Image Segmentation Keras : Implementation of Segnet, FCN, UNet, PSPNet and other models in Keras.

Implementation of various Deep Image Segmentation models in keras.

Link to the full blog post with tutorial : https://divamgupta.com/image-segmentation/2019/06/06/deep-learning-semantic-segmentation-keras.html

Working Google Colab Examples:

Python Interface: https://colab.research.google.com/drive/1q_eCYEzKxixpCKH1YDsLnsvgxl92ORcv?usp=sharing
CLI Interface: https://colab.research.google.com/drive/1Kpy4QGFZ2ZHm69mPfkmLSUes8kj6Bjyi?usp=sharing

Our Other Repositories

Attention based Language Translation in Keras
Ladder Network in Keras model achives 98% test accuracy on MNIST with just 100 labeled examples

Top Contributors

Models

Following models are supported:

model_name	Base Model	Segmentation Model
fcn_8	Vanilla CNN	FCN8
fcn_32	Vanilla CNN	FCN8
fcn_8_vgg	VGG 16	FCN8
fcn_32_vgg	VGG 16	FCN32
fcn_8_resnet50	Resnet-50	FCN32
fcn_32_resnet50	Resnet-50	FCN32
fcn_8_mobilenet	MobileNet	FCN32
fcn_32_mobilenet	MobileNet	FCN32
pspnet	Vanilla CNN	PSPNet
vgg_pspnet	VGG 16	PSPNet
resnet50_pspnet	Resnet-50	PSPNet
unet_mini	Vanilla Mini CNN	U-Net
unet	Vanilla CNN	U-Net
vgg_unet	VGG 16	U-Net
resnet50_unet	Resnet-50	U-Net
mobilenet_unet	MobileNet	U-Net
segnet	Vanilla CNN	Segnet
vgg_segnet	VGG 16	Segnet
resnet50_segnet	Resnet-50	Segnet
mobilenet_segnet	MobileNet	Segnet

Example results for the pre-trained models provided :

Input Image	Output Segmentation Image

Getting Started

Prerequisites

Keras ( recommended version : 2.4.3 )
OpenCV for Python
Tensorflow ( recommended version : 2.4.1 )

apt-get install -y libsm6 libxext6 libxrender-dev
pip install opencv-python

Installing

Install the module

pip install keras-segmentation

pip install --upgrade git+https://github.com/divamgupta/image-segmentation-keras

or

git clone https://github.com/divamgupta/image-segmentation-keras
cd image-segmentation-keras
python setup.py install

Pre-trained models:

from keras_segmentation.pretrained import pspnet_50_ADE_20K , pspnet_101_cityscapes, pspnet_101_voc12

model = pspnet_50_ADE_20K() # load the pretrained model trained on ADE20k dataset

model = pspnet_101_cityscapes() # load the pretrained model trained on Cityscapes dataset

model = pspnet_101_voc12() # load the pretrained model trained on Pascal VOC 2012 dataset

# load any of the 3 pretrained models

out = model.predict_segmentation(
    inp="input_image.jpg",
    out_fname="out.png"
)

Preparing the data for training

You need to make two folders

Images Folder - For all the training images
Annotations Folder - For the corresponding ground truth segmentation images

The filenames of the annotation images should be same as the filenames of the RGB images.

The size of the annotation image for the corresponding RGB image should be same.

For each pixel in the RGB image, the class label of that pixel in the annotation image would be the value of the blue pixel.

Example code to generate annotation images :

import cv2
import numpy as np

ann_img = np.zeros((30,30,3)).astype('uint8')
ann_img[ 3 , 4 ] = 1 # this would set the label of pixel 3,4 as 1

cv2.imwrite( "ann_1.png" ,ann_img )

Only use bmp or png format for the annotation images.

Download the sample prepared dataset

Download and extract the following:

https://drive.google.com/file/d/0B0d9ZiqAgFkiOHR1NTJhWVJMNEU/view?usp=sharing

You will get a folder named dataset1/

Using the python module

You can import keras_segmentation in your python script and use the API

from keras_segmentation.models.unet import vgg_unet

model = vgg_unet(n_classes=51 ,  input_height=416, input_width=608  )

model.train(
    train_images =  "dataset1/images_prepped_train/",
    train_annotations = "dataset1/annotations_prepped_train/",
    checkpoints_path = "/tmp/vgg_unet_1" , epochs=5
)

out = model.predict_segmentation(
    inp="dataset1/images_prepped_test/0016E5_07965.png",
    out_fname="/tmp/out.png"
)

import matplotlib.pyplot as plt
plt.imshow(out)

# evaluating the model 
print(model.evaluate_segmentation( inp_images_dir="dataset1/images_prepped_test/"  , annotations_dir="dataset1/annotations_prepped_test/" ) )

Usage via command line

You can also use the tool just using command line

Visualizing the prepared data

You can also visualize your prepared annotations for verification of the prepared data.

python -m keras_segmentation verify_dataset \
 --images_path="dataset1/images_prepped_train/" \
 --segs_path="dataset1/annotations_prepped_train/"  \
 --n_classes=50

python -m keras_segmentation visualize_dataset \
 --images_path="dataset1/images_prepped_train/" \
 --segs_path="dataset1/annotations_prepped_train/"  \
 --n_classes=50

Training the Model

To train the model run the following command:

python -m keras_segmentation train \
 --checkpoints_path="path_to_checkpoints" \
 --train_images="dataset1/images_prepped_train/" \
 --train_annotations="dataset1/annotations_prepped_train/" \
 --val_images="dataset1/images_prepped_test/" \
 --val_annotations="dataset1/annotations_prepped_test/" \
 --n_classes=50 \
 --input_height=320 \
 --input_width=640 \
 --model_name="vgg_unet"

Choose model_name from the table above

Getting the predictions

To get the predictions of a trained model

python -m keras_segmentation predict \
 --checkpoints_path="path_to_checkpoints" \
 --input_path="dataset1/images_prepped_test/" \
 --output_path="path_to_predictions"

Video inference

To get predictions of a video

python -m keras_segmentation predict_video \
 --checkpoints_path="path_to_checkpoints" \
 --input="path_to_video" \
 --output_file="path_for_save_inferenced_video" \
 --display

If you want to make predictions on your webcam, don't use --input, or pass your device number: --input 0
--display opens a window with the predicted video. Remove this argument when using a headless system.

Model Evaluation

To get the IoU scores

python -m keras_segmentation evaluate_model \
 --checkpoints_path="path_to_checkpoints" \
 --images_path="dataset1/images_prepped_test/" \
 --segs_path="dataset1/annotations_prepped_test/"

Fine-tuning from existing segmentation model

The following example shows how to fine-tune a model with 10 classes .

from keras_segmentation.models.model_utils import transfer_weights
from keras_segmentation.pretrained import pspnet_50_ADE_20K
from keras_segmentation.models.pspnet import pspnet_50

pretrained_model = pspnet_50_ADE_20K()

new_model = pspnet_50( n_classes=51 )

transfer_weights( pretrained_model , new_model  ) # transfer weights from pre-trained model to your model

new_model.train(
    train_images =  "dataset1/images_prepped_train/",
    train_annotations = "dataset1/annotations_prepped_train/",
    checkpoints_path = "/tmp/vgg_unet_1" , epochs=5
)

Knowledge distillation for compressing the model

The following example shows transfer the knowledge from a larger ( and more accurate ) model to a smaller model. In most cases the smaller model trained via knowledge distilation is more accurate compared to the same model trained using vanilla supervised learning.

from keras_segmentation.predict import model_from_checkpoint_path
from keras_segmentation.models.unet import unet_mini
from keras_segmentation.model_compression import perform_distilation

model_large = model_from_checkpoint_path( "/checkpoints/path/of/trained/model" )
model_small = unet_mini( n_classes=51, input_height=300, input_width=400  )

perform_distilation ( data_path="/path/to/large_image_set/" , checkpoints_path="path/to/save/checkpoints" , 
    teacher_model=model_large ,  student_model=model_small  , distilation_loss='kl' , feats_distilation_loss='pa' )

Adding custom augmentation function to training

The following example shows how to define a custom augmentation function for training.

from keras_segmentation.models.unet import vgg_unet
from imgaug import augmenters as iaa

def custom_augmentation():
    return  iaa.Sequential(
        [
            # apply the following augmenters to most images
            iaa.Fliplr(0.5),  # horizontally flip 50% of all images
            iaa.Flipud(0.5), # horizontally flip 50% of all images
        ])

model = vgg_unet(n_classes=51 ,  input_height=416, input_width=608)

model.train(
    train_images =  "dataset1/images_prepped_train/",
    train_annotations = "dataset1/annotations_prepped_train/",
    checkpoints_path = "/tmp/vgg_unet_1" , epochs=5, 
    do_augment=True, # enable augmentation 
    custom_augmentation=custom_augmentation # sets the augmention function to use
)

Custom number of input channels

The following example shows how to set the number of input channels.

from keras_segmentation.models.unet import vgg_unet

model = vgg_unet(n_classes=51 ,  input_height=416, input_width=608, 
                 channels=1 # Sets the number of input channels
                 )

model.train(
    train_images =  "dataset1/images_prepped_train/",
    train_annotations = "dataset1/annotations_prepped_train/",
    checkpoints_path = "/tmp/vgg_unet_1" , epochs=5, 
    read_image_type=0  # Sets how opencv will read the images
                       # cv2.IMREAD_COLOR = 1 (rgb),
                       # cv2.IMREAD_GRAYSCALE = 0,
                       # cv2.IMREAD_UNCHANGED = -1 (4 channels like RGBA)
)

Custom preprocessing

The following example shows how to set a custom image preprocessing function.

from keras_segmentation.models.unet import vgg_unet

def image_preprocessing(image):
    return image + 1

model = vgg_unet(n_classes=51 ,  input_height=416, input_width=608)

model.train(
    train_images =  "dataset1/images_prepped_train/",
    train_annotations = "dataset1/annotations_prepped_train/",
    checkpoints_path = "/tmp/vgg_unet_1" , epochs=5,
    preprocessing=image_preprocessing # Sets the preprocessing function
)

Custom callbacks

The following example shows how to set custom callbacks for the model training.

from keras_segmentation.models.unet import vgg_unet
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping

model = vgg_unet(n_classes=51 ,  input_height=416, input_width=608 )

# When using custom callbacks, the default checkpoint saver is removed
callbacks = [
    ModelCheckpoint(
                filepath="checkpoints/" + model.name + ".{epoch:05d}",
                save_weights_only=True,
                verbose=True
            ),
    EarlyStopping()
]

model.train(
    train_images =  "dataset1/images_prepped_train/",
    train_annotations = "dataset1/annotations_prepped_train/",
    checkpoints_path = "/tmp/vgg_unet_1" , epochs=5,
    callbacks=callbacks
)

Multi input image input

The following example shows how to add additional image inputs for models.

from keras_segmentation.models.unet import vgg_unet

model = vgg_unet(n_classes=51 ,  input_height=416, input_width=608)

model.train(
    train_images =  "dataset1/images_prepped_train/",
    train_annotations = "dataset1/annotations_prepped_train/",
    checkpoints_path = "/tmp/vgg_unet_1" , epochs=5,
    other_inputs_paths=[
        "/path/to/other/directory"
    ],
    
    
#     Ability to add preprocessing
    preprocessing=[lambda x: x+1, lambda x: x+2, lambda x: x+3], # Different prepocessing for each input
#     OR
    preprocessing=lambda x: x+1, # Same preprocessing for each input
)

Projects using keras-segmentation

Here are a few projects which are using our library :

If you use our code in a publicly available project, please add the link here ( by posting an issue or creating a PR )

psk1998/image-segmentation-keras