A DINOv2-based self-supervised learning framework for automated detection of surface defects in injection molding

The startup procedure and quality monitoring in injection molding, traditionally reliant on expert personnel to ensure defect-free production, have become increasingly challenging due to the shortage of skilled operators, the shift toward automation, and growing aesthetic requirements for plastic products. Surface defects such as flash and sink marks compromise both visual quality and functionality, highlighting the limitations of manual inspection, which is labor-intensive and prone to subjectivity. In this study, we propose a framework based on DINOv2 that leverages self-supervised learning for feature extraction and integrates a supervised segmentation head for automated defect classification. The framework is specifically designed to reduce the volume of labeled training data required while maintaining high detection accuracy. Its performance is evaluated through a comparison with a Vision Transformer–based framework trained in a supervised manner. The results demonstrate the superior feature extraction capability of the self-supervised approach, particularly in data-limited scenarios, which is critical for minimizing experimental effort during production startup. In addition, the impact of data augmentation using RandAugment is investigated by applying both geometric and color-based transformations. The findings show that geometric transformations are particularly effective for detecting flash defects, while color-based transformations significantly improve the segmentation of sink marks.