Implement putText()

[sfe-SparkFro]

Was almost implemented in 2f9b0f7, however there were a couple issues, so it was commented out. The problems:

• The firmware size exploded, from roughly 2.4MB to 4.8MB.
• It a long time to execute.
• The output did not match what was expected.

For example, running the following in normal OpenCV Python:

img = np.zeros((6,8), dtype=np.uint8)
cv2.putText(img, "H i", (0, 4), cv2.FONT_HERSHEY_SIMPLEX, 0.15, (255))

results in the following:

array([[  0,   0,   0,   0,   0,   0,   0,   0],
       [  0, 255,   0, 255,   0,   0, 255,   0],
       [  0, 255, 255, 255,   0,   0, 255,   0],
       [  0, 255,   0, 255,   0,   0, 255,   0],
       [  0, 255,   0, 255,   0,   0, 255,   0],
       [  0,   0,   0,   0,   0,   0,   0,   0]], dtype=uint8)

Running in MicroPython, it takes almost 4 seconds, and the output is different:

array([[0, 0, 0, 0, 7, 0, 0, 0],
       [105, 0, 105, 0, 63, 0, 0, 0],
       [128, 98, 132, 0, 101, 0, 0, 0],
       [106, 0, 106, 0, 100, 0, 0, 0],
       [0, 0, 0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0, 0, 0]], dtype=uint8)

[sfe-SparkFro]

It's possible the difference in output is due to different OpenCV versions. The OpenCV install on my computer is v4.10, whereas the MicroPython implementation is built on v5.0-alpha. Should test to confirm.

[sfe-SparkFro]

It seems the long execution time is only on the first time putText() is called. Subsequent calls execute immediately. The fast execution also persists through a soft-reset, so I wonder if it's just taking a long time to load the fonts or something?

[sfe-SparkFro]

Yep, looks like the reason for the large firmware size and long initial load time is that the font files are huge.

In the OpenCV build files, it generates a couple files called builtin_font_sans.h and builtin_font_italic.h, each of which contain a single large array of bytes, and are almost 1MB in size each. There's another font file called builtin_font_uni.h (Unicode?) that is over 12MB, but looks like that is optional and doesn't get included by default (thank goodness!).

Because the fonts are generated, I wonder if there's a way to optimize it for size, or provide our own font files.

[sfe-SparkFro]

Looks like fonts are changing a lot with OpenCV v5.0. Relevant PR: opencv/opencv#18760

Makes me wonder if we should just stick with OpenCV 4.x. I think the font stuff is a lot simpler, making it possibly better for MicroPython.

[gigapod]

I'd try 4.x and see if it helps with font sizing. Hershey fonts are simple - just a bunch of vectors - and really all that's needed. Once you get into truetype fonts, it gets complicated.

Looking at the openCV build scripts (v5 - I think). It's taking truetype files during the build process to create those font header files.

[sfe-SparkFro]

Confirm that OpenCV 4.x is much better for putText(). Firmware size is back down to about 2.4MB, and putText() runs immediately on the first iteration as expected.

Not sure there's much benefit to us using 5.x, if at all, so will downgrade to 4.11 (latest of 4.x)

[sfe-SparkFro]

Also worth noting, after changing to 4.x, the putText() output in MicroPython matches normal OpenCV Python.

To my knowledge, OpenCV 5.x has completely changed the font rendering, entirely removing the Hershey fonts in favor of TrueType:

https://github.com/opencv/opencv/pull/18760/files#diff-91cabb4f1fe2c96aa3e56872dce0b2a7e1c5db39e83ba31914a2b547ae61afc1

Not good for microcontrollers... Best to stay on 4.x, at least for now, especially since 5.x still isn't officially released.