Compressing images for this website

2019-07-05

I wanted my webpages to load faster, and I was somewhat inspired by this website which runs entirely on a solar powered raspberry pi. They needed to minimize the amount of data transfer to keep energy usage down. I was also inspired a bit by this blog post on their image dithering procedure which converts images to black and white while preserving the essence of what the image is trying to portray. Unfortunately they do their image conversion in Adobe Photoshop, which I didn’t want to use. Instead I decided to see if I could replicate their conversion process and achieve highly compressed dithered images using ImageMagick , which could be written into a reproducible script that I could run on a set of images. I also wanted to see if there was a better method of compressing the images.

ImageMagick runs on the command line and the command I use is convert which is designed to take an image, alter it in a multitude of ways, then save to a new file. mogrify is better for overwriting the same image.

I have an example image here which is currently 6.8 MB.

y first pass at the image is simply to reduce the -quality of the image, which takes a percentage value that pertains to the compression level.

convert bike.jpg -quality 90% output_1.jpg

This produces an image of 3.3 MB.

Here is a plot of the file size as the -quality percentage is reduced. and a table of file sizes. Note that 100% compression on the original image actually made the original image of 6.8 MB larger, to 7.8 MB.

Plot of comperession percentage and image size
Compression (%) Size (MB)
100 7.799
95 4.479
90 3.229
85 2.6
80 2.24
75 1.973
70 1.808
65 1.658
60 1.531
55 1.434
50 1.349
45 1.268
40 1.172
35 1.086
30 0.9793
25 0.8683
20 0.7416
15 0.5998
10 0.4374
5 0.2395
—————– ————

It’s interesting that even at 5% -quality, the image is still not bad to look at.

For the next part I’ll take an arbitrary -quality value of 50%, meaning the base image is 1.349 MB.

Converting the image to grayscale makes the file slightly smaller, but not much

convert bike.jpg -quality 50% -colorspace gray output_21.jpg

The original image was 1,414,879 bytes (1.349 MB) and the grayscale image is 1,325,856 (1.264 MB), a difference of 89,023 bytes, or 87 KB.

I also experimented with other options for reducing the colors, but none of these made any positive difference. -posterize 4 = 1.781 MB, -quantize gray = 1.349 MB.

Next I want to dither the image to hopefully save even more file size. I’ll use the grayscale image I generated earlier, which is 1.264 MB. There are multiple options for dithering so here is a comparison.

convert bike.jpg -quality 50% -colorspace gray -dither floyd-steinberg output_25.jpg

Both methods produced files of exactly the same size as the grascale image so I think I’m doing something wrong. I think I should apply some sort of dithering color reduction flag as well. Here are examples with different color reduction methods:

convert bike.jpg -quality 50% -colorspace gray -dither floyd-steinberg -monochrome output_26.jpg

convert bike.jpg -quality 50% -colorspace gray -dither floyd-steinberg -posterize 10 output_27.jpg

The -monochrome option produced a file of 2.812 MB and the -posterize 10 option produced a file of 1.428 MB.

Posterize is an interesting one to play with. Here is a comparison of -posterize with different numbers of colours:

Size (MB) Number of colours
1.428 10
1.459 9
1.487 8
1.532 7
1.578 6
1.631 5
1.674 4
1.793 3
1.437 2
0.06754 1
———– ——————-

Weirdly, the file size actually increases as the number of colours decreases, up until the number of colours is two.

There are also predefined dithering threshold maps that can be used with -ordered-dither:

convert bike.jpg -quality 50% -colorspace gray -ordered-dither c7x7w output_43.jpg

Here is a comparison of these methods:

Plot of ordered dither method and output size
Code Description Size (MB)
checks Checkerboard 2x1 (dither) 2.678
o2x2 Ordered 2x2 (dispersed) 3.702
o3x3 Ordered 3x3 (dispersed) 4.445
o4x4 Ordered 4x4 (dispersed) 4.557
o8x8 Ordered 8x8 (dispersed) 4.675
h6x6a Halftone 6x6 (angled) 4.824
h8x8a Halftone 8x8 (angled) 1.264
h6x6o Halftone 6x6 (orthogonal) 4.877
h8x8o Halftone 8x8 (orthogonal) 4.831
c5x5b Circles 5x5 (black) 4.879
c5x5w Circles 5x5 (white) 4.879
c6x6b Circles 6x6 (black) 4.831
c6x6w Circles 6x6 (white) 4.823
c7x7b Circles 7x7 (black) 4.915
c7x7w Circles 7x7 (white) 4.899
——– ————————— ———–

So the h8x8a method produced the smallest file size of 1.264 MB, but this is only as small as the undithered grayscale image!

So from my little test so far it seems that the best way of compressing images is to use a grayscale undithered image an adjust the -quality of the image to my liking.

Here are a few other options to reduce file size I came across. Here is a comparison using the undithered grayscale image.

Removing EXIF data, which I would probably want to do regardless of the file size, just for privacy reasons:

convert bike.jpg -quality 50% -colorspace gray -strip output_54.jpg

This reduces the file size from 1.264 MB to 1.242 MB, a difference of 23.2 KB.

Adding some Gaussian blur to reduce the detail level:

convert bike.jpg -quality 50% -colorspace gray -strip -gaussian-blur 0.05 output_54.jpg

This reduces the stripped grayscale image from 1.242 MB to 1.031 MB, a difference of 215 KB.

Here is a comparison of images converted with different levels of gaussian blur.

One thing I did find was that the computation time increased with the radius of the blur.

Plot of gaussian blur and output size
Blur Size (MB)
0.05 1.031
0.1 1.031
0.2 1.031
0.3 1.031
0.4 1.031
0.5 1.031
0.6 1.031
0.7 1.031
0.8 1.031
0.9 1.031
1 1.031
2 0.9472
3 0.9377
4 0.9373
5 0.9373
10 0.9373
20 0.9373
30 0.9373
40 0.9373
50 0.9373
—— ———–

So, finally, the optimal way of compressing my JPEG images seems to be:

convert bike.jpg -quality 20% -colorspace gray -strip -gaussian-blur 3.5 output.jpg

This leads to an image of 538 KB. It looks good, except when zooming in, then it gets a bit hairy, this is where dithering might become useful if I can find a way of dithering while maintaining the small size of the image.

Another thing I learned from reading the comments on this website was that JPEG is the most appropriate format for reducing image file size as it is lossy, unlike PNG, so I can convert all my PNG images to JPEG using:

convert test.png -background white -flatten output.jpg

-background white -flatten is needed to allow the JPEG to properly display the transparent areas of the JPG as white background.

I have a folder on my website with images sorted into directories based on what blog post they refer to. I wrote this shell script to prepare the images to be put onto the website:

#!/bin/bash

shopt -s nullglob

# For each image
for i in img_full/*/*.{jpg,JPG,jpeg,png,PNG}; do

	# If the directory in the compressed images directory doesn't exist, create it
	new_dir=$(dirname ${i} | sed "s/img_full/img/g")

	if [ ! -d $new_dir ]; then
  		mkdir -p $new_dir;
	fi

	echo $dir

	# Create new file path for output
	new_path=$(echo "${i}" | sed "s/img_full/img/g")

	echo $new_path

	# Convert image files
	if [ ! -f $new_path ]; then
		convert $i -quality 20% -colorspace gray -strip -background white -flatten -gaussian-blur 3.5 $new_path 
	fi

done

# Maintain full size 404 image 
cp img_full/404/404.png img/404/404.png