hocr-eval-lines

#!/usr/bin/python

# compute statistics about the quality of the geometric segmentation
# at the level of the given OCR element

import sys,os,string,re,getopt
import xml
from xml.dom.ext.reader import HtmlLib
from xml.xpath import Evaluate as xquery
from pylab import array,zeros,reshape

################################################################
### misc library code
################################################################

def assoc(key,list):
    for k,v in list:
        if k==key: return v
    return None

def get_text(node):
    textnodes = xquery(".//text()",node)
    s = string.join([node.nodeValue for node in textnodes])
    return re.sub(r'\s+',' ',s)

simp_re = re.compile(r'[^a-zA-Z0-9.,!?:;]+')

def normalize(s):
    s = simp_re.sub(' ',s)
    s = s.strip()
    return s

def edit_distance(a,b,threshold=999999):
    if a==b: return 0
    m = len(a)
    n = len(b)
    distances = zeros((m+1,n+1))
    distances[:,:] = threshold
    distances[:,0] = array(range(m+1))
    distances[0,:] = array(range(n+1))
    for i in range(1,m+1):
        for j in range(1,n+1):        
            if a[i-1] == b[j-1]:
                cij = 0
            else:
                cij = 1
            d = min(
                distances[i-1,j] + 1, 
                distances[i,j-1] + 1, 
                distances[i-1,j-1] + cij
            )
            if d>=threshold: return d
            distances[i,j] = d
    return distances[m,n]

################################################################
### main program
################################################################

if len(sys.argv)<3:
    print "usage: %s [-v] true-lines.txt hocr-actual.html"%sys.argv[0]
    sys.exit(0)
optlist,args = getopt.getopt(sys.argv[1:],"v")

verbose = (assoc('-v',optlist)=='')
truth_lines = open(args[0]).read().split('\n')
actual_doc = HtmlLib.Reader().fromString(open(args[1]).read())
actual_lines = [get_text(node) for node in xquery("//*[@class='ocr_line']",actual_doc)]

truth_lines = [normalize(s) for s in truth_lines]
truth_lines = [s for s in truth_lines if s!=""]
actual_lines = [normalize(s) for s in actual_lines]
actual_lines = [s for s in truth_lines if s!=""]

remaining = []+truth_lines
ocr_errors = 0
for actual_line in actual_lines:
    min_d = 999999
    min_i = -1
    for index in range(len(remaining)):
        true_line = remaining[index]
        d = edit_distance(true_line,actual_line,min_d)
        if d<min_d:
            min_d = d
            min_i = index
    if verbose and min_d>0:
        print "distance",min_d
        print "\t"+actual_line
        print "\t"+remaining[min_i]
    assert min_i>=0
    del remaining[min_i]
    ocr_errors += min_d

segmentation_errors = 0
for s in remaining: segmentation_errors += len(s)

print "segmentation_errors",segmentation_errors
print "ocr_errors",ocr_errors