JPEG-in-TIF Conversion: Preserve identical array representation - image

Question

I have a JPEG image. I would like to convert the image to a TIF with JPEG compression (i.e. a "JPEG-in-TIF"). I am using GDAL to do this.
I require that the array representations of the original JPEG and converted TIF are identical.
I have found a method to do this but I am trying to understand a.) why it appears to work and b.) whether it will fail or have significant additional downsides. (If there is an alternative, sure-fire method for accomplishing this task then that would also be a great answer.)
What I have found so far
I found that doing the following command results in a TIF where perhaps 5% of pixels are different (full code sample below):
gdal_translate in.jpg out.tif -co COMPRESS=JPEG
The GDAL docs say that this command will result in a "lossless" conversion. I take it that "lossless" here is not the same thing as a guarantee that the arrays will be identical?
Specifying the block sizes to be equal to the size of the image appears to work?
If I additionally specify block size arguments in the conversion command, then the resulting arrays are identical (in my sample images):
# Suppose that in.jpg is 400x245
gdal_translate in.jpg out.tif -co COMPRESS=JPEG -co BLOCKXSIZE=400 -co BLOCKYSIZE=245
My (very fuzzy) best-guess intuition right now is that specifying this large block size prevents compression of the individual blocks? However,
The docs state that lossless compression will occur "without decompression and compression cycles." So no compression should actually be occurring?
I am confused as to why gdal_translate is happy to accept BLOCKYSIZE=245 when 245 is not a multiple of 8?
The resulting out.tif has 400x245 block(s) for each band. This appears to negatively affect read time in the resulting image.
Block size is only necessary when SOURCE_COLOR_SPACE is not RGB?
In further experiments (see demo code) it appears that - if an image has RGB color space - then it is sufficient to specify -co COMPRESS=JPEG and nothing further in the conversion command.
Demo code
Here is some python code to demonstrate actual commands:
import os
from typing import Tuple
import requests
import numpy as np
import gdal
def read_image_gdal(filepath:str) -> np.ndarray:
try:
f = gdal.Open(filepath)
arr = f.ReadAsArray()
return arr
finally:
f = None
del f
def get_image_size(filepath:str) -> Tuple[int,int]:
try:
f = gdal.Open(filepath)
return f.RasterXSize, f.RasterYSize
finally:
f = None
del f
# Ex. 1 Medium-sized image (400x245) with YCbCr color space ~> TIF_0 is different
SAMPLE_JPEG_URL= 'https://jpeg.org/images/about.jpg'
# # Ex. 2 Small (50x50) image with RGB Colorspace ~> TIF_0 is the SAME
# SAMPLE_JPEG_URL = 'https://raw.githubusercontent.com/OSGeo/gdal/0402b86928e09860e6d24215b6f5611c31fa3c69/autotest/gdrivers/data/jpeg/rgbsmall_rgb.jpg'
# # Ex. 3 Small (50x50) image but has CMYK color space. ~> TIF_0 is different (But see e.g. https://gdal.org/drivers/raster/gtiff.html#raw-mode)
# SAMPLE_JPEG_URL = 'https://raw.githubusercontent.com/OSGeo/gdal/master/autotest/gdrivers/data/jpeg/rgb_ntf_cmyk.jpg'
SAMPLE_JPEG_FILEPATH = "sample_file.jpg"
# Download sample JPEG file from GDAL github
res = requests.get(SAMPLE_JPEG_URL)
with open(SAMPLE_JPEG_FILEPATH, "wb") as file:
file.write(res.content)
# ###########################
# EXPERIMENT: Convert JPEG to TIF with JPEG compression + settings.
# GOAL: TIF's array is same as original JPEG.
# ###########################
# Create a JPEG-in-TIF
TIF_0_FILEPATH = "sample_tif_0.tif"
os.system(f"gdal_translate -co COMPRESS=JPEG {SAMPLE_JPEG_FILEPATH} {TIF_0_FILEPATH}")
# JPEG-in-TIF BUT specify block size explicitly as the size of the image.
TIF_BS_FILEPATH = "sample_tif_bs.tif"
xs, ys = get_image_size(SAMPLE_JPEG_FILEPATH)
print(f"Size of image: {xs}x{ys}")
os.system(f"gdal_translate -co COMPRESS=JPEG -co BLOCKXSIZE={xs} -co BLOCKYSIZE={ys} {SAMPLE_JPEG_FILEPATH} {TIF_BS_FILEPATH}")
# Read and compare resulting arrays
arr = read_image_gdal(SAMPLE_JPEG_FILEPATH)
arr_0 = read_image_gdal(TIF_0_FILEPATH)
arr_bs = read_image_gdal(TIF_BS_FILEPATH)
print("Share Pixels different")
print(" JPEG-in-TIF: ", (arr != arr_0).mean())
print(" Explicit block size: ", (arr != arr_bs).mean())
Output is:
Size of image: 400x245
Share Pixels different
JPEG-in-TIF: 0.051727891156462584
Explicit block size: 0.0
The two additional SAMPLE_JPEG_URLs try to suss out whether the SOURCE_COLOR_SPACE matters. (An alternative explanation is that example #2 is a small image and that this affects the results.)