Machine learning data enhancement method based on feature set

Abstract

A machine learning data enhancement method based on feature set comprises the following steps: S1, feature extraction: obtaining an image data set, and cutting each picture to a region of interest corresponding to the aspect of an inspected component; s2, feature concentration: performing feature concentration on the marked data and the non-marked data, and respectively generating feature concentration objective functions of the marked data and the non-marked data; s3, target training: generating a training target model according to a feature set target function of the marked data and the non-marked data, and completing training; s4, experiment and test: selecting a data set containing the same number of normal and defect images for verification and test; according to the method, feature sets of marked data and unmarked data are simultaneously applied, so that the problems that the number of normal and bad image samples is unbalanced and data marking is difficult are solved, and the image classification precision is effectively improved.

Description

technical field [0001] The invention relates to the technical field of automatic visual inspection, in particular to a method for enhancing machine learning data based on feature concentration. Background technique [0002] Automatic visual inspection is the key to production efficiency, and convolutional neural network (CNN) outperforms previous methods for processing visual tasks and has become the standard method for processing visual inspection images in the industry. However, the use of CNNs for visual inspection is still limited by the availability of training data. One is the quantity imbalance between normal sample images and bad sample images; the other is the difficulty of data labeling. Obtaining consistent labels requires professional inspectors. It is easy to obtain a large number of unlabeled detection images, but the labeling costs are expensive. This increases the difficulty of image classification. Therefore, it is an urgent problem to develop effective met...

AI Technical Summary

Problems solved by technology

but is still limited by the availability of training data in visual inspection using CNNs

One is the quantity imbalance between normal sample images and bad sample images; the other is the difficulty of data labeling, obtaining consistent labels requires professional inspectors, it is easy to obtain a large number of unlabeled detection images, but the labeling costs are expensive

This increases the difficulty of image classification. Therefore, it is an urgent problem to develop effective methods for the problems of unbalanced number of images and difficult data labeling that affect the accuracy of image classification.

[0003] Due to technical reasons, most technical researchers now solve the problem that the number of images is unbalanced and data labeling is difficult, which affects the accuracy of image classification. The method is to use oversampling and undersampling in CNN, and their combination methods, as well as supervision and Semi-supervised methods, and these methods have the following problems: the accuracy of image classification is low

Images