Defect Detection on 3D Print Products and in Concrete Structures Using Image Processing and Convolution Neural Network

Selorm Garfo,M.A Muktadir,Sun Yi

doi:10.3844/jmrsp.2020.74.84

Abstract

This paper explores the automated detection of surface defects on 3-D printed products and concrete structures. They are the main factors to evaluate their quality in addition to dimension and roughness. Traditional detection by human inspectors is far from satisfactory. Manual inspection is time-consuming, error-prone and often leads to loss of resources. For this purpose, image processing and deep learning-based object detection adopted by Google Cloud Machine Learning (ML) Engine is used to detect surface defects. In the case of image processing, two approaches are presented in this paper. In both cases, pixels are being considered to differentiate a smooth or rough surface from a picture taken by a USB camera. For the deep learning- based solution, MobileNet -a base convolution neural network treated as an image feature extractor in combination with Single Shot MultiBox Detector (SSD) as an object detector hence MobileNet-SSD. The model was successfully trained on the Google Cloud ML Engine with the dataset of 20000+ images. The review of the results confirms that with the help of MobileNet-SSD can automatically detect surface defects more accurately and rapidly than conventional deep learning methods.

Highlights

There is a rise in the necessity for object detection in civil infrastructure (Jahanshahi and Masri, 2012) and the manufacturing sector (Delli and Chang, 2018) in recent years
In efforts to train for object detection using a local Computer Processing Unit (CPU) computer, we found that the computational loads are substantially heavy which takes CPU days to train. (Training refers to the computer memory task in preparation for object recognition)
The deep learning model was successfully trained on the Google Cloud Machine Learning (ML) Engine with the dataset of 20000+ images

Summary

Introduction

There is a rise in the necessity for object detection in civil infrastructure (Jahanshahi and Masri, 2012) and the manufacturing sector (Delli and Chang, 2018) in recent years. Detecting defects by performing quality monitoring at various (critical) stages of the printing process helps in assuring corrective measures and eliminates the waste of printing bad parts. Jovančević et al used a similarity measure to compare projected geometric features from CAD models with detected ones in actual images. Present similarity measure for segments is modified to ellipses. This comparison enables the detection and data association processes for navigation and inspection tasks on aircraft parts (Jovančević et al, 2016). Roberson et al (2013) developed a decision making and ranking model for selecting an appropriate 3D printer using Deng’s Similarity approach based on accuracy, printing time and product surface smoothness. Wu et al widely presented a machine learning and image classification method to detect the infill defects in the 3D printing process. The method explored feature extraction and implementation of Naive Bayes Classifier and J48 Decision Trees algorithms (Wu et al, 2016)

Methods

Results

Conclusion