Merge branch 'jackmiller1-multiple-fonts'. Fixes #1.

README.md

@@ -19,8 +19,9 @@ Usage is as follows:
    This step may take a while as it will extract 108,634 images.
 
 2. `./gen.py 1000`: Generate 1000 test set images in `test/`. (`test/` must not
-    already exist.) This step requires `UKNumberPlate.ttf` to be in the current
-    directory, which can be [downloaded here](http://www.dafont.com/uk-number-plate.font).
+    already exist.) This step requires `UKNumberPlate.ttf` to be in the
+    `fonts/` directory, which can be
+    [downloaded here](http://www.dafont.com/uk-number-plate.font).
 
 3. `./train.py`: Train the model. A GPU is recommended for this step. It will
    take around 100,000 batches to converge. When you're satisfied that the
@@ -35,3 +36,8 @@ The project has the following dependencies:
 * OpenCV
 * NumPy
 
+Different typefaces can be put in `fonts/` in order to match different type
+faces.  With a large enough variety the network will learn to generalize and
+will match as yet unseen typefaces. See
+[#1](https://github.com/matthewearl/deep-anpr/issues/1) for more information.
+

gen.py

@@ -33,6 +33,7 @@
 )
 
 
+import itertools
 import math
 import os
 import random
@@ -47,18 +48,18 @@
 
 import common
 
-FONT_PATH = "UKNumberPlate.ttf"
+FONT_DIR = "./fonts"
 FONT_HEIGHT = 32  # Pixel size to which the chars are resized
 
 OUTPUT_SHAPE = (64, 128)
 
 CHARS = common.CHARS + " "
 
 
-def make_char_ims(output_height):
+def make_char_ims(font_path, output_height):
     font_size = output_height * 4
 
-    font = ImageFont.truetype(FONT_PATH, font_size)
+    font = ImageFont.truetype(font_path, font_size)
 
     height = max(font.getsize(c)[1] for c in CHARS)
 
@@ -255,27 +256,34 @@ def generate_im(char_ims, num_bg_images):
     return out, code, not out_of_bounds
 
 
-def generate_ims(num_images):
-    """
-    Generate a number of number plate images.
+def load_fonts(folder_path):
+    font_char_ims = {}
+    fonts = [f for f in os.listdir(folder_path) if f.endswith('.ttf')]
+    for font in fonts:
+        font_char_ims[font] = dict(make_char_ims(os.path.join(folder_path,
+                                                              font),
+                                                 FONT_HEIGHT))
+    return fonts, font_char_ims
+
 
-    :param num_images:
-        Number of images to generate.
+def generate_ims():
+    """
+    Generate number plate images.
 
     :return:
         Iterable of number plate images.
 
     """
     variation = 1.0
-    char_ims = dict(make_char_ims(FONT_HEIGHT))
+    fonts, font_char_ims = load_fonts(FONT_DIR)
     num_bg_images = len(os.listdir("bgs"))
-    for i in range(num_images):
-        yield generate_im(char_ims, num_bg_images)
+    while True:
+        yield generate_im(font_char_ims[random.choice(fonts)], num_bg_images)
 
 
 if __name__ == "__main__":
     os.mkdir("test")
-    im_gen = generate_ims(int(sys.argv[1]))
+    im_gen = itertools.islice(generate_ims(), int(sys.argv[1]))
     for img_idx, (im, c, p) in enumerate(im_gen):
         fname = "test/{:08d}_{}_{}.png".format(img_idx, c,
                                                "1" if p else "0")

train.py

@@ -113,8 +113,9 @@ def wrapped(*args, **kwargs):
 
 @mpgen
 def read_batches(batch_size):
+    g = gen.generate_ims()
     def gen_vecs():
-        for im, c, p in gen.generate_ims(batch_size):
+        for im, c, p in itertools.islice(g, batch_size):
             yield im, code_to_vec(p, c)
 
     while True:
@@ -233,6 +234,7 @@ def do_batch():
         try:
             last_batch_idx = 0
             last_batch_time = time.time()
+            print "1"
             batch_iter = enumerate(read_batches(batch_size))
             for batch_idx, (batch_xs, batch_ys) in batch_iter:
                 do_batch()