Multi-target Domain-Based Hierarchical Dynamic Instance Segmentation Method for Steel Defects Detection
Chi Zhang,Xi Zhang
DOI: https://doi.org/10.1007/s00521-022-07990-z
2022-01-01
Abstract:In today’s large-scale popularity of machinery manufacturing, people’s demand for the detection of steel plate surface defects has gradually increased over time. Nowadays, the production of steel plates in various fields has to be in line with the actual product application after multiple rounds of testing. However, due to material, machine, and processing technology problems, defects usually appear on the surface of the steel plate during the process of mass production of steel plates. These defects not only affect the beauty and the esthetics, but also they have an unfavorable impact on the performance of the steel plate, and can even lead to a series of safety problems. Thus, we need to improve the requirements for the detection of defects on the surface of the steel plate. With the improvement of computer computing power and automation technology, deep learning has taken advantage of this development. It is one of the most representative fields of artificial intelligence, excelling in image classification, target detection, segmentation, and tracking of targets. For defect detection tasks with a small number of labeled samples, a large number of unlabeled samples, and a wide range of data sources, the unsupervised domain adaptive method was adopted to realize knowledge transfer between data in different domains to solve the above problems. The unsupervised domain adaptive method can train the defect segmentation model effectively by using labeled source domain data and unlabeled target domain data. In this way, the model can obtain better segmentation results in the target domain without increasing other costs. From the point of view of scene setting, most of the unsupervised domain adaptive methods are focused on a single-source single-target domain. To adapt to the real factory environment, we extended the domain adaptive method of steel plate surface defect feature learning to multi-source single-target and multi-source multi-target domains. Hence, we realized the effective migration of defect features among data in different domains and provided corresponding solutions for segmentation detection of multi-source data sets with few samples and no labels. Multi-target domain technology stands out in the field of related technologies with its powerful memory capability, nonlinear mapping capability, self-learning capability, and robustness. For training, we fed a small group of sample data into the network of generative adversarial networks. In this training, the low-level network could automatically learn the detailed features of the data, and the high-level network could automatically learn the abstract features. To improve industrial automation for the effective detection of steel plate surface defects, a deep learning-based algorithm was proposed for the detection of steel plate surface defects. The method was based on multi-target domain instance segmentation, which could well retain the texture and background integrated features in the visual features of steel plates; the dynamic random operation of the final pooling layer eliminated the multi-scale problem of steel plate vision; the multi-dimensional feature fusion of hierarchy ensured the diversity of features; thus, the recognition accuracy of steel plate surface defects was improved. In the image dataset of the detection of steel plate surface defects, the proposed method achieved outstanding performance compared with the other methods, which proved the effectiveness of the method.