Training method for image recognition model, spinneret plate detection method and apparatus

Patent metadata

Field	Value
Publication number	US-2025166360-A1
Application number	US-202418981165-A
Country	US
Kind code	A1
Filing date	Dec 13, 2024
Priority date	Nov 21, 2023
Publication date	May 22, 2025
Grant date	—

Abstract

Provided is a training method for an image recognition model, a spinneret plate detection method, a spinneret plate detection device, and a storage medium, relating to the fields of image recognition and deep learning technologies. The training method includes: processing a spinneret plate sample image based on a first image recognition model to obtain annotation state information of each micro hole in the spinneret plate sample image, wherein the spinneret plate sample image is a projection image formed on an imaging member after light emitted by a light source passes through each micro hole of a spinneret plate sample and is amplified by an amplifying member, the annotation state information is used to characterize a shape state of a micro hole corresponding thereto; and training a preset image recognition model based on the spinneret plate sample image and the annotation state information to obtain a target image recognition model.

First claim

What is claimed is:

1 . A training method for an image recognition model, comprising:
processing a spinneret plate sample image based on a first image recognition model to obtain annotation state information of each micro hole in the spinneret plate sample image; wherein the spinneret plate sample image is a projection image formed on an imaging member after light emitted by a light source passes through each micro hole of a spinneret plate sample and is amplified by an amplifying member; the first image recognition model comprises a first sub model and a second sub model different from the first sub model, the first sub model is configured to process the spinneret plate sample image to obtain a first information set of each micro hole, and the second sub model is configured to process the spinneret plate sample image to obtain a second information set of each micro hole; and the annotation state information of each micro hole is obtained based on the first information set and the second information set and is used to characterize a shape state of each micro hole corresponding thereto; and training a preset image recognition model based on the spinneret plate sample image and the annotation state information of each micro hole to obtain a target image recognition model, wherein the target image recognition model is used to process a spinneret plate projection image to obtain state information of each micro hole in a spinneret plate.

2 . The method of claim 1 , wherein processing the spinneret plate sample image based on the first image recognition model to obtain the annotation state information of each micro hole in the spinneret plate sample image comprises:
inputting the spinneret plate sample image into the first sub model in the first image recognition model to obtain the first information set, wherein the first information set comprises first state information of each micro hole, first position information of each micro hole, and a first confidence level corresponding to the first state information of each micro hole; inputting the spinneret plate sample image into the second sub model in the first image recognition model to obtain the second information set, wherein the second information set comprises second state information of each micro hole, second position information of each micro hole, and a second confidence level corresponding to the second state information of each micro hole; and determining the annotation state information of each micro hole based on the first information set and the second information set.

3 . The method of claim 2 , wherein determining the annotation state information of each micro hole based on the first information set and the second information set comprises:
determining a state information subset of each micro hole based on the first position information in the first information set and the second position information in the second information set, wherein the state information subset comprises the first state information and the second state information of a same micro hole; for the state information subset of a first micro hole of micro holes, determining a comprehensive confidence level of the first micro hole based on the first confidence level corresponding to the first state information of the first micro hole and the second confidence level corresponding to the second state information of the first micro hole, in a case where the first state information of the first micro hole does not match the second state information of the first micro hole; determining the annotation state information of the first micro hole based on the comprehensive confidence level of the first micro hole and a confidence threshold; and determining the annotation state information of each micro hole at least based on the annotation status information of the first micro hole.

4 . The method of claim 1 , wherein training the preset image recognition model based on the spinneret plate sample image and the annotation state information of each micro hole to obtain the target image recognition model comprises:
obtaining prediction state information of each micro hole by processing the spinneret plate sample image based on the preset image recognition model; determining a first loss function based on the prediction state information of each micro hole and the annotation state information of each micro hole; adding a temperature scalar to the first loss function to obtain a second loss function, wherein the temperature scalar is used to balance a sensitivity of the preset image detection model to a difference between micro holes in different shape states; and adjusting a parameter of the preset image recognition model based on the second loss function to obtain the target image recognition model.

5 . The method of claim 4 , wherein adding the temperature scalar to the first loss function to obtain the second loss function comprises:
determining a dynamic temperature function, which is used to characterize a changing relationship of the temperature scalar over time, based on the temperature scalar; and adding the dynamic temperature function to the first loss function to obtain the second loss function.

6 . The method of claim 4 , wherein determining the first loss function based on the prediction state information of each micro hole and the annotation state information of each micro hole comprises:
determining the first loss function based on the prediction state information of each micro hole, the annotation state information of each micro hole and a loss function of the first image recognition model.

7 . A spinneret plate detection method applied to a spinneret plate detection device, which comprises a clamping member, a light source, an imaging member and an amplifying member, the clamping member is configured to clamp a spinneret plate, the light source and the imaging member are located on both sides of the clamping member, and the amplifying member is located between the imaging member and the clamping member, light emitted by the light source can pass through each micro hole to be detected of the spinneret plate, be amplified by the amplifying member and then from a spinneret plate projection image on the imaging member, the method comprises:
obtaining the spinneret plate projection image on the imaging member; processing the spinneret plate projection image by using a target image recognition model to obtain state information of each micro hole to be detected in the spinneret plate, wherein the target image recognition model is obtained by training based on the training method of claim 1 ; and determining a detection result of the spinneret plate based on the state information of each micro hole to be detected.

8 . The method of claim 7 , further comprising:
controlling a mechanical arm in the spinneret plate detection device to move the spinneret plate to a first area for placing qualified spinneret plates, in a case where the detection result satisfies a preset qualification condition.

9 . The method of claim 7 , further comprising:
controlling a mechanical arm in the spinneret plate detection device to move the spinneret plate to a second area for placing unqualified spinneret plates, in a case where the detection result does not satisfy a preset qualification condition; and sending a first notification for characterizing disqualification of the spinneret plate, wherein the first notification comprises a serial number of the spinneret plate.

10 . The method of claim 7 , wherein the spinneret plate detection device comprises a first capturing member and a second capturing member that are respectively positioned at two sides of a central axis of the im

Assignees

• Zhejiang Hengyi Petrochemical Co Ltd

Inventors

• Wang Peng
• Peng Xiantao
• Qiu Yibo
• Liu Mingyi
• Li Dake
• Wang Dandan
• Shen Jun

Classifications

• G06V20/70

  Labelling scene content, e.g. deriving syntactic or semantic representations · CPC title

• G06V10/776 Primary

  Validation; Performance evaluation · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/774 Primary

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title