Why can't MATLAB resize an image nicely?

So I have an image with some text on it that I want to display in MATLAB for whatever reason, sounds easy enough, but it turns out that for some reason I hope to find out, MATLAB is apparently awful at rendering images to actually look nice.

I have a feeling there are performance reasons for this, but it seems like users should be able to resize an image with text on it and still actually read the image.

I have attached a sample image showing the original image, a highly distorted but still readable version of the image in Microsoft Word, and then an ever so slightly distorted version of the image in MATLAB which still looks awful.

If anyone can either explain how to fix this, or why its unfixable and why I should like it that way, I would really appreciate it!

[Shaun VanWeelden]

Here is the actual attachment! http://imgur.com/HbYMLX1

And here is the original image: http://www.antigrain.com/research/font_rasterization/msword_text_rendering.png

Answer 1

[Alex Taylor]

Would you please provide some code as to how you are displaying, resizing, and exporting the data from MATLAB so that we can better understand what you are trying to do?

One thing I would try is that I suspect you are using image/imagesc instead of imshow to display that data. If you aren't careful with the data aspect ratio in your axes, this will distort the text in display. If you cannot use imshow (IPT dependency), try this:

A = imread('http://www.antigrain.com/research/font_rasterization/msword_text_rendering.png');

image(A)
truesize

This will display your data at the original aspect ratio, which is definitely something you need when working with text.

[Shaun VanWeelden]

Hey Alex, great question about the code. Literally at this point it is just: img=imread('myImage.png') image(img)

and then I resize the figure

[Shaun VanWeelden]

I accepted your answer because I appreciate the help, I did kind of find a work around myself that delivers very high quality at the cost of performance. I did not know (or forgot) about image truesize though, very cool!

Answer 2

[Image Analyst]

Try using the 'nearest' option in imresize(). Though if a line is just one pixel in your image, it's either going to be there or not if you use nearest. If you use bicubic then you'll get partial intensity depending on how much of the output pixel lands on your input pixel.

[Shaun VanWeelden]

The default though is bicubic isn't it? I think 'nearest' would make it look absolutely awful honestly, much worse than bicubic. The entire idea of nearest is that it doesn't interp anything right?

I mean it seems like one possible workaround is use interp to make the image seem much bigger than it is and kind of smooth out the image when MATLAB awkwardly tried to render it.

I would like to avoid the image processing toolbox if possible also.

This seems like a renderer or axes property setting should fix this, but it looks bad in painters and only remotely better in opengl.

[Image Analyst]

Using nearest will give you sharper edges than bicubic, though at the cost of having jaggies on your edges and lines. Alex is the expert here so he would be able to help you more than me.

Answer 3

[Shaun VanWeelden]

A robust way I figured out how to do it was resize the image to 10x (or whatever you want, benefits seem to cut off at 8-10 times) its original size. When MATLAB puts the image into an axis of a different size than the original image (almost always), it looks MUCH MUCH better - granted this does kinda kill performance.

This would be very possible to do without the IPT, using interp2, its just not a simple one liner then.

Answer 4

[DGM]

For passers-by:

Normally, imshow() displays images using nearest-neighbor interpolation. That leads to thin features disappearing completely.

It wasn't available at the time of this post, but imshow() also supports bilinear display interpolation:

imshow(myimg,'interpolation','bilinear')