CN-121982444-A - Florence 2-based intelligent labeling method and system for process design image

CN121982444ACN 121982444 ACN121982444 ACN 121982444ACN-121982444-A

Abstract

The invention provides a process design image intelligent labeling method and a process design image intelligent labeling system based on Florence2 in the technical field of computer vision, wherein the method comprises the following steps of S1, obtaining a large number of historical process design images for screening, labeling and cross checking; the method comprises the steps of S2, preprocessing each historical process design image to construct a process image set, S3, creating a process image labeling model based on a Florence2 network, configuring pre-training weights and input layer parameters of the process image labeling model, S4, training, verifying and testing the process image labeling model through the process image set, S5, obtaining a target process design image, preprocessing the target process design image, and inputting the process image labeling model to obtain a process design image labeling result. The method has the advantage that the labeling efficiency and consistency of the process design image are greatly improved.

Inventors

LIN XING
ZHUANG HAIJUN

Assignees

福州市数字产业互联科技有限责任公司

Dates

Publication Date: 20260505
Application Date: 20251218

Claims (10)

1. The intelligent labeling method of the process design image based on Florence2 is characterized by comprising the following steps: Step S1, acquiring a large number of historical process design images, and screening, labeling and cross checking each historical process design image; S2, preprocessing at least including size adjustment, image enhancement and data cleaning is carried out on each historical process design image so as to construct a process image set; Step S3, creating a process image annotation model based on a Florence2 network, and configuring pre-training weights and input layer parameters of the process image annotation model; S4, training, verifying and testing the process image annotation model through the process image set; s5, acquiring a target process design image, preprocessing the target process design image, and inputting a process image labeling model to obtain a process design image labeling result.
2. The intelligent labeling method of the process design image based on Florence2 of claim 1, wherein the step S1 is specifically: Acquiring a large number of historical process design images through a data source at least comprising a professional website, a design database and an enterprise album, screening each historical process design image based on three dimensions of style, process and material, marking each screened historical process design image at least comprising a key area, a pattern, material, a process, a tooth edge and meshes, and carrying out preset round cross check on each marked historical process design image so as to ensure the marking accuracy and consistency.
3. The intelligent labeling method of the process design image based on Florence2 of claim 1, wherein the step S2 is specifically: scaling the historical process design images in equal proportion through a Python Piclow library, filling preset colors through edges to unify the sizes of the historical process design images, and further finishing size adjustment; Performing random rotation, random scaling, horizontal overturning and brightness adjustment on each of the historical process design images subjected to size adjustment in parallel to complete image enhancement; Performing definition evaluation on each history process design image after image enhancement based on a definition evaluation algorithm, and removing the history process design images with definition lower than a preset definition threshold value to complete preprocessing; And constructing a process image set based on each preprocessed historical process design image.
4. The intelligent labeling method of the process design image based on Florence2 of claim 1, wherein the step S4 is specifically: Dividing the process image set into a training set, a verification set and a test set according to the ratio of 7:2:1, configuring a loss function of the process image annotation model as a cross entropy loss function, and configuring an optimizer as Adam, an initial learning rate, a learning rate adjustment strategy and a maximum training round; And after each training round is finished, calculating F1 scores by using the verification set to evaluate the performance of the model until the F1 scores of 10 training rounds are not improved continuously or training to the maximum training round, calculating F1 scores, accuracy and recall rate by using the test set to test the trained process image annotation model, and deploying the process image annotation model after the test is passed.
5. The intelligent labeling method of the process design image based on Florence2 of claim 1, wherein the step S5 is specifically: Obtaining a target process design image to be marked, preprocessing the target process design image at least comprising size adjustment, image enhancement and data cleaning, inputting the preprocessed target process design image into a deployed process image marking model for reasoning, and obtaining a process design image marking result.
6. The intelligent process design image labeling system based on Florence2 is characterized by comprising the following modules: the historical process design image acquisition module is used for acquiring a large number of historical process design images, and screening, labeling and cross checking are carried out on each historical process design image; the process image set construction module is used for carrying out pretreatment at least comprising size adjustment, image enhancement and data cleaning on each historical process design image so as to construct a process image set; The process image annotation model creation module is used for creating a process image annotation model based on the Florence2 network, and configuring the pre-training weight and the input layer parameters of the process image annotation model; The process image annotation model training module is used for training, verifying and testing the process image annotation model through the process image set; The process design image labeling module is used for acquiring a target process design image, preprocessing the target process design image and inputting the target process design image into the process image labeling module to obtain a process design image labeling result.
7. The intelligent labeling system of process design images based on Florence2 as set forth in claim 6, wherein said historical process design image acquisition module is specifically configured to: Acquiring a large number of historical process design images through a data source at least comprising a professional website, a design database and an enterprise album, screening each historical process design image based on three dimensions of style, process and material, marking each screened historical process design image at least comprising a key area, a pattern, material, a process, a tooth edge and meshes, and carrying out preset round cross check on each marked historical process design image so as to ensure the marking accuracy and consistency.
8. The intelligent labeling system of process design images based on Florence2 of claim 6, wherein the process image set construction module is specifically configured to: scaling the historical process design images in equal proportion through a Python Piclow library, filling preset colors through edges to unify the sizes of the historical process design images, and further finishing size adjustment; Performing random rotation, random scaling, horizontal overturning and brightness adjustment on each of the historical process design images subjected to size adjustment in parallel to complete image enhancement; Performing definition evaluation on each history process design image after image enhancement based on a definition evaluation algorithm, and removing the history process design images with definition lower than a preset definition threshold value to complete preprocessing; And constructing a process image set based on each preprocessed historical process design image.
9. The intelligent labeling system for process design images based on Florence2 of claim 6, wherein the process image labeling model training module is specifically configured to: Dividing the process image set into a training set, a verification set and a test set according to the ratio of 7:2:1, configuring a loss function of the process image annotation model as a cross entropy loss function, and configuring an optimizer as Adam, an initial learning rate, a learning rate adjustment strategy and a maximum training round; And after each training round is finished, calculating F1 scores by using the verification set to evaluate the performance of the model until the F1 scores of 10 training rounds are not improved continuously or training to the maximum training round, calculating F1 scores, accuracy and recall rate by using the test set to test the trained process image annotation model, and deploying the process image annotation model after the test is passed.
10. The intelligent labeling system for process design images based on Florence2 of claim 6, wherein the process design image labeling module is specifically configured to: Obtaining a target process design image to be marked, preprocessing the target process design image at least comprising size adjustment, image enhancement and data cleaning, inputting the preprocessed target process design image into a deployed process image marking model for reasoning, and obtaining a process design image marking result.

Description

Florence 2-based intelligent labeling method and system for process design image Technical Field The invention relates to the technical field of computer vision, in particular to a Florence 2-based intelligent labeling method and system for process design images. Background In the process of process design and industrial application, accurate, standard and standardized labeling of process design images is a key link for connecting the whole flows of design, production, circulation and the like. The labeling information usually includes core elements such as pattern type (e.g. dots, hollowed-out, laces, etc.), process characteristics (e.g. embroidery, crocheting, jacquard, etc.), material properties, size parameters, and style positioning. The information is not only the basis of efficient communication and collaborative creation among designers, but also the important basis for guiding the production process formulation, the material purchase of a supply chain and the introduction of products in terminal sales, and directly influences the efficiency and quality of the conversion of the process products from the conceptual design to the market, such as lace products. At present, the labeling work of the process design image is mainly finished manually. The pattern has high requirements on the professional ability of the labeling personnel, the labeling personnel needs to have systematic process design knowledge, can accurately distinguish different pattern styles and process details, and is technically described by applying industry standard terms. However, manual labeling has several significant drawbacks in practical operation: 1. The efficiency is low, the manual labeling is required to identify and input information piece by piece in the face of massive process design images brought by rapid updating, the labor cost is high, and the timeliness requirements of rapid landing and efficient order response of the design are difficult to meet. 2. The labeling consistency is difficult to ensure, and because of the difference between the experience level and subjective judgment of labeling personnel, the pattern classification or process identification of the same process image may be inconsistent, so that the selection and material matching of the subsequent production process are affected, the quality problem of the product is even caused, and the reworking and communication cost is increased. With the rapid development of artificial intelligence technology, technologies based on image recognition and intelligent labeling have shown efficient and stable application effects in a plurality of fields such as medical imaging, industrial quality inspection, electronic commerce commodity management and the like. The visual-language model based on deep learning is used as a core technology of intelligent image annotation, and automatic identification of image content and text description generation can be realized through training. However, in the field of image annotation of technical design at present, a customized intelligent annotation scheme suitable for the special characteristics (such as complex pattern structure and various process types) of the image annotation is not yet presented, and the bottleneck of manual annotation in terms of efficiency and accuracy is still difficult to break through by means of the existing AI technology. Therefore, how to provide a method and a system for intelligent labeling of process design images based on Florence2, so as to improve labeling efficiency and consistency of the process design images, is a technical problem to be solved urgently. Disclosure of Invention The invention aims to solve the technical problem of providing a Florence 2-based intelligent labeling method and system for process design images, which can improve the labeling efficiency and consistency of the process design images. In a first aspect, the invention provides a Florence 2-based intelligent labeling method for a process design image, which comprises the following steps: Step S1, acquiring a large number of historical process design images, and screening, labeling and cross checking each historical process design image; S2, preprocessing at least including size adjustment, image enhancement and data cleaning is carried out on each historical process design image so as to construct a process image set; Step S3, creating a process image annotation model based on a Florence2 network, and configuring pre-training weights and input layer parameters of the process image annotation model; S4, training, verifying and testing the process image annotation model through the process image set; s5, acquiring a target process design image, preprocessing the target process design image, and inputting a process image labeling model to obtain a process design image labeling result. Further, the step S1 specifically includes: Acquiring a large number of historical process design images through a data source at least