Image Segmentation [Code]

Image segmentation using modified UNET. Dataset is obtained from Kaggle. Data is split into train and test set in 80:20 ratio.

Example of original data

Model Architecture

• MobileNetV2 is used as downsampler of the modified UNET model.
• Upsampler part of the model consist of 4 Conv2DTranspose + BatchNormalization + ReLU layers initialized through tensorflow_examples.models.pix2pix.upsampler() function. Each layer is initialized with filters of size 3x3 with:
  • Upsample Layer 1: 512 filters
  • Upsample Layer 2: 256 filters
  • Upsample Layer 3: 128 filters
  • Upsample Layer 4: 64 filters
• Model is trained for 20 epochs, with batch size 16.
• Best accuracy (0.9937) is obtained on epoch 19.

Model architecture summary

Prediction at end of epoch 1

Prediction at end of epoch 5

Prediction at end of epoch 10

Prediction at end of epoch 15

Prediction at end of epoch 20

Predictions on test data

Accuracy

Loss