U Net Lowered with Keras

Complete U-net Implementation with keras

Original Paper Link : https://arxiv.org/abs/1505.04597

Special Implementations :

The model is implemented using the original paper. But I have changed the number of filters of the layers. The implemented number of layers are reduced to 25% of the original paper.

Original Model Architecture :

Dataset :

The dataset has been taken from kaggle . It had a specific directory tree, but it was tough to execute dataset building from it, so I prepared an usable dat directory.

Link : https://www.kaggle.com/azkihimmawan/chest-xray-masks-and-defect-detection

Primary Directory Tree :

.
└── root/
    ├── train_images/
    │   └── id/
    │       ├── images/
    │       │   └── id.png
    │       └── masks/
    │           └── id.png
    └── test_images/
        └── id/
            └── id.png

Given Images :

Image	Mask

Supporting Libraries :

Numpy	opencv	Matplotlib

Library Versions :

All versions are up to date as per 14th June 2021.

Dataset Directory Generation :

We have performed operations to ceate the data directory like this :

              .
              └── root/
                  ├── train/
                  │   ├── images/
                  │   │   └── id.png
                  │   └── masks/
                  │       └── id.png
                  └── test/
                      └── id.png

Model Architecture ( U-Net Lowered ):

Model: “UNet-Lowered”

Data Preparation :

Taken single channels of both image and mask for training.

Hyperparameters :

      Image Shape : (512 , 512 , 1)
      Optimizer : Adam ( Learning Rate : 1e-4 )
      Loss : Binary Cross Entropy 
      Metrics : Accuracy
      Epochs on Training : 100
      Train Validation Ratio : ( 85%-15% )
      Batch Size : 10

Model Evaluation Metrics :

Model Performance on Train Data :

Model Performance on Validation Data :

One task left : Will update the tutorial notebooks soon ;)

Conclusion :

The full model on the simpliefied 1 channel images was giving bad overfitted accuracy. But this structure shows better and efficient tuning over the data.

STAR the repository if this was helpful :) Also follow me on kaggle and Linkedin.

THANK YOU for visiting :)