A Surface Anomaly Detection Method Based on Hybrid Supervised Learning

Abstract

The invention specifically discloses a surface abnormality detection method based on mixed supervision learning. The method includes the following steps: firstly acquiring normal samples with normal surfaces and abnormal samples with abnormal surfaces and preprocessing them; and constructing a neural network model including an anomaly localization network, a self-attention network and an anomaly discriminating network, and then preprocessing the data. Input the neural network model for training to obtain the abnormality detection neural network model; finally, input the image data to be tested into the abnormality detection neural network model, and then automatically determine whether the image to be tested is abnormal and locate the abnormal area. In the present invention, only a small number of abnormal samples need to be roughly labeled, and there is no need to provide a large number of finely labeled defect samples as training samples, which reduces the need for high-precision labeling of fully supervised learning, greatly reduces labeling costs, improves detection efficiency, and can be accurate and efficient. Complete industrial surface inspection tasks.

Description

technical field [0001] The invention relates to the technical field of deep learning and industrial detection, and in particular to a surface abnormality detection method based on hybrid supervision learning. Background technique [0002] Anomaly detection refers to detection that can detect heterogeneous or unexpected patterns in a homogeneous set of natural images. Currently, anomaly detection has a large number of applications, including visual industrial inspection. Quality inspection of products and their surfaces is an important part of industrial production. Quality inspection methods include traditional quality inspection methods and deep learning-based quality inspection methods. Traditional machine vision methods have been widely used in automated visual inspection. Process, with the continuous advancement of industrial automation, new detection methods based on deep learning have also begun to be used. Traditional machine learning models rely on specific visual de...

AI Technical Summary

Problems solved by technology

The quality inspection of products and their surfaces is an important link in industrial production. Quality inspection methods include traditional quality inspection methods and quality inspection methods based on deep learning. Traditional machine vision methods have been widely used in automated visual inspection. With the continuous advancement of industrial automation, new detection methods based on deep learning have also begun to be used. Traditional machine learning models rely on specific visual detection tasks obtained by manual analysis and extraction of defect features, and then use rule-based advanced Based on empirical knowledge or learning-based classifiers to make decisions, such as support vector machines, neural networks and decision numbers, etc., in this method, the system performance is heavily dependent on the accurate representation of specific feature types, so this method is very inefficient, subjective and Time-consuming; unlike traditional machine learning methods, deep learning models are more suitable for anomaly detection tasks because they can automatically learn features from low-level data and have strong modeling capabilities for complex features without manual intervention. The success of the algorithm largely depends on the labeled images used to train an effective deep network. Since abnormal samples are very rare in industrial production lines, and the high cost of pixel-level labeling leads to lack of training data, it will seriously affect the detection performance of the model. Therefore, it is of great significance to improve the economic benefits of industrial products by reducing the required labeling amount and reducing the expected labeling accuracy to minimize the labeling work, and at the same time improve the detection efficiency and detection accuracy.

Images