Generating thumbnails from a collection of pictures is an easy task and there are many tools for the job. But I didn’t find a tool who can clip out a centered box of my pictures to get quadratic thumbs of my originals. Because I am a developer and I dont want to try out many tools: I will hack a short python script (with help of the PIL) myself.

simple solution

The following python script can be find on many webpages. It will take all command line arguments as image filenames and convert it to 200×200 thumbnails with the default extension “_thumb.jpg” (build from the original filename):

import Image, os, sys

for filename in sys.argv[1:]:
    img = Image.open(filename).resize( (200,200) )
    out = file(os.path.splitext(filename)[0]+"_thumb.jpg", "w")
    try:
        img.save(out, "JPEG")
    finally:
        out.close()

With the command line tool find all images and converted:

find . -regex .*jpg | xargs python resize.py

But the result is not ok:

the only good part: the result image is 200×200 pixel
aspect ratio ignored
ugly quality
performance bad: 1.025 sec (example image on my eeeBox)
no value to choose the jpg quality for saving

For my web gallery I need identical thumbnails for a clean design and the aspect ratio should not be ignored!

Challenge: Find the biggest box with correct aspect ratio and clip the image, resize it fast and with the best quality to the given thumbnail size!

resize tips you should know

When choosing an image libary (with python is the build-in variant PIL good enough, but Samuel will add an article for java image libraries) for resizing images try to find out the following three parts:

choose the right algorythm

There should be a option to choose the image transforming algorithm. “Fast and ugly” or “slow and high quality”.

With the python imaging library there are included algorithm for image translations (ordered from fastest to best quality): NEAREST, BILINEAR, BICUBIC, ANTIALIAS. If you resize an image with NEAREST the quality will be poor but the difference in calculation time to ANTIALIAS is factor 3 (depending on images). The first example has the NEAREST as default value.

img = Image.open(filename).resize( (200,200), Image.ANTIALIAS)

Resizing the image with the best algorithm will produce a good quality, but needs more than 3 sec.

use two resize steps

The trick for resizing an 1600×1600 image to 200×200:

first resize it from 1600×1600 to double the target size (400×400) with the fastest allgorithm
resize the smaller image to the target resolution with the best algorithm

You get the fast transformation for the big amount and the last step to the target resolution with the best quality.

img = Image.open(filename).resize( (400,400) )
img = img.resize( (200,200), Image.ANTIALIAS)

This 2-pass resize part runs in exactly the same time (1.02 sec) like the first example with better quality.

resize vs. thumbnail

Because our picture collection contain 90% JPEG there is an other optimization step. To resize JPEGs you dont need all pixels for downsizing with factor 8. Its adequate to read only the starting point of an minimum coded Unit (MCU -> wikipedia). You have to try out if it is supported of your image lib – the transformation time is the answer.

The python image library offer this option with the Image.thumbnail-method. An other aspect is that the thumbnail parameter mean a target box size instead the target image size – so the aspect ratio will be saved.

img = Image.open(filename)
img.thumbnail( (200,200) )

The thumbnail method works on the image object and dont produce a new image object. Its faster: 0.25 sec.

img = Image.open(filename)
img.thumbnail( (400,400) )
img.thumbnail( (200,200), Image.ANTIALIAS)

The two-pass variant is slower but has better quality on the edges. Its runs 0.42 sec.

Finding the center box

I will construct two extreme examples for defining the clipping algorithm. If the two examples works the quadratic thumbnail work too.

Get from a portait picture a landscape thumb

Original: bmw.jpg (2048×3072 pixel)
thumbnail-box: 200×100 pixel

The clipping algorithm should find the centered box:

(0, 1024, 2048, 2048).

This picture is ideal for a wide thumbnail because the dominant radiator grill of the old bmw is THE part of the picture!

Get from a landscape picture a portait thumb

Second Example: ford.jpg (3072×2048 pixel)
thumbnail-box: 100×200 pixel

The clipping algorithm should find the centered box:

(1024, 515, 2048, 2562).

This old ford T-Model image has to many people on the image – a portrail thumbnail crop some from the front.

the complete function

def resize(img, box, fit, out):
    '''Downsample the image.
    @param img: Image -  an Image-object
    @param box: tuple(x, y) - the bounding box of the result image
    @param fix: boolean - crop the image to fill the box
    @param out: file-like-object - save the image into the output stream
    '''
    #preresize image with factor 2, 4, 8 and fast algorithm
    factor = 1
    while img.size[0]/factor > 2*box[0] and img.size[1]*2/factor > 2*box[1]:
        factor *=2
    if factor > 1:
        img.thumbnail((img.size[0]/factor, img.size[1]/factor), Image.NEAREST)

    #calculate the cropping box and get the cropped part
    if fit:
        x1 = y1 = 0
        x2, y2 = img.size
        wRatio = 1.0 * x2/box[0]
        hRatio = 1.0 * y2/box[1]
        if hRatio > wRatio:
            y1 = int(y2/2-box[1]*wRatio/2)
            y2 = int(y2/2+box[1]*wRatio/2)
        else:
            x1 = int(x2/2-box[0]*hRatio/2)
            x2 = int(x2/2+box[0]*hRatio/2)
        img = img.crop((x1,y1,x2,y2))

    #Resize the image with best quality algorithm ANTI-ALIAS
    img.thumbnail(box, Image.ANTIALIAS)

    #save it into a file-like object
    img.save(out, "JPEG", quality=75)
#resize

Additional tip: Cropping from the smaller image will save memory and memcopy-operation.

Line 13: fast pre-resize
Line 27: cropping the pre-sized image with the calculcated cropping values
Line 30: resize the image to the target box (aspecti ratio with using of method thumbnail is considerd).
Line 33: Saving a thumbnail with 75% is acceptable for the quality

Download area

the complete script with doctest and command line option is includet here in the article for free download.
the bmw and ford t model pictures are taken while the Leo Classic 2009 (Leonberg – Germany) and are for free use (download here).

How to do this in java

The questions for efficent and high quality resizing with java libs is open for the next article … be free to give improvements in the comments or trackbacks!

Bio
Latest Posts

Christian Harms

I am working as a software developer and project manager in germany. I have a strong background with python and javascript coding and design/develop web services with all aspects starting with http headers, scaling and fail-over issues. I use my knowledge as project manager and team motivator in the successful german automotive branche.

Latest posts by Christian Harms (see all)

google code jam 2013 – tic-tac-toe-Tomek solution - April 16, 2013
Google code jam 2013 – the lawnmower - April 14, 2013
code puzzles and permutations - April 11, 2013

united-coders.com

Image resizing tips general and for python