Image Segmentation on a large scale fish dataset

Performed image segmenatation on the large scale fish dataset

link to Notebook

Data Preprocessing

Images in the dataset were already augmented.

• Used Scikit-learn to split the dataset into train/test sets.

• Used PyTorch's Dataset class to create my own Custom Fish Dataset as the images and their masks were not arranged as per my needs.

• Resized the Images to 224x224 pixels and converted them to Tensors, to perform these trasnformations I used a library called Albumentations

• Created a custom DataLoader with a batch size of 32

Neural Network

Used a UNet architecture to implement image segmentation.

Training and Testing

The model was trained for 5 Epochs over the train dataset yielding:

Average Epoch Accuracy	Average Epoch Loss	Epoch
61.4%	0.209	1
83.5%	0.0788	2
85.8%	0.0673	3
86.5%	0.0635	4
87%	0.061	5

Average training Accuracy was 80.85%

To predict the accuracy I used the IoU (Intersection over Union) metric

The model was then tested using the test dataset yielding:

Average accuracy of 86.69%

List of libraries used

• numpy
• pandas
• PyTorch
• tqdm
• albumentations
• scikit-learn
• pathlib
• pillow