ML Community Day is November 9! Join us for updates from TensorFlow, JAX, and more Learn more

Image segmentation

View on View source on GitHub Download notebook

This tutorial focuses on the task of image segmentation, using a modified U-Net.

What is image segmentation?

In an image classification task the network assigns a label (or class) to each input image. However, suppose you want to know the shape of that object, which pixel belongs to which object, etc. In this case you will want to assign a class to each pixel of the image. This task is known as segmentation. A segmentation model returns much more detailed information about the image. Image segmentation has many applications in medical imaging, self-driving cars and satellite imaging to name a few.

This tutorial uses the Oxford-IIIT Pet Dataset (Parkhi et al, 2012). The dataset consists of images of 37 pet breeds, with 200 images per breed (~100 each in the training and test splits). Each image includes the corresponding labels, and pixel-wise masks. The masks are class-labels for each pixel. Each pixel is given one of three categories:

  • Class 1: Pixel belonging to the pet.
  • Class 2: Pixel bordering the pet.
  • Class 3: None of the above/a surrounding pixel.
pip install git+
import tensorflow as tf

import tensorflow_datasets as tfds
from tensorflow_examples.models.pix2pix import pix2pix

from IPython.display import clear_output
import matplotlib.pyplot as plt

Download the Oxford-IIIT Pets dataset

The dataset is available from TensorFlow Datasets. The segmentation masks are included in version 3+.

dataset, info = tfds.load('oxford_iiit_pet:3.*.*', with_info=True)

In addition, the image color values are normalized to the [0,1] range. Finally, as mentioned above the pixels in the segmentation mask are labeled either {1, 2, 3}. For the sake of convenience, subtract 1 from the segmentation mask, resulting in labels that are : {0, 1, 2}.

def normalize(input_image, input_mask):
  input_image = tf.cast(input_image, tf.float32) / 255.0
  input_mask -= 1
  return input_image, input_mask
def load_image(datapoint):
  input_image = tf.image.resize(datapoint['image'], (128, 128))
  input_mask = tf.image.resize(datapoint['segmentation_mask'], (128, 128))

  input_image, input_mask = normalize(input_image, input_mask)

  return input_image, input_mask

The dataset already contains the required training and test splits, so continue to use the same splits.

TRAIN_LENGTH = info.splits['train'].num_examples
train_images = dataset['train'].map(load_image,
test_images = dataset['test'].map(load_image,

The following class performs a simple augmentation by randomly-flipping an image. Go to the Image augmentation tutorial to learn more.

class Augment(tf.keras.layers.Layer):
  def __init__(self, seed=42):
    # both use the same seed, so they'll make the same random changes.
    self.augment_inputs = tf.keras.layers.RandomFlip(mode="horizontal", seed=seed)
    self.augment_labels = tf.keras.layers.RandomFlip(mode="horizontal", seed=seed)

  def call(self, inputs, labels):
    inputs = self.augment_inputs(inputs)
    labels = self.augment_labels(labels)
    return inputs, labels

Build the input pipeline, applying the Augmentation after batching the inputs.

train_batches = (

test_batches = test_images.batch(BATCH_SIZE)

Visualize an image example and its corresponding mask from the dataset.

def display(display_list):
  plt.figure(figsize=(15, 15))

  title = ['Input Image', 'True Mask', 'Predicted Mask']

  for i in range(len(display_list)):
    plt.subplot(1, len(display_list), i+1)
for images, masks in train_batches.take(2):
  sample_image, sample_mask = images[0], masks[0]
  display([sample_image, sample_mask])
Corrupt JPEG data: 240 extraneous bytes before marker 0xd9
Corrupt JPEG data: premature end of data segment



2021-10-27 01:31:30.714975: W tensorflow/core/kernels/data/] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset  will be discarded. This can happen if you have an input pipeline similar to `dataset.cache().take(k).repeat()`. You should use `dataset.take(k).cache().repeat()` instead.

Define the model

The model being used here is a modified U-Net. A U-Net consists of an encoder (downsampler) and decoder (upsampler). In-order to learn robust features and reduce the number of trainable parameters, you will use a pretrained model - MobileNetV2 - as the encoder. For the decoder, you will use the upsample block, which is already implemented in the pix2pix example in the TensorFlow Examples repo. (Check out the pix2pix: Image-to-image translation with a conditional GAN tutorial in a notebook.)

As mentioned, the encoder will be a pretrained MobileNetV2 model which is prepared and ready to use in tf.keras.applications. The encoder consists of specific outputs from intermediate layers in the model. Note that the encoder will not be trained during the training process.

base_model = tf.keras.applications.MobileNetV2(input_shape=[128, 128, 3], include_top=False)

# Use the activations of these layers
layer_names = [
    'block_1_expand_relu',   # 64x64
    'block_3_expand_relu',   # 32x32
    'block_6_expand_relu',   # 16x16
    'block_13_expand_relu',  # 8x8
    'block_16_project',      # 4x4
base_model_outputs = [base_model.get_layer(name).output for name in layer_names]

# Create the feature extraction model
down_stack = tf.keras.Model(inputs=base_model.input, outputs=base_model_outputs)

down_stack.trainable = False
Downloading data from
9412608/9406464 [==============================] - 0s 0us/step
9420800/9406464 [==============================] - 0s 0us/step

The decoder/upsampler is simply a series of upsample blocks implemented in TensorFlow examples.

up_stack = [
    pix2pix.upsample(512, 3),  # 4x4 -> 8x8
    pix2pix.upsample(256, 3),  # 8x8 -> 16x16
    pix2pix.upsample(128, 3),  # 16x16 -> 32x32
    pix2pix.upsample(64, 3),   # 32x32 -> 64x64
def unet_model(output_channels:int):
  inputs = tf.keras.layers.Input(shape=[128, 128, 3])

  # Downsampling through the model
  skips = down_stack(inputs)
  x = skips[-1]
  skips = reversed(skips[:-1])

  # Upsampling and establishing the skip connections
  for up, skip in zip(up_stack, skips):
    x = up(x)
    concat = tf.keras.layers.Concatenate()
    x = concat([x, skip])

  # This is the last layer of the model
  last = tf.keras.layers.Conv2DTranspose(
      filters=output_channels, kernel_size=3, strides=2,
      padding='same')  #64x64 -> 128x128

  x = last(x)

  return tf.keras.Model(inputs=inputs, outputs=x)

Note that the number of filters on the last layer is set to the number of output_channels. This will be one output channel per class.

Train the model

Now, all that is left to do is to compile and train the model.

Since this is a multiclass classification problem, use the tf.keras.losses.CategoricalCrossentropy loss function with the from_logits argument set to True, since the labels are scalar integers instead of vectors of scores for each pixel of every class.

When running inference, the label assigned to the pixel is the channel with the highest value. This is what the create_mask function is doing.


model = unet_model(output_channels=OUTPUT_CLASSES)

Have a quick look at the resulting model architecture:

tf.keras.utils.plot_model(model, show_shapes=True)


Try out the model to check what it predicts before training.

def create_mask(pred_mask):
  pred_mask = tf.argmax(pred_mask, axis=-1)
  pred_mask = pred_mask[..., tf.newaxis]
  return pred_mask[0]
def show_predictions(dataset=None, num=1):
  if dataset:
    for image, mask in dataset.take(num):
      pred_mask = model.predict(image)
      display([image[0], mask[0], create_mask(pred_mask)])
    display([sample_image, sample_mask,
             create_mask(model.predict(sample_image[tf.newaxis, ...]))])


The callback defined below is used to observe how the model improves while it is training.

class DisplayCallback(tf.keras.callbacks.Callback):
  def on_epoch_end(self, epoch, logs=None):
    print ('\nSample Prediction after epoch {}\n'.format(epoch+1))
VALIDATION_STEPS = info.splits['test'].num_examples//BATCH_SIZE//VAL_SUBSPLITS

model_history =, epochs=EPOCHS,


Sample Prediction after epoch 20
loss = model_history.history['loss']
val_loss = model_history.history['val_loss']

plt.plot(model_history.epoch, loss, 'r', label='Training loss')
plt.plot(model_history.epoch, val_loss, 'bo', label='Validation loss')
plt.title('Training and Validation Loss')
plt.ylabel('Loss Value')
plt.ylim([0, 1])


Make predictions

Now, make some predictions. In the interest of saving time, the number of epochs was kept small, but you may set this higher to achieve more accurate results.

show_predictions(test_batches, 3)




Optional: Imbalanced classes and class weights

Semantic segmentation datasets can be highly imbalanced meaning that particular class pixels can be present more inside images than that of other classes. Since segmentation problems can be treated as per-pixel classification problems, you can deal with the imbalance problem by weighing the loss function to account for this. It's a simple and elegant way to deal with this problem. Refer to the Classification on imbalanced data tutorial to learn more.

To avoid ambiguity, does not support the class_weight argument for inputs with 3+ dimensions.

  model_history =, epochs=EPOCHS,
                            class_weight = {0:2.0, 1:2.0, 2:1.0})
  assert False
except Exception as e:
  print(f"Expected {type(e).__name__}: {e}")
Expected ValueError: `class_weight` not supported for 3+ dimensional targets.

So, in this case you need to implement the weighting yourself. You'll do this using sample weights: In addition to (data, label) pairs, also accepts (data, label, sample_weight) triples. propagates the sample_weight to the losses and metrics, which also accept a sample_weight argument. The sample weight is multiplied by the sample's value before the reduction step. For example:

label = [0,0]
prediction = [[-3., 0], [-3, 0]] 
sample_weight = [1, 10] 

loss = tf.losses.SparseCategoricalCrossentropy(from_logits=True,
loss(label, prediction, sample_weight).numpy()
array([ 3.0485873, 30.485874 ], dtype=float32)

So to make sample weights for this tutorial you need a function that takes a (data, label) pair and returns a (data, label, sample_weight) triple. Where the sample_weight is a 1-channel image containing the class weight for each pixel.

The simplest possible implementation is to use the label as an index into a class_weight list:

def add_sample_weights(image, label):
  # The weights for each class, with the constraint that:
  #     sum(class_weights) == 1.0
  class_weights = tf.constant([2.0, 2.0, 1.0])
  class_weights = class_weights/tf.reduce_sum(class_weights)

  # Create an image of `sample_weights` by using the label at each pixel as an 
  # index into the `class weights` .
  sample_weights = tf.gather(class_weights, indices=tf.cast(label, tf.int32))

  return image, label, sample_weights

The resulting dataset elements contain 3 images each:
(TensorSpec(shape=(None, 128, 128, 3), dtype=tf.float32, name=None),
 TensorSpec(shape=(None, 128, 128, 1), dtype=tf.float32, name=None),
 TensorSpec(shape=(None, 128, 128, 1), dtype=tf.float32, name=None))

Now you can train a model on this weighted dataset:

weighted_model = unet_model(OUTPUT_CLASSES)
10/10 [==============================] - 3s 42ms/step - loss: 0.3264 - accuracy: 0.5658
<keras.callbacks.History at 0x7f75d076b810>

Next steps

Now that you have an understanding of what image segmentation is and how it works, you can try this tutorial out with different intermediate layer outputs, or even different pretrained models. You may also challenge yourself by trying out the Carvana image masking challenge hosted on Kaggle.

You may also want to see the Tensorflow Object Detection API for another model you can retrain on your own data. Pretrained models are available on TensorFlow Hub