Building a Fused Deposition Modelling (FDM) 3D Printing Visual Defect Detection System, Part 1: Creation of a Dynamic Imaging System, a Pixel-wise Segmentation Dataset with a Hybrid Synthetic Data Creation Method, and a Semantic Segmentation Algorithm with the SegFormer Deep Learning Model

Abstract

As a part of an effort to develop a surface defect detection system for FDM 3D printed objects, this work project studies the application of the SegFormer network to semantically segment 3D printed objects. The project also showcases an affordable and accessible imaging system designed for the surface defect detection system, to support the decisions made during the segmentation task and to be used to evaluate the segmentation models. To achieve this, the first-ever pixel-wise annotation dataset of 3D-printed object images was created. Model-O1, a SegFormer MiT-B0 model trained on this dataset with minimal data augmentation resulted in an Intersection-over-Union score of 87.04%. A synthetic data creation method that caters to the nature of 3D printed objects was also proposed, which expands upon existing synthetic data creation methods. The model trained on this dataset, Model-A2, achieved an IoU score of 89.31%, the best performance achieved among the models developed in this project. During the evaluation of the model based on the inference results, Model-A2 was also identified to be the most practical model for building a surface defect detection system.