CN113902939A

CN113902939A - Industrial product large defect detection method based on twin network

Info

Publication number: CN113902939A
Application number: CN202110833024.5A
Authority: CN
Inventors: 许琦
Original assignee: Shenzhen Deepvision Creative Technology Ltd
Current assignee: Shenzhen Deepvision Creative Technology Ltd
Priority date: 2021-07-22
Filing date: 2021-07-22
Publication date: 2022-01-07

Abstract

The invention provides a twin network-based industrial product large defect detection method, which comprises the following steps: directly dividing the images into two categories of OK and NG so as to construct a plurality of image pairs; training the twin network using the image pairs; the image pair respectively passes through the convolution layer of the twin network and the full-connection network to extract a characteristic vector of the image; calculating a distance from the two feature vectors of the pair of images; if the distance is greater than the threshold, then calculating a loss function; if the loss function is 0, not updating the network parameters, and if the loss function is more than 0, updating the network parameters; and (4) identifying the image by using the trained twin network so as to detect the large defects of the product. The network for testing can be trained without manually marking pictures, and a single model can test a plurality of products instead of one product corresponding to one model.

Description

Industrial product large defect detection method based on twin network

Technical Field

The invention relates to the field of defect detection, in particular to a method for detecting large defects of industrial products based on a twin network.

Background

The rapid development of deep learning widens the application boundary of computer vision in industrial products, such as target positioning, identification, defect detection and the like. In an actual application scenario, deep learning belongs to supervised learning, and training is often performed by relying on a large amount of manual labeling data. Deep learning is mainly information of fitting data, and the deviation of the model can be caused by weak change of the data. In actual production, pictures of industrial products often need high-resolution pictures, so that the pictures occupy a large memory, and the pictures cannot be directly input into a model but input into a small picture. In actual production, due to illumination change of the photos, abrasion of equipment and the like, large defects such as large chipping, large cracks and the like can be caused, so that the model can be invisible to the defects, and large defects are missed. These large defect missed detections are very undesirable for the customer and affect the experience of using the model.

Generally, certain problems can be relieved by adding the large defects into model retraining, but due to the fact that the field environment is severe, pictures with the large defects are often unpredictable and small in quantity, and retraining requires a large amount of manual labeling and time cost. Such an operation is time-consuming, labor-consuming and easy, and has half the result. Certainly, the conventional image processing method is used for solving the problem, but the manual design algorithm can only solve one kind of defects of one product, and has no universality and poor stability.

Disclosure of Invention

The invention provides a twin network-based industrial product large defect detection method, which aims to solve at least one technical problem.

To solve the above problems, as an aspect of the present invention, there is provided a twin network-based industrial product macro defect detection method, including:

directly dividing the images into two categories of OK and NG so as to construct a plurality of image pairs;

training the twin network using the image pairs;

the image pair respectively passes through the convolution layer of the twin network and the full-connection network to extract a characteristic vector of the image;

calculating a distance from the two feature vectors of the pair of images;

if the distance is greater than the threshold, then calculating a loss function;

if the loss function is 0, not updating the network parameters, and if the loss function is more than 0, updating the network parameters;

and (4) identifying the image by using the trained twin network so as to detect the large defects of the product.

Preferably, the threshold value is dynamically settable to control the accuracy of the pair of similar pictures and the pair of dissimilar pictures.

Due to the adoption of the technical scheme, the invention has the following advantages: (1) for obvious defects, a network for testing can be trained without manually marking pictures. (2) A single model may test multiple products instead of one product for each model.

Drawings

Fig. 1 schematically shows a network framework diagram of the present invention.

Detailed Description

The following detailed description of embodiments of the invention, but the invention can be practiced in many different ways, as defined and covered by the claims.

In industrial production, the deep learning network often cannot see the large defect, resulting in missed detection of the large defect. Aiming at the problem in the industry, the invention provides a method for using a twin network to calculate the image similarity to judge whether the product has large defects or not, thereby reducing the omission of the large defects.

The method uses the twin network to compare two pictures to calculate the similarity of the pictures. And training a twin network by using the selected image pair, extracting the characteristics of the two images, and calculating the similarity of the characteristic pair. The method does not need to additionally label the image by using a tool, and can directly classify the image into two categories of OK and NG.

The image pairs are respectively input into the same twin network, then the distance between the characteristics of the two images is calculated, the images are dissimilar when the distance is larger than a certain threshold value, the images are similar when the distance is smaller than the certain threshold value, and the distance between the images with large defects and the images without the defects is larger. In practice, threshold values can be set as required to detect defects of different degrees. For example, in practical applications, the threshold value may be set to 1.5. If the distance between the two pictures is larger than the threshold value of 1.5, the two pictures are judged to be larger, the two pictures belong to a large defect, and if the distance between the two pictures is smaller than the threshold value of 1.5, the two pictures are judged to be similar and are normal pictures.

The invention uses a twin network to input images which are well paired in advance, the OK pictures of the same batch of products are paired with each other to form a positive sample, and the OK pictures and the NG pictures of the same batch are paired with each other to form a negative sample for training a model. And extracting picture characteristics of the two pictures through the convolution layer and the full-connection network respectively, and then calculating the distance. The convolutional network feature extraction may use a commonly used network structure.

Firstly, the paired pictures respectively pass through the convolution layer and the full-connection layer, and the distance is calculated for a group of obtained feature vectors, namely the similarity degree of the two pictures is that the smaller the distance is, the more similar the distance is, and the larger the distance is otherwise. And then calculating a loss function, wherein if the loss function is 0, the network parameters are not updated, and if the loss function is more than 0, the network parameters are updated.

Second, the threshold of the network may be set to 1.0. If the calculated distance is larger than 1.0 when the two normal pictures are judged to be wrong and the loss function is larger than 0, updating the network parameters by using a gradient descent algorithm to find out the parameters of a better network; if the two normal pictures are the same, the distance between the two pictures is calculated to be less than 1.0, and the network parameters are not updated. If the calculated distance is larger than 1.0, the parameters are not updated, otherwise, the calculated distance is smaller than 1.0, the loss function is larger than 0, and the network parameters are updated.

Preferably, different thresholds may be dynamically set to control the accuracy of the similar and dissimilar picture pairs. For example, in the test, since there is a difference in the distance between pictures, for example, the distance between some normal pictures is at most 2.0, the threshold may be set to be 2.0, and if the distance is at most 1.0, the threshold may be set to be 1.0.

The network frame diagram is shown in fig. 1, where the result output indicates the similarity between two pictures, and image 1 and image 2 are the paired pictures:

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A twin network-based industrial product large defect detection method is characterized by comprising the following steps:

training the twin network using the image pairs;

calculating a distance from the two feature vectors of the pair of images;

2. The twin network based industrial product macro defect detection method of claim 1, wherein a threshold value can be dynamically set to control the accuracy of the similar picture pair and the dissimilar picture pair.