Problem with PNG Image (grayscale): 12 bits from 16 bits

Hi,

I´m working in a project to get images from DDSM ( http://marathon.csee.usf.edu/Mammography/Database.html). I´ve got the step to convert (LJPEG to LJPEG.1) and now I´m with the step to convert LJPEG.1 to PNG, but the problem is the PNG image must be of 12 bits (because the original image of DDSM has 12 bits).

I know it´s not possible to obtain directly a 12 bit image from PNG, so I´ve used 2 ways:

1. Divide the matrix by 16. e.g. I4=I3/16;
2. Shift the image 4 positions. e.g. I5=bitshift (I3,-4);

With this two options, I get the same result, a matrix with max(I5(:)) = 4095, so the process is OK and I could see it in a figure, but the problem comes when I want to save in a png file, because the image must be saved in 16-bit and get me a dark image with a big size (four times the 8 bit png image). I also checked the value of bits: 13, 14, 15 and 16 and their values are 0.

I´ve used the PNG option of 'SignificantBits' with 12, but I obtained the same result.

Any ideas to solve the problem? I´ve found many articles about it but any of them gives a solution.

Thanks & Regards

Alejo

[Image Analyst]

I believe Ashish Uthama is the "image format guy" at the Mathworks. Try calling him if he doesn't answer here. I think he might hang out at the newsgroup more than there so try there also. Steve Eddins also has been willing to answer format questions on occasion in the past.

[Cheng Zhao]

This may be of great interest to you! Someone has already done it.

http://www.irma-project.org/datasets_en.php?SELECTED=00010#00010.dataset

Answer 1

[Walter Roberson]

Do not downscale the data: it must be full-range for the bitdepth (16). Use 'SignificantBits' of 12.

Grayscale, 16 bits per channel, 12 significant bits, is the way that PNG represents 12-bit grayscale images. Decoders are allowed to ignore the 12 bit specification

See 4.2.4.3. sBIT Significant bits in http://www.libpng.org/pub/png/spec/1.2/PNG-Chunks.html

[Alejo]

Walter, thanks for the response, but I continue with the problem.

My code is:

fid = fopen(filename,'r','ieee-be'); 
DDSM_img=fread(fid,[image_size(1) image_size(2)],'uint16');
DDSM_img=DDSM_img-min(DDSM_img(:)); % shift data
DDSM_img=DDSM_img/max(DDSM_img(:)); % normalize the shifted data to 1 
imwrite(DDSM_img,'Temp.png','bitdepth',16,'SignificantBits',12);
%%With this 'imwrite' with 'SignificantBits' the last four bits have values.

On the other hand, the size of 16 Bits in PNG is 25.541 KB, whereas in 8 Bit is 7.872 KB, it´s not correct, so I´ve applied the 2 ways reported before:

Divide the matrix by 16. e.g. I4=I3/16; Shift the image 4 positions. e.g. I5=bitshift (I3,-4); %%The last four bits are empty, so this is OK.

However, with those matrix (I4 or I5), I tried to save in 16 Bits PNG Image, but I've obtained a dark image with a big size (30.000 KB aprox.)... it´s in this point, where I have the problem. I think I should obtain a file of 4000 KB (1/2 size of 8 Bits).

Regards.

[Walter Roberson]

Please test class(DDSM_img) after the data manipulations to be sure that it is still uint16()

[Alejo]

The test class for the output is double (uint16() for input), because I´ve found problems if I choose uint16(): DDSM_img=fread(fid,[image_size(1) image_size(2)],'*uint16');

The result is a black image of 105 Kb and I can´t work with it.

[Walter Roberson]

See this part of the imwrite() documentation:

If the input array is of class double, and the image is a grayscale or RGB color image, imwrite assumes the dynamic range is [0,1] and automatically scales the data by 255 before writing it to the file as 8-bit values.

[Alejo]

I´ve seen this part, but I´m not able to obtain good results with uint16 in the output and I don´t know why. Anyway, I´ve tested with other values: ubit12, ubit16 with bad results.

My specifications are: LEFT_CC LINES 5872 PIXELS_PER_LINE 3696 BITS_PER_PIXEL 12 RESOLUTION 50 OVERLAY

The other problem ocurrs when I obtain the image with 12 bits and I want to save it as 16-bit png the max value is 4096 and I don´t know how to indicate that this (4095) is the max value and not 65536 (16 bits), for this reason my result is a black image.

I can see (with imshow()) the matrix with max value= 4095 and it´s OK, but I´m not able to save as 12 bit png image.

[Alejo]

Hi,

Finally, I´ve got an output image of uint16 and its maximum value is 4095, so that´s right, but now my problem is when I tried to save this image with imwrite.

If I use 'imshow', it works fine, but with imwrite: imwrite(A,'myimage.png', 'BitDepth', 16,'SignificantBits',12);

I obtain a black image with wrong gray levels, I´ve tried to use the command 'imadjust', but it doesn´t work properly, because the result is a "white" image with with wrong gray levels.

Do you know how I can change the gray leves (with max: 4095 (12 bits), not 65536 (16 bits)) with imwrite or another way to get it?

Thanks!!

[Deyaa]

I have the same problem here trying to read DDSM database images, can you provide me with your code solution ?

Answer 2

[Anmol Sharma]

I have written a small utility to easily download, convert, view and get annotations from DDSM data, without the need for any manual work. You just need to download the required cases and then follow the instructions on how to use my utility to convert the LJPEG files to LJPEG1 and then to other formats like PNG, JPEG, TIF, GIF etc. 

You can download my utility from here - https://github.com/trane293/DDSMUtility 

The detailed instructions along with images are given in the Tutorial PDF file.