Construction method, system and device of surface scratch detection neural network and medium

Abstract

The invention discloses a construction method, system and device of a surface scratch detection neural network and a medium. The method comprises the steps of obtaining image data with scratches on the surface of an object, and screening the image data based on a preset resolution threshold to obtain a scratch image sample; constructing a basic neural network through a feature extraction module, asegmentation network module and a decision network module; and finally, training the basic neural network through the scratch image sample to obtain a surface scratch detection neural network. The system comprises an acquisition unit, a screening unit, a construction unit and a training unit. By using the method provided by the invention, an object image with a scratch defect is identified, and training of a small number of samples and high-precision detection are realized while the network complexity is low. The method can be widely applied to the technical field of image recognition.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to a construction method, system, device and medium of a surface scratch detection neural network. Background technique [0002] With the rapid development of the field of deep learning, automatic surface defect detection technology has become an important detection technology in industrial applications. However, since only a small number of defect samples can be provided in general industrial production, how to use limited samples to obtain the required features for training and construct a suitable model to achieve higher accuracy is an important issue. [0003] However, the existing surface defect detection technology based on deep learning often has problems such as large amount of calculation and difficult training in the process of pursuing detection accuracy. For example, the LBP feature extraction method, although it can extract image information well when the amo...

AI Technical Summary

Problems solved by technology

However, since only a small number of defect samples can be provided in general industrial actual production, how to use limited samples to obtain the required features for training and construct a suitable model to achieve higher accuracy is an important issue.

[0003] However, the existing surface defect detection technology based on deep learning often has problems such as large amount of calculation and difficult training in the process of pursuing detection accuracy.

For example, the LBP feature extraction method, although it can extract image information well when the amount of data is small, the algorithm is computationally complex and it is difficult to select a suitable sparse dictionary when calculating the corresponding relationship.

In the method based on the convolutional neural network, the network depth is usually deepened on the basis of the shallow network, and the detection accuracy is improved with a more complex network, but this reduces the training efficiency of the network, and the detection takes a long time.

At present, there is still a lack of a good neural network construction technology in the existing schemes to solve the above problems

Images