megadetector 5.0.5__py3-none-any.whl → 5.0.7__py3-none-any.whl

data_management/ocr_tools.py

@@ -0,0 +1,829 @@
+########
+#
+# ocr_tools.py
+#
+# Use OCR (via the Tesseract package) to pull metadata (particularly times and
+# dates from camera trap images).
+#
+# The general approach is:
+#
+# * Crop a fixed percentage from the top and bottom of an image, slightly larger
+#   than the largest examples we've seen of how much space is used for metadata.
+#
+# * Define the background color as the median pixel value, and find rows that are
+#   mostly that color to refine the crop.
+#
+# * Crop to the refined crop, then run pytesseract to extract text.
+#
+# * Use regular expressions to find time and date.
+#
+# Prior to using this module:
+#
+# * Install Tesseract from https://tesseract-ocr.github.io/tessdoc/Installation.html
+#
+# * pip install pytesseract
+#    
+# Known limitations:
+#
+# * Semi-transparent overlays (which I've only seen on consumer cameras) usually fail.
+#    
+########
+
+#%% Notes to self
+
+"""
+
+* To use the legacy engine (--oem 0), I had to download an updated eng.traineddata file from:
+    
+  https://github.com/tesseract-ocr/tessdata
+  
+"""
+
+#%% Constants and imports
+
+import os
+import json
+import numpy as np
+import datetime
+import re
+
+from functools import partial
+from dateutil.parser import parse as dateparse
+
+import cv2
+from PIL import Image, ImageFilter
+from tqdm import tqdm
+
+from md_utils.path_utils import find_images
+from md_visualization import visualization_utils as vis_utils
+from md_utils import write_html_image_list        
+from md_utils.path_utils import open_file
+
+# pip install pytesseract
+#
+# Also install tesseract from: https://github.com/UB-Mannheim/tesseract/wiki, and add
+# the installation dir to your path (on Windows, typically C:\Program Files (x86)\Tesseract-OCR)
+import pytesseract
+
+
+#%% Extraction options
+
+class DatetimeExtractionOptions:
+            
+    def __init__(self):
+        
+        # Using a semi-arbitrary metric of how much it feels like we found the 
+        # text-containing region, discard regions that appear to be extraction failures
+        self.p_crop_success_threshold = 0.5
+        
+        # Pad each crop with a few pixels to make tesseract happy
+        self.crop_padding = 10        
+        
+        # Discard short text, typically text from the top of the image
+        self.min_text_length = 4
+        
+        # When we're looking for pixels that match the background color, allow some 
+        # tolerance around the dominant color
+        self.background_tolerance = 2
+            
+        # We need to see a consistent color in at least this fraction of pixels in our rough 
+        # crop to believe that we actually found a candidate metadata region.
+        self.min_background_fraction = 0.3
+        
+        # What fraction of the [top,bottom] of the image should we use for our rough crop?
+        self.image_crop_fraction = [0.045 , 0.045]
+        # self.image_crop_fraction = [0.08 , 0.08]
+        
+        # Within that rough crop, how much should we use for determining the background color?
+        self.background_crop_fraction_of_rough_crop = 0.5
+        
+        # A row is considered a probable metadata row if it contains at least this fraction
+        # of the background color.  This is used only to find the top and bottom of the crop area, 
+        # so it's not that *every* row needs to hit this criteria, only the rows that are generally
+        # above and below the text.
+        self.min_background_fraction_for_background_row = 0.5
+        
+        # psm 6: "assume a single uniform block of text"
+        # psm 13: raw line
+        # oem: 0 == legacy, 1 == lstm
+        # tesseract_config_string = '--oem 0 --psm 6'
+        #
+        # Try these configuration strings in order until we find a valid datetime
+        self.tesseract_config_strings = ['--oem 1 --psm 13','--oem 0 --psm 13',
+                                         '--oem 1 --psm 6','--oem 0 --psm 6']
+        
+        self.force_all_ocr_options = False
+        
+        self.apply_sharpening_filter = True
+        
+        # Tesseract should be on your system path, but you can also specify the
+        # path explicitly.
+        #
+        # os.environ['PATH'] += r';C:\Program Files\Tesseract-OCR'
+        # self.tesseract_cmd = 'r"C:\Program Files\Tesseract-OCR\tesseract.exe"'
+        self.tesseract_cmd = 'tesseract.exe'
+
+
+#%% Support functions
+
+def make_rough_crops(image,options=None):
+    """
+    Crops the top and bottom regions out of an image, returns a dict with fields
+    'top' and 'bottom', each pointing to a PIL image.
+    
+    [image] can be a PIL image or a file name.
+    """
+    
+    if options is None:
+        options = DatetimeExtractionOptions()
+        
+    if isinstance(image,str):
+        image = vis_utils.open_image(image)
+        
+    w = image.width
+    h = image.height
+    
+    crop_height_top = round(options.image_crop_fraction[0] * h)
+    crop_height_bottom = round(options.image_crop_fraction[1] * h)
+    
+    # l,t,r,b
+    #
+    # 0,0 is upper-left
+    top_crop = image.crop([0,0,w,crop_height_top])
+    bottom_crop = image.crop([0,h-crop_height_bottom,w,h])
+    return {'top':top_crop,'bottom':bottom_crop}
+    
+# ...def make_rough_crops(...)
+
+
+def crop_to_solid_region(rough_crop,crop_location,options=None):
+    """       
+    Given a rough crop from the top or bottom of an imaeg, find the background color
+    and crop to the metadata region.
+    
+    rough_crop should be PIL Image, crop_location should be 'top' or 'bottom'.
+    
+    Within a region of an image (typically a crop from the top-ish or bottom-ish part of 
+    an image), tightly crop to the solid portion (typically a region with a black background).
+
+    The success metric is just a binary indicator right now: 1.0 if we found a region we believe
+    contains a solid background, 0.0 otherwise.
+    
+    Returns cropped_image,p_success,padded_image
+    """
+    
+    if options is None:
+        options = DatetimeExtractionOptions()    
+
+    crop_to_solid_region_result = {}
+    crop_to_solid_region_result['crop_pil'] = None
+    crop_to_solid_region_result['padded_crop_pil'] = None
+    crop_to_solid_region_result['p_success'] = 0.0
+    
+    # pil --> cv2        
+    rough_crop_np = np.array(rough_crop) 
+    rough_crop_np = rough_crop_np[:, :, ::-1].copy()         
+    
+    # Search *part* of the crop for the background value (the part closest to the top or bottom
+    # of the image)
+    rows_to_use_for_background_search = int(rough_crop_np.shape[0] * \
+                                            options.background_crop_fraction_of_rough_crop)
+    
+    if crop_location == 'top':
+        background_search_image = rough_crop_np[0:rows_to_use_for_background_search,:,:]
+    elif crop_location == 'bottom':
+        background_search_image = rough_crop_np[-rows_to_use_for_background_search:,:,:]
+    else:
+        raise ValueError('Unrecognized crop location: {}'.format(crop_location))
+                
+    background_search_image = cv2.cvtColor(background_search_image, cv2.COLOR_BGR2GRAY)
+    background_search_image = background_search_image.astype('uint8')     
+    background_search_image = cv2.medianBlur(background_search_image,3) 
+    pixel_values = background_search_image.flatten()
+    counts = np.bincount(pixel_values)
+    background_value = int(np.argmax(counts))
+    
+    # Did we find a sensible mode that looks like a background value?
+    background_value_count = int(np.max(counts))
+    p_background_value = background_value_count / np.sum(counts)
+    
+    if (p_background_value < options.min_background_fraction):
+        return crop_to_solid_region_result
+    else:
+        p_success = 1.0
+        
+    analysis_image = cv2.cvtColor(rough_crop_np, cv2.COLOR_BGR2GRAY)
+    analysis_image = analysis_image.astype('uint8')     
+    analysis_image = cv2.medianBlur(analysis_image,3) 
+    
+    # This will now be a binary image indicating which pixels are background
+    analysis_image = cv2.inRange(analysis_image,
+                                background_value-options.background_tolerance,
+                                background_value+options.background_tolerance)
+    
+    # Use row heuristics to refine the crop        
+    h = analysis_image.shape[0]
+    w = analysis_image.shape[1]
+    
+    min_x = 0
+    min_y = -1
+    max_x = w
+    max_y = -1
+    
+    # Find the first and last row that are mostly the background color
+    for y in range(h):
+        row_count = 0
+        for x in range(w):
+            if analysis_image[y][x] > 0:
+                row_count += 1
+        row_fraction = row_count / w
+        if row_fraction > options.min_background_fraction_for_background_row:
+            if min_y == -1:
+                min_y = y
+            max_y = y
+    
+    assert (min_y == -1 and max_y == -1) or (min_y != -1 and max_y != -1)
+    
+    if min_y == -1:
+        return crop_to_solid_region_result
+    
+    if max_y == min_y:
+        return crop_to_solid_region_result
+    
+    x = min_x
+    y = min_y
+    w = max_x-min_x
+    h = max_y-min_y
+    
+    x = min_x
+    y = min_y
+    w = max_x-min_x
+    h = max_y-min_y
+
+    # Crop the image
+    crop_np = rough_crop_np[y:y+h,x:x+w]
+      
+    # Tesseract doesn't like characters really close to the edge, so pad a little.
+    crop_padding = options.crop_padding
+    padded_crop_np = cv2.copyMakeBorder(crop_np,crop_padding,crop_padding,crop_padding,crop_padding,
+                                     cv2.BORDER_CONSTANT,
+                                     value=[background_value,background_value,background_value])
+
+    crop_pil = Image.fromarray(crop_np)
+    padded_crop_pil = Image.fromarray(padded_crop_np)
+    
+    crop_to_solid_region_result['crop_pil'] = crop_pil
+    crop_to_solid_region_result['padded_crop_pil'] = padded_crop_pil
+    crop_to_solid_region_result['p_success'] = p_success
+    
+    return crop_to_solid_region_result
+    
+# ...crop_to_solid_region(...)    
+
+
+def find_text_in_crops(rough_crops,options=None,tesseract_config_string=None):
+    """
+    Find all text in each Image in the dict [rough_crops]; those images should be pretty small 
+    regions by the time they get to this function, roughly the top or bottom 20% of an image.
+    """
+    
+    if options is None:
+        options = DatetimeExtractionOptions()
+    
+    if tesseract_config_string is None:
+        tesseract_config_string = options.tesseract_config_strings[0]
+        
+    find_text_in_crops_results = {}
+    
+    # crop_location = 'top'
+    # crop_location = 'bottom'
+    for crop_location in ('top','bottom'):
+
+        find_text_in_crops_results[crop_location] = {}
+        find_text_in_crops_results[crop_location]['text'] = ''
+        find_text_in_crops_results[crop_location]['crop_to_solid_region_results'] = None
+        
+        rough_crop = rough_crops[crop_location]
+        
+        # Crop to the portion of the rough crop with a solid background color
+        crop_to_solid_region_results = crop_to_solid_region(rough_crop,crop_location,options)
+        
+        find_text_in_crops_results[crop_location]['crop_to_solid_region_results'] = \
+            crop_to_solid_region_results
+        
+        # Try cropping to a solid region; if that doesn't work, try running OCR on the whole
+        # rough crop.
+        if crop_to_solid_region_results['p_success'] >= options.p_crop_success_threshold:
+            padded_crop_pil = crop_to_solid_region_results['padded_crop_pil']
+        else:            
+            # continue
+            padded_crop_pil = rough_crop        
+            
+        if options.apply_sharpening_filter:
+            padded_crop_pil = padded_crop_pil.filter(ImageFilter.SHARPEN)
+        
+        # Find text in the padded crop
+        pytesseract.pytesseract.tesseract_cmd = options.tesseract_cmd
+        text = pytesseract.image_to_string(padded_crop_pil, lang='eng', 
+                                           config=tesseract_config_string)
+        
+        text = text.replace('\n', ' ').replace('\r', '').strip()
+
+        find_text_in_crops_results[crop_location]['text'] = text        
+                
+    # ...for each cropped region
+    
+    return find_text_in_crops_results
+    
+# ...def find_text_in_crops(...)
+    
+
+def datetime_string_to_datetime(matched_string):
+    """
+    Takes an OCR-matched datetime string, does a little cleanup, and parses a date
+    from it.
+    
+    By the time a string gets to this function, it should be a proper date string, with
+    no extraneous characters other than spaces around colons or hyphens.
+    """
+    
+    matched_string = matched_string.replace(' -','-')
+    matched_string = matched_string.replace('- ','-')
+    matched_string = matched_string.replace(' :',':')
+    matched_string = matched_string.replace(': ',':')
+    try:
+        extracted_datetime = dateparse(matched_string)
+    except Exception:
+        extracted_datetime = None
+    return extracted_datetime
+
+
+def get_datetime_from_strings(strings,options=None):
+    """
+    Given a string or list of strings, search for exactly one datetime in those strings.
+    using a series of regular expressions.
+    
+    Strings are currently just concatenated before searching for a datetime.
+    """
+    
+    if options is None:
+        options = DatetimeExtractionOptions()    
+    
+    if isinstance(strings,str):
+        s = strings
+    else:
+        s = ' '.join(strings).lower()
+    s = s.replace('—','-')    
+    s = ''.join(e for e in s if e.isalnum() or e in ':-/' or e.isspace())
+        
+    ### AM/PM
+    
+    # 2013-10-02 11:40:50 AM
+    m = re.search('(\d\d\d\d)\s?-\s?(\d\d)\s?-\s?(\d\d)\s+(\d+)\s?:?\s?(\d\d)\s?:\s?(\d\d)\s*([a|p]m)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))        
+    
+    # 04/01/2017 08:54:00AM
+    m = re.search('(\d\d)\s?/\s?(\d\d)\s?/\s?(\d\d\d\d)\s+(\d+)\s?:\s?(\d\d)\s?:\s?(\d\d)\s*([a|p]m)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))    
+    
+    # 2017/04/01 08:54:00AM
+    m = re.search('(\d\d\d\d)\s?/\s?(\d\d)\s?/\s?(\d\d)\s+(\d+)\s?:\s?(\d\d)\s?:\s?(\d\d)\s*([a|p]m)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))        
+    
+    # 04/01/2017 08:54AM
+    m = re.search('(\d\d)\s?/\s?(\d\d)\s?/\s?(\d\d\d\d)\s+(\d+)\s?:\s?(\d\d)\s*([a|p]m)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))    
+    
+    # 2017/04/01 08:54AM
+    m = re.search('(\d\d\d\d)\s?/\s?(\d\d)\s?/\s?(\d\d)\s+(\d+)\s?:\s?(\d\d)\s*([a|p]m)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))        
+    
+    ### No AM/PM
+    
+    # 2013-07-27 04:56:35
+    m = re.search('(\d\d\d\d)\s?-\s?(\d\d)\s?-\s?(\d\d)\s*(\d\d)\s?:\s?(\d\d)\s?:\s?(\d\d)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))        
+    
+    # 07-27-2013 04:56:35
+    m = re.search('(\d\d)\s?-\s?(\d\d)\s?-\s?(\d\d\d\d)\s*(\d\d)\s?:\s?(\d\d)\s?:\s?(\d\d)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))        
+    
+    # 2013/07/27 04:56:35
+    m = re.search('(\d\d\d\d)\s?/\s?(\d\d)\s?/\s?(\d\d)\s*(\d\d)\s?:\s?(\d\d)\s?:\s?(\d\d)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))        
+    
+    # 07/27/2013 04:56:35
+    m = re.search('(\d\d)\s?/\s?(\d\d)\s?/\s?(\d\d\d\d)\s*(\d\d)\s?:\s?(\d\d)\s?:\s?(\d\d)',s)
+    if m is not None:        
+        return datetime_string_to_datetime(m.group(0))        
+    
+    return None
+    
+# ...def get_datetime_from_strings(...)
+
+
+def get_datetime_from_image(image,include_crops=True,options=None):
+    """
+    Find the datetime string (if present) in [image], which can be a PIL image or a 
+    filename.  Returns a dict:
+        
+    datetime: Python datetime object, or None
+    
+    text_results: length-2 list of strings
+    
+    all_extracted_datetimes: if we ran multiple option sets, this will contain the 
+    datetimes extracted for each option set
+    
+    ocr_results: detailed results from the OCR process, including crops as PIL images;
+    only included if include_crops is True.
+        
+    [options] can be None, a DatetimeExtractionOptions object, or a list of 
+    DatetimeExtractionOptions objects to try for each image.    
+    """
+    
+    if options is None:
+        options = DatetimeExtractionOptions()
+    
+    if isinstance(image,str):
+        image = vis_utils.open_image(image)
+
+    # Crop the top and bottom from the image
+    rough_crops = make_rough_crops(image,options)
+    assert len(rough_crops) == 2
+    
+    all_extracted_datetimes = {}
+    all_text_results = []
+    all_ocr_results = []
+    
+    extracted_datetime = None
+    
+    # Find text, possibly trying all config strings
+    #
+    # tesseract_config_string = options.tesseract_config_strings[0]
+    for tesseract_config_string in options.tesseract_config_strings:
+        
+        ocr_results = find_text_in_crops(rough_crops,options,tesseract_config_string)
+        all_ocr_results.append(ocr_results)
+        
+        text_results = [v['text'] for v in ocr_results.values()]
+        assert len(text_results) == 2
+        all_text_results.append(text_results)
+            
+        # Find datetime
+        extracted_datetime_this_option_set = get_datetime_from_strings(text_results,options)
+        assert isinstance(extracted_datetime_this_option_set,datetime.datetime) or \
+            (extracted_datetime_this_option_set is None)
+        
+        all_extracted_datetimes[tesseract_config_string] = \
+            extracted_datetime_this_option_set
+            
+        if extracted_datetime_this_option_set is not None:
+            if extracted_datetime is None:
+                extracted_datetime = extracted_datetime_this_option_set
+            if not options.force_all_ocr_options:
+                break        
+            
+    # ...for each set of OCR options
+    
+    if extracted_datetime is not None:        
+        assert extracted_datetime.year <= 2023 and extracted_datetime.year >= 1990
+
+    to_return = {}
+    to_return['datetime'] = extracted_datetime
+    
+    to_return['text_results'] = all_text_results
+    to_return['all_extracted_datetimes'] = all_extracted_datetimes
+    
+    if include_crops:
+        to_return['ocr_results'] = all_ocr_results
+    else:
+        to_return['ocr_results'] = None
+        
+    return to_return
+
+# ...def get_datetime_from_image(...)
+
+
+def is_iterable(x):
+    try:
+        _ = iter(x)
+    except:
+       return False
+    return True
+
+
+def try_get_datetime_from_image(filename,include_crops=False,options=None):
+    """
+    Try/catch wrapper for get_datetime_from_image, optionally trying multiple option sets
+    until we find a datetime.
+    """
+    
+    if options is None:
+        options = DatetimeExtractionOptions()
+
+    if not is_iterable(options):
+        options = [options]
+    
+    result = {}
+    result['error'] = None
+    
+    for i_option_set,current_options in enumerate(options):
+        try:
+            result = get_datetime_from_image(filename,include_crops=include_crops,options=current_options)
+            result['options_index'] = i_option_set
+            if 'datetime' in result and result['datetime'] is not None:
+                break
+        except Exception as e:
+            result['error'] = str(e)
+    
+    return result
+
+
+def get_datetimes_for_folder(folder_name,output_file=None,n_to_sample=-1,options=None):
+    """
+    Retrieve metadata from every image in [folder_name], and 
+    write the results to the .json file [output_file].
+    
+    [options] can be None, a DatetimeExtractionOptions object, or a list of 
+    DatetimeExtractionOptions objects to try for each image.
+    
+    Returns a dict mapping filenames to datetime extraction results.  Optionally writes
+    results to the .json file [output_file].
+    """
+    
+    if options is None:
+        options = DatetimeExtractionOptions()
+    
+    image_file_names = \
+        find_images(folder_name,convert_slashes=True,
+                    return_relative_paths=False,recursive=True)
+    
+    if n_to_sample > 0:
+        import random
+        random.seed(0)
+        image_file_names = random.sample(image_file_names,n_to_sample)
+        
+    n_cores = 16
+    use_threads = False
+    
+    if n_cores <= 1:
+        
+        all_results = []
+        for fn_abs in tqdm(image_file_names):
+            all_results.append(try_get_datetime_from_image(fn_abs,options=options))
+            
+    else:    
+        
+        # Don't spawn more than one worker per image
+        if n_cores > len(image_file_names):
+            n_cores = len(image_file_names)
+            
+        if use_threads:
+            from multiprocessing.pool import ThreadPool
+            pool = ThreadPool(n_cores)
+            worker_string = 'threads'        
+        else:
+            from multiprocessing.pool import Pool
+            pool = Pool(n_cores)
+            worker_string = 'processes'
+            
+        print('Starting a pool of {} {}'.format(n_cores,worker_string))
+        
+        all_results = list(tqdm(pool.imap(
+            partial(try_get_datetime_from_image,options=options),image_file_names),
+            total=len(image_file_names)))
+    
+    filename_to_results = {}
+    
+    # fn_relative = image_file_names[0]
+    for i_file,fn_abs in enumerate(image_file_names):
+        filename_to_results[fn_abs] = all_results[i_file]
+    
+    if output_file is not None:
+        with open(output_file,'w') as f:
+            json.dump(filename_to_results,f,indent=1,default=str)
+
+    return filename_to_results
+
+
+#%% Interactive driver
+
+if False:
+    
+    #%% Process images
+    
+    folder_name = r'g:\temp\island_conservation_camera_traps'
+    # folder_name = r'g:\camera_traps\camera_trap_images'
+    output_file = r'g:\temp\ocr_results.json'
+    from md_utils.path_utils import insert_before_extension
+    output_file = insert_before_extension(output_file)
+    n_to_sample = -1
+    assert os.path.isdir(folder_name)
+    options_a = DatetimeExtractionOptions()
+    options_b = DatetimeExtractionOptions()
+    options_b.image_crop_fraction = [0.08 , 0.08]
+    options_a.force_all_ocr_options = False
+    options_b.force_all_ocr_options = False
+    # all_options = [options_a,options_b]
+    all_options = [options_a]
+    filename_to_results = get_datetimes_for_folder(folder_name,output_file,
+                                                   n_to_sample=n_to_sample,options=all_options)
+    
+
+    #%% Load results
+    
+    # output_file = r"G:\temp\ocr_results.2023.10.31.07.37.54.json"
+    with open(output_file,'r') as f:
+        filename_to_results = json.load(f)
+    filenames = sorted(list(filename_to_results.keys()))
+    print('Loaded results for {} files'.format(len(filename_to_results)))
+    
+    
+    #%% Scrap cell
+    
+    fn = 'g:/camera_traps/camera_trap_images/2018.07.02/newcam/people/DSCF0273.JPG'
+    # fn = r'g:\camera_traps\camera_trap_images\2022.01.29\cam0\coyote\DSCF0057.JPG'
+    # fn = 'g:/temp/island_conservation_camera_traps/chile/frances01/frances012013/chile_frances012013_02012013105658.jpg'
+    # fn = 'g:/temp/island_conservation_camera_traps/dominicanrepublic/camara06/cam0618junio2016/dominicanrepublic_cam0618junio2016_20160614_114115_img_0013.jpg'
+    # fn = os.path.join(folder_name,r'dominicanrepublic\camara22\cam228noviembre2015\dominicanrepublic_cam228noviembre2015_20151105_071226_img_0132.jpg')
+    # fn = 'g:/camera_traps/camera_trap_images/2021.06.06/camera01/empty/DSCF0873.JPG'
+    include_crops = False
+    options_a = DatetimeExtractionOptions()
+    options_b = DatetimeExtractionOptions()
+    options_b.image_crop_fraction = [0.08 , 0.08]
+    image = vis_utils.open_image(fn) # noqa    
+    result = try_get_datetime_from_image(fn,options=[options_a,options_b]) # noqa
+    print(result)
+    
+    # open_file(fn)
+    # rough_crops = make_rough_crops(image,options=options)
+        
+        
+    #%% Look for OCR or parsing failures
+    
+    bad_tokens = ()
+    
+    files_with_disagreements = set()
+            
+    # i_fn = 0; fn = filenames[i_fn]
+    for i_fn,fn in enumerate(filenames):
+        
+        image = fn
+        results = filename_to_results[fn]
+        
+        if 'text_results' not in results:
+            raise Exception('no results available for {} ({})'.format(i_fn,fn))
+            print('Skipping {}, no results'.format(i_fn))
+            continue
+        
+        s = ' '.join([x[0] for x in results['text_results']])
+        
+        known_bad = False
+        for bad_token in bad_tokens:
+            if bad_token in s:
+                known_bad = True
+        if known_bad: 
+            continue
+                
+        extracted_datetime = results['datetime']
+        
+        # If we have a datetime, make sure all successful OCR results agree
+        if extracted_datetime is not None:
+            for config_string in results['all_extracted_datetimes']:
+                if results['all_extracted_datetimes'][config_string] is not None:
+                    if results['all_extracted_datetimes'][config_string] != extracted_datetime:
+                        files_with_disagreements.add(fn)
+        else:
+            print('Falling back for {} ({})'.format(i_fn,fn))
+            ocr_results = get_datetime_from_image(fn)
+            extracted_datetime = ocr_results['datetime']
+        
+        if extracted_datetime is None:
+            print('Failure at {}: {}'.format(i_fn,s))
+        
+        # open_file(fn)
+        # get_datetime_from_image(fn)
+    
+    
+    #%% Write results to an HTML file for testing
+          
+    n_to_sample = 5000
+    if (n_to_sample >= 0) and (len(filename_to_results) > n_to_sample):
+        filenames = sorted(list(filename_to_results.keys()))        
+        import random
+        random.seed(0)
+        keys = random.sample(filenames,n_to_sample)
+        filename_to_results = {k: filename_to_results[k] for k in keys}
+
+    preview_dir = r'g:\temp\ocr-preview'
+    os.makedirs(preview_dir,exist_ok=True)
+    
+    def resize_image_for_preview(fn_abs):
+        fn_relative = os.path.relpath(fn_abs,folder_name)        
+        resized_image = vis_utils.resize_image(fn_abs,target_width=600)
+        resized_fn = os.path.join(preview_dir,fn_relative)
+        os.makedirs(os.path.dirname(resized_fn),exist_ok=True)
+        resized_image.save(resized_fn)
+        return resized_fn
+        
+    # Resize images in parallel
+    n_rendering_workers = 16
+        
+    if n_rendering_workers <= 1:
+        for fn_abs in tqdm(filename_to_results.keys()):
+            resize_image_for_preview(fn_abs)
+    else:
+        # from multiprocessing.pool import Pool as RenderingPool; worker_string = 'processes'
+        from multiprocessing.pool import ThreadPool as RenderingPool; worker_string = 'threads'
+        pool = RenderingPool(n_rendering_workers)
+        
+        print('Starting rendering pool with {} {}'.format(n_rendering_workers,worker_string))
+    
+        _ = list(tqdm(pool.imap(resize_image_for_preview,filename_to_results.keys()),
+                      total=len(filename_to_results)))
+    
+    
+    def make_datetime_preview_page(filenames,html_file):
+        
+        html_image_list = []
+        html_options = write_html_image_list.write_html_image_list()
+        html_options['maxFiguresPerHtmlFile'] = 2500
+        html_options['defaultImageStyle'] = 'margin:0px;margin-top:5px;margin-bottom:30px;'
+        
+        # fn_abs = filenames[0]
+        for fn_abs in filenames:
+            
+            fn_relative = os.path.relpath(fn_abs,folder_name)        
+            # resized_fn = os.path.join(preview_dir,fn_relative)
+            results_this_image = filename_to_results[fn_abs]
+                
+            extracted_datetime = results_this_image['datetime']
+            title = 'Image: {}<br/>Extracted datetime: {}'.format(fn_relative,extracted_datetime)
+            html_image_list.append({'filename':fn_relative,'title':title})
+            
+            # ...for each crop
+        
+        # ...for each image
+                
+        html_options['makeRelative'] = True
+        write_html_image_list.write_html_image_list(html_file,
+                                                    html_image_list,
+                                                    html_options)
+        open_file(html_file)
+        return html_image_list
+    
+    failed_files = []
+    for fn_abs in filename_to_results:
+        results_this_image = filename_to_results[fn_abs]
+        if results_this_image['datetime'] is None:
+            failed_files.append(fn_abs)
+    
+    print('Found {} failures'.format(len(failed_files)))
+    
+    output_summary_file = os.path.join(preview_dir,'summary.html')
+    html_image_list = make_datetime_preview_page(sorted(list(filename_to_results.keys())),output_summary_file)
+    
+    failure_summary_file = os.path.join(preview_dir,'failures.html')    
+    html_image_list_failures = make_datetime_preview_page(failed_files,failure_summary_file)
+    
+    filenames = failed_files
+    html_file = failure_summary_file
+
+    
+    #%% Other approaches to getting dates from strings
+    
+    # ...that didn't really work out.
+    
+    # pip install dateparser
+    import dateparser
+
+    # pip install datefinder
+    import datefinder
+
+    from dateparser.search import search_dates # noqa
+
+    dateparser_settings = {'PREFER_DATES_FROM':'past','STRICT_PARSING':True}
+
+    dateparser_result = dateparser.search.search_dates(s, settings=dateparser_settings)
+    
+    if dateparser_result is not None:
+        assert len(dateparser_result) == 1
+        extracted_datetime = dateparser_result[0][1]
+    else:
+        matches = datefinder.find_dates(s,strict=False)
+        matches_list = [m for m in matches]
+        if len(matches_list) == 1:
+            extracted_datetime = matches_list[0]
+        else:
+            extracted_datetime = None
+        
+    if extracted_datetime is not None:        
+        assert extracted_datetime.year <= 2023 and extracted_datetime.year >= 1990