Mitigating Overfitting in Fine-Pattern Quality Classification: Enabling Automated Projection Image Quality Inspection

This paper presents a deep learning-based method for automated projection image quality inspection. Traditional inspection relies on manual visual evaluation, which is timeconsuming, labor-intensive, and inconsistencies due to subjective judgment and operator fatigue. Mainstream object detection models such as the YOLO series are not suited for this task due to their low accuracy on small objects and the requirement to resize high-resolution images an approach typically adopted to reduce computational load and inference time given the complexity of these models which ultimately compromises the quality of fine-pattern features. To overcome these limitations, we propose a two-stage inference framework that integrates template-based object detection with a ResNet classification network. In the first stage, template matching is employed to detect 1,458 fine-patterns from high-resolution projection images. In the second stage, each fine-pattern is classified into one of ten image quality levels using a ResNet50-based model. A key contribution of this work is the introduction of a quality-preserving data augmentation strategy that effectively mitigates overfitting without compromising fine-pattern quality. This approach improves the model's robustness to positional deviations commonly observed in corner fine-patterns of projection images. Experimental results demonstrate that the proposed method achieves high classification accuracy across various geometric conditions and significantly outperforms the baseline in challenging corner regions. Furthermore, the system provides visual feedback through color-coded bounding boxes overlaid on the projection image, enabling intuitive interpretation of spatial image quality. The proposed method offers a practical solution for improving the efficiency and reliability of quality inspection in real-world projection images.