Find optimal parameters for tesseract OCR, given a set of input files and expected output.

find_tesseract_param.py

#!/usr/bin/env python3
import Levenshtein
import scipy.optimize
import subprocess
import glob
import sys

p = subprocess.Popen(["tesseract", "--print-parameters"], stdout=subprocess.PIPE)
args, _ = p.communicate()

args_name=[]
args_vec=[]
for argline in args.decode().split("\n"):
    try:
        [argname, val, desc] = argline.split("\t")
        val = float(val)
        if "debug" not in argname:
            args_name.append(argname)
            args_vec.append(val)
    except: pass

objective_text = "\n".join(open(f).read() for f in glob.glob("test_data/*txt"))
additional_args = sys.argv[1:]

def eval(vec):
    decoded = ""
    processes = []
    # Run processes in parallel
    for image in glob.glob("test_data/*png"):
        cmd_line = ["tesseract", image, "stdout"] + additional_args
        for nameval in zip(args_name, vec):
            cmd_line.append("-c")
            cmd_line.append("%s=%f" % nameval) 
        p = subprocess.Popen(cmd_line, stdout=subprocess.PIPE)
        processes.append(p)
    for p in processes:
        res, _ = p.communicate()
        decoded += res.decode()
    print(decoded)
    return Levenshtein.distance(decoded, objective_text)
        

result = scipy.optimize.fmin_powell(eval, args_vec, maxiter=1e4, maxfun=1e4)
print(result)