shui-yu / image-classification

Image-Classification

This repository includes four Jupyter notebooks regarding image classification of pipe damages.

Original Dataset:

• 3 classes: Dent/Slit/Hole
• Training size: 286 (Devided into training and validation set)
• Test size: 71

01_cnn.ipynb:

• Constructed 4 Constructed a Convolutional Neural Network models
• Improved model performance step by step conducting parameter tuning and adding regularization.
• Obtained best model which has training accuracy 98.6% and test accuracy 83.1%

02_cnn_with_data_augmentation.ipynb

• Implemented data augmentation by randomly applying width/height shift, brightness and channel shift
• Resulted in doubline the training set size
• Achieved training accuracy 100% and test accuracy 88.7%

03_trasfer_learning.ipynb

• Conducted transfer learning on VGG-16 model, adjusting the fully connected layers to accommondate my problem
• Achieved training accuracy 98.1% and test accuracy 90.1%

04_image_cropping.ipynb

• Transformed original dataset to image sequences by cropping 20 images from each original image

05_convlstm.ipynb

• Experimented a ConvLSTM model to classify image sequences