Thanks for your contribution. Question: do you have a use case where this actually increases your throughput? If so, how big are your input images?

I ran a few benchmarks, and the throughput improvement is nearly linear per core (I have 8 cores for the numbers below). For testing out the 200x193 I used 10k images and for the 1919x2670 I used 100 images.

resolution (200, 193)
time 0.6368 seconds after unlocking GIL vs 4.279 seconds on master
resolution (1919, 2670)
time 1.227 seconds after unlocking GIL vs 9.349 seconds on master

pool = ThreadPoolExecutor(max_workers=os.cpu_count())
t1 = time.time()
output = list(pool.map(lambda x: zxingcpp.read_barcode(x), ims))
t2 = time.time()

Not sure how familiar you are with Python -- essentially with the GIL locked no other python code can execute in any other threads.

My specific application is automated inspection in electronics manufacturing -- camera array scanning labels from panels of circuit boards. I need to scan ~40 barcodes in real-time, but it's grossly parallel.

Thanks for your benchmarking information. Makes your change look obviously very worthwhile. :). If you care about performance, make sure you only look for symbologies (BarcodeFormats) that you are actually interested in. It sounds like you have a rather standardized set of input images. I'd be interested to see a sample if possible, maybe I can provide more tips on how to improve the performance for that type of input.