Compress Images Using TensorFlow tf.nn.max_pool(): A Beginner Example

tf.nn.max_pool() function can implement a max pool operation on a input data, in this tutorial, we will introduce how to use it to compress an image.

To understand how to use tensorflow tf.nn.max_pool(), you can read the tutorial:

Understand TensorFlow tf.nn.max_pool(): Implement Max Pooling for Convolutional Network

Read an image using tensorflow

We read the image data using tensorflow first.

Here is an example:

import tensorflow as tf
import numpy as np
from PIL import Image

fpath = 'logo.png'
img = tf.read_file(fpath)

img_arr = tf.image.decode_png(img, channels=3) #rgb

with tf.Session() as sess:
    imgdata = sess.run(img_arr)
    print(imgdata.shape)

Run this code, we can get:

(248, 250, 3)

Which means logo.png image contains data with (248, 250, 3)

Here we use tensorflow to read an image, you also can use python pillow to read an image into a numpy array. Here is the tutorial:

Python Pillow Read Image to NumPy Array: A Step Guide – Python Pillow Tutorial

In order to make tf.nn.max_pool() can process image, we should reshape it.

size = tf.shape(img_arr)

img_4d = tf.cast(tf.reshape(img_arr, [1, size[0], size[1], 3]), tf.float32)

Then we can use tf.nn.max_pool() to process image.

Compress images using tf.nn.max_pool()

pool = tf.nn.max_pool(img_4d, [1, 5, 5, 1], [1, 1, 1, 1], 'SAME')

Here ksize = [1, 5, 5, 1]. We will get max value from a 5 * 5 matrix.

Then we can convert image matrix data to an image file

Save image

with tf.Session() as sess:
    imgdata, pool = sess.run([img_arr, pool])
    print(imgdata.shape)
    print(pool.shape)
    Image.fromarray(np.uint8(pool.reshape(pool.shape[1:4]))).save('logo-processed.png')

We can use python pillow Image.fromarray() function to save numpy array to a file.

Run this code, we can find:

The source logo.png is:

file size: 5,403 byte

We can use different ksize to compress this image.

ksize	image	file size	*heightwidht**
[1, 5, 5, 1]		3,377 byte	248 * 250
[1, 10, 10, 1]		3,207 byte	248 * 250
[1, 15, 15, 1]		3,021 byte	248 * 250