Development and Implementation of a Defect Detection Model for Microstructures Using Image Processing Methods

The aim of this research is to develop and implement artificial intelligence models for the automatic detection of defects in the microstructures of austempered ductile iron (ADI). Our research used three different approaches, representing various categories of machine learning tasks: image classification (ResNet), pixel-wise segmentation (UNet), and object detection (YOLO). Each of the models were adapted to the specific characteristics of the dataset and tested on a collection of microstructural images prepared within the scope of the research. The data preparation process included clustering using the k-means method, morphological operations, generation of binary masks, conversion of labels into formats required by each architecture, and data augmentation to increase the diversity of training samples. The results demonstrated that ResNet achieved very high classification accuracy but did not provide spatial information about defect localization. UNet produced precise segmentation masks of martensitic regions, allowing for quantitative analysis of samples, although it required significantly higher computational resources and struggled with detecting very small defects. YOLO, in turn, enabled fast detection of defects in the form of bounding boxes. In summary, each model proved effective in a different context: ResNet for preliminary classification, UNet for detailed laboratory analysis, and YOLO for industrial detection tasks.