and optimizing the multi-view sorting and distinguishing model, and the sorting and distinguishing model is constructed through a multi-view image merging algorithm and the depth convolution neural network, so that the accuracy and the efficiency of intelligent logistics sorting are improved.

Inventors

GAO XIAOER
FU MINGDONG
ZHONG YANHUA
JIANG WEI
DAI CANLIN

Assignees

元宇智信息科技(昆山)有限公司

Dates

Publication Date: 20260505
Application Date: 20260205

Claims (10)

1. A visual identification system for sorting transportation of goods, comprising: the image acquisition module is used for acquiring original object images of goods in transmission according to different visual angles; the image processing module is used for cleaning and standardizing the original object image to obtain an object multi-view combined image and extract logistics distribution positions; The model construction module is used for aggregating the historical object images into a training set and carrying out iterative training by combining a target detection algorithm to generate a multi-view sorting discrimination model; The screening and identifying module is used for judging the multi-view joint images of the articles according to the multi-view sorting and judging model, and determining article sorting conclusion by combining the logistics distribution positions; And the updating and optimizing module is used for filling the multi-view joint images of the articles into the article database according to the article sorting conclusion and optimizing the multi-view sorting judgment model.
2. The visual identification system for sorting transportation of goods of claim 1, wherein the image processing module comprises: the image correction unit is used for performing geometric correction and gray correction on the original article image to obtain a corrected article image; the gain compensation unit is used for carrying out gain compensation and gray balance on the corrected article image to obtain a positive-color article image; the image denoising unit is used for filtering, denoising, interpolating and filling the positive color object image to obtain a noiseless object image; The feature dividing unit is used for clustering and region growing the noiseless article images to obtain article feature regions and determining article center coordinates; the edge segmentation unit is used for carrying out edge detection and example segmentation on the article characteristic region and extracting a single article characteristic image; the template matching unit is used for matching the single article characteristic images according to the logistics sheet template and screening logistics sheet article common views; The image association unit is used for carrying out optical flow association on the logistics single article common view and the single article characteristic image according to the article center coordinate and the acquisition time stamp to obtain an article multi-view joint image; and the information extraction unit is used for carrying out text detection on the common view of the logistics single objects and extracting logistics distribution positions.
3. The visual identification system for sorting transportation of goods of claim 1, wherein the model building module comprises: The image aggregation unit is used for performing multi-view aggregation on the historical object images and constructing a training set, a verification set and a test set according to a preset dividing proportion; the feature extraction unit is used for extracting features of the multi-view joint images of the articles in the training set according to the first extraction layer to obtain a multi-view global feature matrix; extracting features of the single views of the articles in the training set according to a second extraction layer to obtain a local view feature matrix; performing feature extraction on the common view of the logistics single objects in the training set according to a third extraction layer to obtain a logistics single location feature matrix; The iterative training unit is used for carrying out joint training on the multi-view global feature matrix and the local view feature matrix by combining the article view labels according to the first training layer to obtain a morphological view weight matrix; According to a second training layer, cross training is carried out on the multi-view global feature matrix and the logistics unit location feature matrix by combining a spatial attention mechanism, so as to obtain a view positioning weight matrix; Performing global refinement filtering on the morphological view weight matrix by utilizing a multi-view pseudo tag algorithm according to a first branch of a third training layer to obtain a global view weight matrix; the second branch performs structure joint operation on the view positioning weight matrix by using a decoupling semi-supervision algorithm to obtain a local positioning weight matrix; And the third branch performs hierarchical training on the global view weight matrix and the local positioning weight matrix by using 2 up-sampling blocks, a cross entropy loss function and a global attention mechanism to obtain an article comprehensive judgment matrix.
4. The visual recognition system for sorting transportation of cargo of claim 3, wherein the model building module further comprises: the self-correction verification unit is used for detecting and identifying the verification set according to the comprehensive article judgment matrix, determining an article view and an article sorting position, comparing the article view with an original article view label, and calculating accuracy and positioning accuracy; If the accuracy and the positioning accuracy do not reach the corresponding threshold values, carrying out weighting operation on the accuracy and the positioning accuracy, correcting the original learning rate to obtain a new learning rate, and carrying out retraining on the comprehensive judgment matrix by combining a target detection algorithm until the maximum iteration number is reached to obtain a standard-combination comprehensive judgment matrix; and the identification optimizing unit is used for detecting and identifying the test set according to the label combination comprehensive judgment matrix, outputting the article view and the logistics single view, comparing with the article view label, calculating and drawing a PR curve, correcting a cross entropy loss function according to the ratio point of the PR curve to obtain an optimal article judgment matrix, and generating a multi-view sorting judgment model.
5. The visual identification system for sorting transportation of goods of claim 1, wherein said screening identification module comprises: The identification matching unit is used for identifying the multi-view joint image of the article according to the multi-view sorting discrimination model to obtain an article view and a logistics list view; if the logistics single view is incomplete, extracting and dividing logistics single part bitmaps in all object views, and performing image stitching fusion to obtain logistics single images; if the logistics sheet image cannot be analyzed, matching the object view according to a preset object view sorting database to obtain a similar object form view, and extracting an object transfer position; If not, carrying out information interpretation on the logistics sheet image to determine a logistics sorting position; And if the logistics single view is complete, performing information interpretation on the current logistics single view, and determining the logistics sorting position.
6. The visual identification system for sorting transportation of goods of claim 5, wherein said screening identification module further comprises: The transfer verification unit is used for identifying the logistics bill information again when the article is transferred to the article transfer position or the logistics sorting position, reading the logistics distribution position and matching the transportation sorting position; if the article transfer position or the logistics sorting position is different from the conveying sorting position, continuing to convey the current article, and extracting a multi-view combined image of the article; And the conclusion construction unit is used for carrying out aggregation marking on the conveying sorting position and the actual multi-view of the articles to generate article sorting conclusions.
7. The visual identification system for sorting transportation of goods of claim 1, wherein the update optimization module comprises: The information deconstructing unit is used for analyzing the article sorting conclusion, extracting and conveying sorting positions and actual multi-view of the articles; the space-time association unit is used for carrying out association alignment on the multi-view joint images of the articles according to the view angle characteristics to obtain actual joint images of the articles and extract view level characteristics; the semantic coding unit is used for coding and packaging the conveying sorting positions according to sorting mouth identification and sorting results to generate article sorting codes; the information filling unit is used for carrying out similarity calculation on the view level characteristics according to an article database, screening similar article views, carrying out neighbor filling on the actual joint images of the articles, and establishing a space-time neighbor index; The quality enhancement unit is used for carrying out feature comparison and data enhancement on the similar object view and the object actual joint image according to the space-time neighbor index to obtain object specific enhancement features; And the increment optimization unit is used for carrying out graph aggregation on the article specific enhancement features and the article sorting codes to obtain multi-view fusion features, and carrying out feature iterative optimization on the multi-view sorting discrimination model by combining a triplet loss function to obtain an optimal sorting discrimination model.
8. A visual recognition method applied to the system of any one of claims 1 to 7, comprising: acquiring original object images of goods in transmission according to different visual angles; Cleaning and standardizing the original article image to obtain an article multi-view combined image, and extracting logistics distribution positions; Aggregating historical object images as a training set, and performing iterative training by combining a target detection algorithm to generate a multi-view sorting discrimination model; Judging the multi-view combined drawing of the article according to the multi-view sorting judgment model, and determining an article sorting conclusion by combining the logistics distribution position; and filling the multi-view joint images of the articles into an article database according to the article sorting conclusion, and optimizing the multi-view sorting judgment model to obtain an optimal sorting judgment model.
9. An electronic device, comprising: One or more processors; A memory for storing one or more programs; the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of claim 8.
10. A storage medium having stored therein at least one instruction, at least one program, a set of codes, or a set of instructions, the at least one instruction, the at least one program, the set of codes, or the set of instructions being loaded and executed by a processor to implement the visual identification method of claim 8.

Description

Visual identification system and method for goods sorting and transportation Technical Field The application relates to the technical field of logistics transportation control, in particular to a visual identification system and method for goods sorting transportation. Background In the era background of the vigorous development of digital economy, the technology of the Internet of things is deeply remodelling the logistics industry pattern. In a complete logistics link from a merchant to a customer, the package needs to undergo a complex process of "pick-up-transfer-sort-delivery", wherein the sorting links of the terminal transfer center and the delivery network are used as key nodes, and the efficiency directly affects the overall logistics aging. Current sorting techniques mainly take on two paradigms: ‌ manual sorting mode ‌, wherein the distribution range of each website is memorized by an operator, and the website codes are marked by manually identifying the face sheet addresses. The mode has the obvious disadvantages of high labor cost, low address information matching accuracy (subjective deviation exists due to manual identification), and difficulty in coping with the requirement of large-scale order processing. ‌ The intelligent sorting system ‌ trains the mapping relation between addresses and network points based on historical delivery data by constructing a machine learning model. The system realizes automatic sorting and prediction of the face sheet addresses by configuring keyword rules of province and city areas, streets and the like or utilizing a deep learning algorithm to mine distribution rules. The mode remarkably improves the treatment efficiency and becomes the main stream development direction of the industry. Although intelligent sorting systems have advantages in matching rate and processing speed, there is a deadly disadvantage of ‌ model iteration lag ‌. To ensure system coverage, the data acquisition cycle typically lasts for months. When business rules change (e.g., regional distribution point adjustments), the model undergoes a lengthy retraining process (typically over 30 days), resulting in a significant number of mis-sorts during the transition. This not only incurs additional mistakes handling costs, but also causes a significant reduction in sorting timeliness, ultimately translating into substantial impairment of the user experience. The prior patent discloses a training method and device of an object recognition model, and an object recognition method and system. The object recognition model training method comprises the steps of obtaining a training set, wherein the training set comprises a point cloud labeling data set of an object collected by a laser radar, the point cloud labeling data set is provided with a truth value boundary box, generating a top view according to the point cloud labeling data set, utilizing a feature extractor to extract a plurality of feature graphs with different resolutions from the top view, determining the size of an anchor frame and the position of the anchor frame on the feature graphs, generating anchor frames with different sizes including the size and the aspect ratio by taking each pixel of the feature graphs as a center, matching the anchor frame with the truth value boundary box on the feature graphs with different resolutions to determine the sample types of the anchor frame, and training the object recognition model based on the contribution of the anchor frames with different sample types to a loss function of a convolutional neural network. The prior art scheme has the following defects that 1, the traditional visual recognition model only recognizes faces and automobiles, the recognition degree is not high in the logistics field, and when irregular-shaped packages circulate on a sorting line, stacking and rolling can occur, so that a logistics sheet is blocked, and therefore, the sorting position of the packages cannot be determined by reading information of the logistics sheet in real time. Disclosure of Invention Aiming at the defects of the prior art, the application aims to optimize the matching speed and efficiency of a newly shot photo in a database in a six-view classification mode by classifying the articles in transportation into six views, output article transfer judgment by combining AI identification and analysis of article characteristics, and further determine the actual sorting position of the articles by triple screening compared with similar articles in the database so as to realize free model training of a logistics article database and improve the precision and efficiency of subsequent logistics identification. The method is realized by adopting the following technical scheme: In a first aspect, the present application provides a visual identification system for sorting transportation of goods, comprising: the image acquisition module is used for acquiring original object images of goods in transmission ac