Screw absence classification on aluminum plate via feature based transfer learning models
Screw, the little element that use to join two or more objects together. It is able to hold soft material such as wood or plastic or hard material such as metal or concrete. It is also widely used in industries to fasten objects together. Manual inspection processes by humans often lead to human fai...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | en |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://umpir.ump.edu.my/id/eprint/46733/1/Screw%20absence%20classification%20on%20aluminum%20plate%20via%20feature%20based%20transfer%20learning%20models.pdf https://umpir.ump.edu.my/id/eprint/46733/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Screw, the little element that use to join two or more objects together. It is able to hold soft material such as wood or plastic or hard material such as metal or concrete. It is also widely used in industries to fasten objects together. Manual inspection processes by humans often lead to human failure due to tiredness, lack of focus or even distraction during work. It is very common as humans will get bored with 8 hours of repetitive work daily. Screw absences classification on aluminum plate via feature-based transfer learning models is the target of today’s study. Nevertheless, extracting and attaining significations features from collected datasets is also somewhat quite challenging process. From literature wise, it shows that such screw detection can be seamlessly extracted in variety different applications using deep learning, especially transfer learning (TL). Limited study is made based on screw detection using MVTec Halcon Software or Transfer Learning accompanied by classical Machine Learning (ML) pipelines. 200 datasets are collected from TT Vision Technologies Sdn Bhd which included 100 images of presence screw and 100 images of absence screw. The collected datasets are then undergoes the features extraction process from different TL models. Then extracted features are then classified through four classical ML models, namely random forest, decision tree classifier, support vector machine and logistic regression to determine the optimal pipeline of extracted features. The hyperparameters of those ML models are matched with 5-fold cross-validation technique through an extensive grid search approach. The training, validation and testing of the model are split into a stratified ratio of 60:20:20. The results obtained from the transfer learning and machine learning pipeline are evaluated in terms of classification accuracy and inference time. On the other hand, seventeen transfer learning models which are, VGG19, VGG16, MobileNet, MobileNetV2, ResNet50, ResNet101, ResNet152, ResNet50 V2, ResNet101 V2, ResNet152 V2, DenseNet 121, DenseNet 169, DenseNet 201, NasNet Large, NasNet Mobile, Inception V3 and Xception is used in this study. Whilst it was observed that random forest with xception transfer learning model achieve the classification accuracy of 100% for training, validation and testing with inference time of 0.28s. In conclusion, xception transfer learning model with random forest is the most suitable for screw detection. |
|---|