Final Year Project

MobileNetV2 SSD on Fabric Defect Detection

Report: https://hdl.handle.net/10356/175198

Scripts used for FYP. Ran in Kaggle.

Datasets: From Roboflow

Pretrained models: https://www.kaggle.com/datasets/lax4nd/pretrained-models/data

The pretrained models are model weights generated from the scripts that are used for futher training due to the 12hr GPU runtime limit on Kaggle or for finetuning.

Note that the recent update to Tensorflow 2.16.1 may lead to issues when cloning the tensorflow models repository. Should this error occur, remove the code that clones the tensorflow models repository and replace the tfr_labels function with the tfr_labels function in focal-mobilenetv2-ssd.ipynb.

How to run:

1. Clone the repository or download the script
2. Create Roboflow account and get your Roboflow API key
3. Add your API key to the scripts
4. Change the path to the pretrained model and necessary parameters

Base: MobileNetV2 SSD trained on Capstone (comparing image resolution)

Backbone: MobileNetV2, DenseNet121, SEResNeXt50, EfficientNetB0

Feature Fusion: Concatenate or Elemental Sum (80+40 or 80+40, 40+20 or 80+40+20 or 80+40+20, 40+20)

Detection: Feature Pyramid, FPN, PAFPN, BiFPN

Transfer Learning: Tilda, Daffodil, Thesis using MobileNetV2 SSD pretrained on Capstone

Best: SE ResNeXt50 BiFPN and MobileNetV2 BiFPN

Credit to the following GitHub repositories:

• https://github.com/FurkanOM/tf-ssd/tree/master
• https://github.com/flash10042/EfficientDet/tree/master
• https://github.com/SunYanCN/TensorFlow/blob/master/TensorFlow-Code-PPT/lesson26-fasterRCNN/detection/models/necks/fpn.py
• https://github.com/TanyaChutani/SSD-Tensorflow2.0/tree/master