Towards Sim-to-Real Industrial Parts Classification with Synthetic Dataset

Abstract

Automated industrial part classification is important for intelligent manufacturing, robotic inspection, and automated quality control, where accurate visual recognition supports efficient and consistent component identification. However, conventional approaches often require large collections of real-world labeled images, while variations in illumination, background, orientation, and image quality can adversely affect model generalization. The Synthetic Industrial Parts Dataset (SIP-17) is utilized as the source dataset, with the Synthetic data with background subset used for model training and validation and the Realdatam test subset reserved for real-world testing. Images are resized and augmented using random horizontal and vertical flipping, rotation, color jittering, tensor conversion, and normalization. ResNet50, ConvNeXt Tiny, EfficientNet-B0, EfficientNet-B2, Swin Transformer Tiny, ViT-B/16, MobileNetV3, CondConv-EfficientNet, MaxViT-Tiny, and Xception are comparatively evaluated. Model performance is assessed using accuracy, precision, recall, and F1-score on validation and real-world test data to examine sim-to-real generalization. Among the evaluated architectures, MaxViT-Tiny achieves the best performance, attaining 99.8% validation accuracy and 91% accuracy on the real-world test dataset. The resulting system integrates FastAPI-based inference with Grad-CAM++ explanations, providing practical and interpretable industrial part classification under real-world domain variation.