A Few-Sample Defect Detection Method Based on Metric Learning

Abstract

The invention discloses a small sample defect detection method based on metric learning, comprising step (1): through G 2 ‑GAN adversarial network performs data enhancement on the small sample defect data set; Step (2): The convolutional neural network SKM‑CNN based on convolution kernel adaptation extracts the features of the defect data set similar to the small sample defect data set to be detected, Generate a pre-trained model; Step (3): Migrate the pre-trained model to a small-sample defect detection network S based on metric learning 2 D 2 In N, the rapid identification and localization of defects is realized by the method of first target feature extraction and then metric learning. The invention completes defect detection quickly and efficiently, and solves the problems of low detection accuracy, low reliability, overfitting and the like of a model trained under the condition of lack of sample data.

Description

technical field [0001] The invention belongs to the technical field of defect detection of large assembly parts, in particular to a small sample defect detection method based on metric learning. Background technique [0002] At present, the three major tasks of computer vision include detection, classification and segmentation, which are all implemented based on a large number of labeled images. The deep learning algorithm has the advantages of good versatility, high detection accuracy, strong robustness, and good generalization ability. Large-scale high-end equipment for aviation, aerospace, and navigation ("three aviation") is a national defense security guarantee and an important part of my country's equipment manufacturing industry. The product quality of high-end equipment directly affects the final combat performance and international competitiveness. [0003] Automatic defect detection on the surface of large assembly parts is of great significance to ensure the norma...

AI Technical Summary

Problems solved by technology

[0005] Defects seriously affect the safety performance and service life of high-end equipment

The visual inspection method has a large workload, takes a long time, has low efficiency, is easily affected by subjective experience, and will introduce many additional human detection errors

This method is difficult to ensure the efficiency and accuracy of defect detection, which brings uncontrollable factors to the assembly speed of subsequent large-scale assembly parts, and the quality of high-end equipment will also be affected to a certain extent

[0006] Before deep learning, defect detection algorithms often used the method of manually extracting features. This method showed shortcomings such as low efficiency, inaccurate extracted features, and low intelligence, which greatly reduced the practicability and generalization of the model.

Deep learning algorithms rely on a large amount of trainable data, and deep learning algorithms cannot play their true role in the context of small sample data sets

Images