Industrial Vision Inspection System

Abstract

This project comprises a vision inspection system for automatic defect detection employing a Convolutional Neural Network (CNN) built from scratch. The purpose is to improve quality control in industrial settings by detecting products, such as electronic chips, mechanical parts, tiles, pharmaceuticals and others ... with the highest accuracy and zero human intervention. The CNN architecture created in this project is unlike other traditional approaches of machine learning or pre-trained models, where it was hypothesized end-to-end training on a dataset of labeled images of defective and non-defective samples was the way to go.The setup involves a preprocessing pipeline for normalizing and augmenting data followed by a multi-layer CNN specially designed for feature extraction and binary classification. A Qt/QML front-end application serves as an interface for real-time image annotation, from which various defect types can be identified and serial codes tracked via Firebase. Thus, the integration of a bespoke deep-learning model with an interactive application interface creates a well-formed and scalable solution to the automated visual inspection task.