CN-122023325-A - Mobile phone motherboard appearance defect detection method adopting multi-scale feature fusion

CN122023325ACN 122023325 ACN122023325 ACN 122023325ACN-122023325-A

Abstract

The invention discloses a mobile phone motherboard appearance defect detection method adopting multi-scale feature fusion, which is executed by a computer and comprises the steps of receiving defect types input by a user, reordering a plurality of mobile phone motherboard appearance images based on the defect types, displaying the reordered plurality of mobile phone motherboard appearance images to the user, obtaining an optimized mobile phone motherboard appearance image based on the mobile phone motherboard appearance image selected by the user, and providing the optimized mobile phone motherboard appearance image to a server, wherein the server trains a deep convolutional neural network based on a multi-scale feature fusion module by utilizing the optimized mobile phone motherboard appearance image. The invention can improve the detection precision and robustness.

Inventors

LIAO DONGQI
LI WEI

Assignees

深圳市阿龙通讯技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (10)

1. A method for detecting appearance defects of a mobile phone motherboard by adopting multi-scale feature fusion, wherein the method is executed by a computer and comprises the following steps: Receiving a defect type input by a user; reordering the appearance images of the mobile phone mainboards based on the defect types, wherein the appearance images of the mobile phone mainboards are obtained through different types of sensors; displaying the reordered appearance images of the plurality of mobile phone mainboards to a user; obtaining an optimized mobile phone motherboard appearance image based on the mobile phone motherboard appearance image selected by the user, and And providing the optimized mobile phone motherboard appearance image to a server, wherein the server trains a deep convolutional neural network based on a multi-scale feature fusion module by utilizing the optimized mobile phone motherboard appearance image.
2. The method of claim 1, wherein the plurality of phone motherboard appearance images comprises a first phone motherboard appearance image set acquired by a first type of sensor and a second phone motherboard appearance image set acquired by a second type of sensor; wherein reordering the plurality of mobile phone motherboard appearance images based on the defect type comprises: Determining a type of sensor associated with the defect type based on the defect type; If it is determined that the type of sensor associated with the defect type is the first type, reordering the first set of handset motherboard appearance images before the second set of handset motherboard appearance images; If it is determined that the type of sensor associated with the defect type is the second type, the second set of handset motherboard appearance images is reordered before the first set of handset motherboard appearance images.
3. The method of claim 1, wherein obtaining an optimized phone motherboard appearance image based on the user-selected phone motherboard appearance image comprises: Determining a defect area in the appearance image of the mobile phone motherboard selected by the user; determining whether the defect area is positioned at the edge of the mobile phone motherboard appearance image selected by the user; If the defect area is determined to be positioned at the edge of the mobile phone motherboard appearance image selected by the user, a first command is sent to a first sensor for acquiring the mobile phone motherboard appearance image selected by the user; And receiving a re-acquired first mobile phone motherboard appearance image sent by a first sensor, wherein the re-acquired first mobile phone motherboard appearance image is re-acquired by the first sensor based on the first command, and the defect area is positioned in the center of the re-acquired first mobile phone motherboard appearance image in the re-acquired first mobile phone motherboard appearance image.
4. The method of claim 3, wherein obtaining an optimized phone motherboard appearance image based on the user-selected phone motherboard appearance image further comprises: determining the proportion of the area of the defect area to the area of the mobile phone motherboard appearance image selected by the user; If the ratio is determined to be less than a threshold value, a second command is sent to the first sensor; and receiving a re-acquired second mobile phone motherboard appearance image sent by the first sensor, wherein the re-acquired second mobile phone motherboard appearance image is re-acquired by the first sensor based on the second command, and the ratio of the area of the defect area to the area of the mobile phone motherboard appearance image selected by the user in the re-acquired second mobile phone motherboard appearance image is greater than a threshold value.
5. The method of claim 1, wherein the method further comprises: And performing defect detection on the appearance of the mobile phone motherboard by using the trained deep convolutional neural network.
6. A mobile phone motherboard appearance defect detection system adopting multiscale feature fusion, the system includes: A receiving module configured to receive a defect type input by a user; A reordering module configured to reorder a plurality of mobile phone motherboard appearance images based on the defect type, wherein the plurality of mobile phone motherboard appearance images are acquired by different kinds of sensors; A display module configured to display the reordered plurality of mobile phone motherboard appearance images to a user; An optimization module configured to obtain an optimized mobile phone motherboard appearance image based on the mobile phone motherboard appearance image selected by the user, and A transmitting module configured to provide the optimized handset motherboard appearance image to a server, wherein the server trains a deep convolutional neural network based on a multi-scale feature fusion module using the optimized handset motherboard appearance image.
7. The system of claim 6, wherein the plurality of phone motherboard appearance images comprises a first set of phone motherboard appearance images acquired by a first type of sensor and a second set of phone motherboard appearance images acquired by a second type of sensor; wherein reordering the plurality of mobile phone motherboard appearance images based on the defect type comprises: Determining a type of sensor associated with the defect type based on the defect type; If it is determined that the type of sensor associated with the defect type is the first type, reordering the first set of handset motherboard appearance images before the second set of handset motherboard appearance images; If it is determined that the type of sensor associated with the defect type is the second type, the second set of handset motherboard appearance images is reordered before the first set of handset motherboard appearance images.
8. The system of claim 7, wherein obtaining an optimized phone motherboard appearance image based on the user-selected phone motherboard appearance image comprises: Determining a defect area in the appearance image of the mobile phone motherboard selected by the user; determining whether the defect area is positioned at the edge of the mobile phone motherboard appearance image selected by the user; If the defect area is determined to be positioned at the edge of the mobile phone motherboard appearance image selected by the user, a first command is sent to a first sensor for acquiring the mobile phone motherboard appearance image selected by the user; And receiving a re-acquired first mobile phone motherboard appearance image sent by a first sensor, wherein the re-acquired first mobile phone motherboard appearance image is re-acquired by the first sensor based on the first command, and the defect area is positioned in the center of the re-acquired first mobile phone motherboard appearance image in the re-acquired first mobile phone motherboard appearance image.
9. The system of claim 8, wherein obtaining an optimized phone motherboard appearance image based on the user-selected phone motherboard appearance image further comprises: determining the proportion of the area of the defect area to the area of the mobile phone motherboard appearance image selected by the user; If the ratio is determined to be less than a threshold value, a second command is sent to the first sensor; and receiving a re-acquired second mobile phone motherboard appearance image sent by the first sensor, wherein the re-acquired second mobile phone motherboard appearance image is re-acquired by the first sensor based on the second command, and the ratio of the area of the defect area to the area of the mobile phone motherboard appearance image selected by the user in the re-acquired second mobile phone motherboard appearance image is greater than a threshold value.
10. The system of claim 9, wherein the system further comprises: And the defect detection module is configured to detect defects on the appearance of the mobile phone motherboard by using the trained deep convolutional neural network.

Description

Mobile phone motherboard appearance defect detection method adopting multi-scale feature fusion Technical Field The invention relates to the technical field of artificial intelligence, in particular to a mobile phone motherboard appearance defect detection method adopting multi-scale feature fusion. Background When detecting the appearance defects of the mobile phone motherboard based on the neural network with multi-scale feature fusion, the to-be-detected image is generally input into a deep convolution network, the network firstly extracts feature graphs with different layers and different scales, then integrates the features through a specific fusion module (such as a feature pyramid FPN), and finally locates and classifies the defects by using the fused rich information. The superior performance of the above-described network is highly dependent on the quality and representativeness of the training samples. However, the defect samples in the actual production of the mobile phone motherboard are rare and various in form, and the model trained by directly using the conventional image is difficult to fully learn the multi-scale characteristics of the tiny and fuzzy defects, so that the detection precision is limited. Therefore, how to efficiently construct a high-quality defect sample set becomes a key premise for improving such detection methods. The prior art lacks a scheme for efficiently constructing a high-quality defect sample set, and the invention aims to solve the problems of the prior art. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a mobile phone motherboard appearance defect detection method adopting multi-scale feature fusion, which guides a screening process by manually designating defect types, a system sorts and displays images acquired by different sensors according to the defect types for a user to select, and then the system re-shoots by controlling the sensors so as to generate an optimized image for training a multi-scale feature fusion neural network. The method directly aims at the problem of insufficient feature learning caused by too small area occupation ratio and off-center position of the defect area in the conventional training image, and the neural network can learn the multi-scale features of the micro defects more effectively by generating the high-quality training sample with prominent defects and obvious features, so that the accuracy and the robustness of the follow-up actual detection are fundamentally improved. The invention provides a method for detecting appearance defects of a mobile phone motherboard by adopting multi-scale feature fusion, which is executed by a computer and comprises the following steps: Receiving a defect type input by a user; reordering the appearance images of the mobile phone mainboards based on the defect types, wherein the appearance images of the mobile phone mainboards are obtained through different types of sensors; displaying the reordered appearance images of the plurality of mobile phone mainboards to a user; obtaining an optimized mobile phone motherboard appearance image based on the mobile phone motherboard appearance image selected by the user, and And providing the optimized mobile phone motherboard appearance image to a server, wherein the server trains a deep convolutional neural network based on a multi-scale feature fusion module by utilizing the optimized mobile phone motherboard appearance image. In a preferred embodiment, the plurality of mobile phone motherboard appearance images includes a first mobile phone motherboard appearance image set and a second mobile phone motherboard appearance image set, wherein the first mobile phone motherboard appearance image set is acquired by a first type of sensor, and the second mobile phone motherboard appearance image set is acquired by a second type of sensor; Wherein, reordering the plurality of mobile phone motherboard appearance images based on the defect type includes: determining a type of sensor associated with the defect type based on the defect type; If it is determined that the type of sensor associated with the defect type is a first type, reordering the first set of handset motherboard appearance images to before the second set of handset motherboard appearance images; If the type of sensor associated with the defect type is determined to be a second type, the second set of handset motherboard appearance images is reordered before the first set of handset motherboard appearance images. In a preferred embodiment, obtaining an optimized mobile phone motherboard appearance image based on the mobile phone motherboard appearance image selected by the user includes: Determining a defect area in the appearance image of the mobile phone motherboard selected by a user; determining whether the defect area is positioned at the edge of the mobile phone motherboard appearance image selected by the user; if the defect area is determin