I've done a bit of fiddling around with PNGs over the last couple days and I am upset with my findings. I'm concluding that the majority of my results deal with compression. So this weekend I'm going to dive into advanced compression articles. I wanted to share my findings so far. To see if anyone has any advice on achieving my goal and to maybe point me in the right direction.
I am currently working on a project where I need to obtain the smallest possible file size within a window of less than 15 seconds.
The majority of the images I am working with are PNG-8bpp with a full 256 color palette. Most of these images I could represent accurately with 5bpp (32 colors).
PNG indexed however only supports 1,2,4, and 8bpp. So my idea was to strip the PNG format to the minimal information I needed and write an encoder/decoder to support IDAT sections with 3,5,6, or 7bpp.
Test 1:
Original File: 61.5KB, 750 * 500, 8pp Palette, 256 colors, No tRNS
After Optimizations (Reductions to 4bpp, Strip Anx Chunks, & PNGOUT): 49.2KB 4bpp, 16 Colors
Human Interpretation: I can see 6 distinguishable colors.
Since I only need six colors to represent the image I decide to encode the IDAT using 3bpp to give me a max palette of 8 colors. First I uncompressed the IDAT which results in a new file size of 368KB. After applying a 3bpp to IDAT my new uncompressed file size is 274KB. I was off to what seemed to be a good start... Next I applied deflate to my new IDAT section. Result... 59KB.
10KB larger than using 4bpp.
Test 2:
Original File: 102KB, 1000 * 750, 8bpp, 256 Colors, tRNS 1 fully transparent color
After Optimization: 79KB, 8bpp, 193 colors, tRNS 1 full transparent color
Human Interpretation: I need about 24 colors to represent this picture.
24 colors can be represented in 5bpp at 32 colors. Using the same technique above I was able to achieve much better results over uncompressed but again I lost at compression. Final size compressed... 84KB. Then I tried at 6,7bpp... same result poorer compression that at 8bpp.
Just to be sure I saved all the uncompressed images and tried several other compression algorithms... LZMA, BZIP2, PAQ8... same result smaller compression size at 8bpp than at 5,6, or 7bpp AND smaller size at 4bpp than at 3bpp.
Why is this occuring? Can I tweak/modify a compression algorithm to target a PNG like format that uses a 5,6, or 7bpp format that beast 8bpp compression? Is it worth the time... and yes saving another 10KB would be worth it.
What you're seeing is that by using odd pixel sizes, your effective compression decreases because of the way PNG compression works. The advantage of PNG compression over just using straight FLATE/ZIP compression is the filtering. PNG compression tries to exploit horizontal and vertical symmetry with a small assortment of pre-processing filters. These filters work on byte boundaries and are effective with pixel sizes of 4/8/16/24/32/48/64 bits. When you move to an odd size pixel (3/5/6/7 bits) you are defeating the filtering because identically colored pixels won't "cancel each other out" horizontally when filtered on 8-bit boundaries.
Even if the filtering weren't an issue, the way that FLATE compression works, reducing the pixel size from 8 to 7 or 6 bits won't have much effect either because it also assumes a symbol size of 8-bits.
In conclusion...the only benefit you can achieve by using odd sizes of pixels is that the uncompressed data will be smaller. By breaking the pixels' byte boundary symmetry, you defeat much of the benefit of PNG compression.
GIF compression supports all pixel sizes from 1 to 8 bits. It defines the symbol size as the pixel size and doesn't use any pre-filtering. An 8-bit GIF image, if compressed as 7-bit pixels, wouldn't suffer less compression, but also wouldn't benefit because the compression depends more on the repetition of the pixels than the symbol size.
DEFLATE compression used by PNG has two main techniques:
finds repeating byte sequences and encodes them as backreferences
encodes bytes using Huffman coding
By changing pixel length from 8-bit you're out of sync with byte boundaries and DEFLATE won't be able to encode repeating pixel runs as repeated bytes.
And thanks to Huffman coding it doesn't matter that 8-bit pixels have unused bits, because the coding will encode bytes with variable-width codes assigning shortest ones to most frequently occurring values.
Related
I'm writing a tool to detect images on our website that should be flagged for manual intervention to reduce file size. If a "large" image is 100K that might be fine, but if a "small" image is 100K, someone forgot to flatten it or compress it.
I'm looking at the "file density" of an image as the ratio filesize/(height x width). Is there a term for this? Is there some guidance about what a reasonable range for this density should be, so that I can flag images? Or am I thinking about this wrong?
Yes, if the file size is given in bits, then that fraction is known as the bitrate in bits per pixel (bpp) - as sascha points out. For example, an uncompressed image is usually 24 bit (8bit/channel * 3 channels (r,g,b)). Anything at this or higher bitrates is (most often) not compressed.
In general, lossless compression can be achieved at bitrates of about 12bpp (a 2:1 compression ratio). Normally you can aim at much lower bitrates (e.g., 1 bit per pixel, 24:1 compression ratio) and expect decent quality, but it'll depend on the images you're dealing with.
Ok, so I tried to use the imagemagick command:
"convert picA.png -interlace line picB.png"
to make an interlace version of my .png images. Most of the time, I got the resulting image is larger than the original one, which is kinda normal. However, on certain image, the resulting image size is smaller.
So I just wonder why does that happen? I really don't want my new image to lose any quality because of the command.
Also, is there any compatibility problem with interlaced .png image?
EDIT: I guess my problem is that the original image was not compressed as best as it could be.
The following only applies to the cases where the pixel size is >= 8 bits. I didn't investigate for other cases but I expect similar outcomes.
A content-identical interlaced PNG image file will almost always be greater because of the additional data for filter type descriptions required to handle the passes scanlines. This is what I explained in details in this web page based on the PNG RFC RFC2083.
In short, this is because the sum of the below number of bytes for interlaced filter types description per interlacing pass is almost always greater than the image height (which is the number of filter types for non-interlaced images):
nb_pass1_lines = CEIL(height/8)
nb_pass2_lines = (width>4?CEIL(height/8):0)
nb_pass3_lines = CEIL((height-4)/8)
nb_pass4_lines = (width>2?CEIL(height/4):0)
nb_pass5_lines = CEIL((height-2)/4)
nb_pass6_lines = (width>1?CEIL(height/2):0)
nb_pass7_lines = FLOOR(height/2)
Though, theoretically, it can be that the data entropy/complexity accidentally gets lowered enough by the Adam7 interlacing so that, with the help of filtering, the usually additional space needed for filter types with interlacing may be compensated through the deflate compression used for the PNG format. This would be a particular case to be proven as the entropy/complexity is more likely to increase with interlacing because the image data is made less consistent through the interlacing deconstruction.
I used the word "accidentally" because reducing the data entropy/complexity is not the purpose of the Adam7 interlacing. Its purpose is to allow the progressive loading and display of the image through a passes mechanism. While, reducing the entropy/complexity is the purpose of the filtering for PNG.
I used the word "usually" because, as shown in the explanation web page, for example, a 1 pixel image will be described through the same length of uncompressed data whether interlaced or not. So, in this case, no additional space should be needed.
When it comes to the PNG file size, a lower size for interlaced can be due to:
Different non-pixel encoding related content embedded in the file such as palette (in the case of color type =! 3) and non-critical chunks such as chromaticities, gamma, number of significant bits, default background color, histogram, transparency, physical pixel dimensions, time, text, compressed text. Note that some of those non-pixel encoding related content can lead to different display of the image depending on the software used and the situation.
Different pixel encoding related content (which can change the image quality) such as bit depth, color type (and thus the use of palette or not with color type = 3), image size,... .
Different compression related content such as better filtering choices, accidental lower data entropy/complexity due to interlacing as explained above (theoretical particular case), higher compression level (as you mentioned)
If I had to check whether 2 PNG image files are equivalent pixel wise, I would use the following command in a bash prompt:
diff <( convert non-interlaced.png rgba:- ) <( convert interlaced.png rgba:- )
It should return no difference.
For the compatibility question, if the PNG encoder and PNG decoder implement the mandatory aspects of the PNG RFC, I see no reason for the interlacing to lead to a compatibility issue.
Edit 2018 11 13:
Some experiments based on auto evolved distributed genetic algorithms with niche mechanism (hosted on https://en.oga.jod.li ) are explained here:
https://jod.li/2018/11/13/can-an-interlaced-png-image-be-smaller-than-the-equivalent-non-interlaced-image/
Those experiments show that it is possible for equivalent PNG images to have a smaller size interlaced than non-interlaced. The best images for this are tall, they have a one pixel width and have pixel content that appear random. Though, the shape is not the only important aspect for the interlaced image to be smaller than the non-interlaced image as random cases with the same shape lead to different size differences.
So, yes, some PNG images can be identical pixel wise and for non-pixel related content but have a smaller size interlaced than non-interlaced.
So I just wonder why does that happen?
From section Interlacing and pass extraction of the PNG spec.
Scanlines that do not completely fill an integral number of bytes are padded as defined in 7.2: Scanlines.
NOTE If the reference image contains fewer than five columns or fewer than five rows, some passes will be empty.
I would assume the behavior your experiencing is the result of the Adam7 method requiring additional padding.
The following is from TIFF 6.0 Specification Section 9: PackBits Compression
That is the essence of the algorithm. Here are some additional rules:
Pack each row separately. Do not compress across row boundaries.
The number of uncompressed bytes per row is defined to be (ImageWidth + 7)
/ 8. If the uncompressed bitmap is required to have an even number of bytes per
row, decompress into word-aligned buffers.
If a run is larger than 128 bytes, encode the remainder of the run as one or more
additional replicate runs
The first and the third items are easy to understand but I am confused about the second one specifically this: The number of uncompressed bytes per row is defined to be (ImageWidth + 7) / 8. Isn't that only true for 1 bit bi-level image. But to my knowledge, packbits is a byte oriented compression algorithm, it could be used for any type of tiff.
Could someone who knows about tiff and packbits give me some hints?
The TIFF document from this site: http://www.fileformat.info/format/tiff/corion-packbits.htm
has the following at the top:
Abstract
This document describes a simple compression scheme for bilevel
scanned and paint type files.
Motivation
The TIFF specification defines a number of compression schemes.
Compression type 1 is really no compression, other than basic
pixel packing. Compression type 2, based on CCITT 1D
compression, is powerful, but not trivial to implement.
Compression type 5 is typically very effective for most bilevel
images, as well as many deeper images such as palette color and
grayscale images, but is also not trivial to implement. PackBits
is a simple but often effective alternative
So it is clear the additional rules are with respect to bilevel images. For some reason, the above abstract and description are missing from the pdf version of TIFF6.0.
JPEG is a lossy image compression which can give a high compression ratio.
As far as I know, information loss takes place in JPEG during quantization.
Are there any other steps in JPEG compression where the loss takes place or can take place?
If it takes place, then where?
There are 3 aspects of JPEG compression which affect the quality and accuracy of images:
1) Loss of precision takes place during the quantization stage. Accuracy of the colors is lost in order to reduce the amount of data generated.
2) Errors are introduced during the conversion to/from the RGB/YCC color spaces.
3) Errors are introduced during the transformation to/from the frequency domain. The Discrete Cosine Transform converts pixels into the frequency domain. This conversion incurs errors in both directions.
Another place where loss can take place in JPEG compression is the chroma subsampling stage.
My understanding is that most JPEG-compressed images use 4:2:0 color subsampling: after converting each pixel from RGB to YCbCr, the Cb values for a 2x2 block of pixels are averaged to a single value, and the Cr values for that 2x2 block of pixels are also averaged to a single value.
The JPEG standard also supports 4:4:4 (no downsampling).
The JPEG compression encoding process splits a given image into blocks of 8x8 pixels, working with these blocks in future lossy and lossless compressions. [source]
It is also mentioned that if the image is a multiple 1MCU block (defined as a Minimum Coded Unit, 'usually 16 pixels in both directions') that lossless alterations to a JPEG can be performed. [source]
I am working with product images and would like to know both if, and how much benefit can be derived from using multiples of 16 in my final image size (say, using an image with size 480px by 360px) vs. a non-multiple of 16 (such as 484x362). In this example I am not interested in further alterations, editing, or recompression of the final image.
To try to get closer to a specific answer where I know there must be largely generalities: Given a 480x360 image that is 64k and saved at maximum quality in Photoshop [example]:
Can I expect any quality loss from an image that is 484x362
What amount of file size addition can I expect (for this example, the additional space would be white pixels)
Are there any other disadvantages to growing larger than the 8px grid?
I know it's arbitrary to use that specific example, but it would still be helpful (for me and potentially any others pondering an image size) to understand what level of compromise I'd be dealing with in breaking the non-8px grid.
The key issue here is a debate I've had is whether 8-pixel divisible images are higher quality than images that are not divisible by 8-pixels.
8 pixels is the cutoff. The reason is because JPEG images are simply an array of 8x8 DCT blocks; if the image resolution isn't mod8 in both directions, the encoder has to pad the sides up to the next mod8 resolution. This in practice is not very expensive bit-wise; what's much worse are the cases when an image has sharp black lines (such as a letterboxed image) that don't lie on block boundaries. This is especially problematic in video encoding. The reason for this being a problem is that the frequency transform of a sharp line is a Gaussian distribution of coefficients--resulting in an enormous number of bits to code.
For those curious, the most common method of padding edges in intra compression (such as JPEG images) is to mirror the lines of pixels before the edge. For example, if you need to pad three lines and line X is the edge, line X+1 is equal to line X, line X+2 is equal to line X-1, and line X+3 is equal to line X-2. This quite effectively minimizes the cost in transform coefficients of the extra lines.
In inter coding, however, the padding algorithms generally simply duplicate the last line, because the mirror method does not work well for inter compression, such as in video compression.
Sometimes you need to use 16 pixel boundaries rather than 8 because of subsampling; every 2nd pixel is thrown away during the encoding process, and those 8x8 DCT blocks started out as 16x16 and will decode back to 16x16. This won't be a problem at the highest quality settings.
A JPG with sizes being multiplies of 8 can also be rotated/flipped with no quality loss. For example gthumb can do this on Linux.
The image dimensions being multiples of 8 or 16 is not going to affect the size on disk very much, but you can get dramatic savings if you can line up the visual contents to the 8x8 pixel grid, such as if there is a repeating pattern or texture in the image.
What Tometzky said. If you don't have the correct multiple, the lossless flip and rotate algorithms don't work. That's because the padding on the right/bottom that can be safely ignored now ends up on the left/top, where it can't.