U Net Lowered with TF-Keras

Complete U-net Implementation with tf-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”

Layer Type	Output Shape	Param	Connected to
input_1 (InputLayer)	[(None, 512, 512, 1)	0
conv2d (Conv2D)	(None, 512, 512, 16)	160	input_1[0][0]
conv2d_1 (Conv2D)	(None, 512, 512, 16)	2320	conv2d[0][0]
max_pooling2d (MaxPooling2D)	(None, 256, 256, 16)	0	conv2d_1[0][0]
conv2d_2 (Conv2D)	(None, 256, 256, 32)	4640	max_pooling2d[0][0]
conv2d_3 (Conv2D)	(None, 256, 256, 32)	9248	conv2d_2[0][0]
max_pooling2d_1 (MaxPooling2D)	(None, 128, 128, 32)	0	conv2d_3[0][0]
conv2d_4 (Conv2D)	(None, 128, 128, 64)	18496	max_pooling2d_1[0][0]
conv2d_5 (Conv2D)	(None, 128, 128, 64)	36928	conv2d_4[0][0]
max_pooling2d_2 (MaxPooling2D)	(None, 64, 64, 64)	0	conv2d_5[0][0]
conv2d_6 (Conv2D)	(None, 64, 64, 128)	73856	max_pooling2d_2[0][0]
conv2d_7 (Conv2D)	(None, 64, 64, 128)	147584	conv2d_6[0][0]
dropout (Dropout)	(None, 64, 64, 128)	0	conv2d_7[0][0]
max_pooling2d_3 (MaxPooling2D)	(None, 32, 32, 128)	0	dropout[0][0]
conv2d_8 (Conv2D)	(None, 32, 32, 256)	295168	max_pooling2d_3[0][0]
conv2d_9 (Conv2D)	(None, 32, 32, 256)	590080	conv2d_8[0][0]
dropout_1 (Dropout)	(None, 32, 32, 256)	0	conv2d_9[0][0]
up_sampling2d (UpSampling2D)	(None, 64, 64, 256)	0	dropout_1[0][0]
conv2d_10 (Conv2D)	(None, 64, 64, 128)	131200	up_sampling2d[0][0]
concatenate (Concatenate)	(None, 64, 64, 256)	0	dropout[0][0] & conv2d_10[0][0]
conv2d_11 (Conv2D)	(None, 64, 64, 128)	295040	concatenate[0][0]
conv2d_12	(Conv2D)	(None, 64, 64, 128)	147584
up_sampling2d_1 (UpSampling2D)	(None, 128, 128, 128)	0	conv2d_12[0][0]
conv2d_13 (Conv2D)	(None, 128, 128, 64)	32832	up_sampling2d_1[0][0]
concatenate_1 (Concatenate)	(None, 128, 128, 128)	0	conv2d_5[0][0] & conv2d_13[0][0]
conv2d_14 (Conv2D)	(None, 128, 128, 64)	73792	concatenate_1[0][0]
conv2d_15 (Conv2D)	(None, 128, 128, 64)	36928	conv2d_14[0][0]
up_sampling2d_2 (UpSampling2D)	(None, 256, 256, 64)	0	conv2d_15[0][0]
conv2d_16 (Conv2D)	(None, 256, 256, 32)	8224	up_sampling2d_2[0][0]
concatenate_2 (Concatenate)	(None, 256, 256, 64)	0	conv2d_3[0][0] & conv2d_16[0][0]
conv2d_17 (Conv2D)	(None, 256, 256, 32)	18464	concatenate_2[0][0]
conv2d_18 (Conv2D)	(None, 256, 256, 32)	9248	conv2d_17[0][0]
up_sampling2d_3 (UpSampling2D)	(None, 512, 512, 32)	0	conv2d_18[0][0]
conv2d_19 (Conv2D)	(None, 512, 512, 16)	2064	up_sampling2d_3[0][0]
concatenate_3 (Concatenate)	(None, 512, 512, 32)	0	conv2d_1[0][0] & conv2d_19[0][0]
conv2d_20 (Conv2D)	(None, 512, 512, 16)	4624	concatenate_3[0][0]
conv2d_21 (Conv2D)	(None, 512, 512, 16)	2320	conv2d_20[0][0]
conv2d_22 (Conv2D)	(None, 512, 512, 2)	290	conv2d_21[0][0]
conv2d_23 (Conv2D)	(None, 512, 512, 1)	3	conv2d_22[0][0]

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 :)