Surface defect inspection system with anomaly detection system

Ling, Androw Kian Quan (2025) Surface defect inspection system with anomaly detection system. Final Year Project, UTAR.

PDF Download (2952Kb)

Abstract

This project proposes a Surface Defect Inspection System with Anomaly Detection Approach to enhance quality control in manufacturing by leveraging deep learning and computer vision techniques. The system utilizes the pretrained ConvNeXt-v2-Tiny architecture as backbone of model that using Fully Convolutional Masked Autoencoder method for self-supervised learning and trained on the MVTec Anomaly Detection Dataset, to classify and detect surface defects such as scratches, cracks, and contamination. By employing this strategy, I fine-tuned the pretrained model that could adapt to MVTec Dataset and extract the good features from "good" samples only, then evaluate the accuracy on test dataset which includes both "good" and "defective" samples to detect the anomalies on that particular samples by comparing the Mahalanobis Distance score with threshold. The final results demonstrate robust performance on detecting both "good" and "defective" samples with correct predicted classes and correct predicted defect types on a well-designed web interface. The system aims to address challenges in traditional inspection methods by offering an automated, scalable, and efficient solution for industrial manufacturing sections.

Actions (login required)