def ocr(trainingset="../chars74k-lite", testingset="../detection-images"):
#########
#
# INITIALIZATION
#
#########
window_character_threshold = 1.0
patches_containing_charater = []
characters = set()
# TRAINING CASES
dg = DataGenerator(dataset=trainingset, normalized=True)
dg.shuffle_data()
# extract the training data from our data-generator
training_cases, training_labels = dg.get_training_partition(percentage=1)
# convert training data from float64 to float32
training_cases = training_cases.astype(np.float32)
# print(training_cases[0].shape) # (400,)
# TESTING CASES
dg = DataGenerator(dataset=testingset, normalized=True)
# extract the testing data from our data-generator
testing_cases, _ = dg.get_data()
testing_labels = training_labels
# print(np.array(testing_cases[0]).shape) # (40000,)
cases_reshaped = []
for idx, case in enumerate(testing_cases):
h, w = dg.get_dimensions(idx)
case = np.array(case)
case = case.reshape(h, w)
cases_reshaped.append(case)
# print(cases_reshaped[0].shape) # (200,200) # (600,300)
# windows = window_stack(cases_reshaped[0], 1, 20)
# print(windows.shape) # (181,4000)
# print(windows[0].shape) # (4000,)
#########
#
# OCR
#
#########
# create the classifier to run over the windows
# Create the Conv model
print("Creating model")
window_width = 20
window_height = 20
test_img = testing_cases[0]
image_width = 200
image_height = 200
model = init_model(window_height, window_width)
# Train the model
print("Training model")
model = train_conv_net(model, training_cases, training_labels)
print("Creating windows")
columns = image_width - window_width
rows = image_height - window_height
windows = []
xy = []
for i in range(columns):
for j in range(rows):
index = i + j * image_width
xy.append((i, j))
windows.append(
get_window(test_img, index, window_width, window_height,
image_width).reshape(1, 400))
windows = np.array(windows)
print("Predicting windows")
template = ('Prediction is "{}" ({:.1f}%)')
predictions = prediction_conv_net(model, windows)
fix, ax = plt.subplots(1)
ax.imshow(np.reshape(test_img, (-1, image_width)))
cmap = cm.autumn
norm = Normalize(vmin=0, vmax=27)
possible = []
for pred_dict, window, pos in zip(predictions, windows, xy):
class_id = pred_dict['class_ids'][0]
probability = pred_dict['probabilities'][class_id]
if probability >= window_character_threshold:
# window_2d = np.reshape(window, (-1, window_width))
# plt.imshow(window_2d, interpolation='nearest')
# plt.title(dg.int_to_char(class_id))
# plt.show()
# print(template.format(dg.int_to_char(class_id), 100 * probability))
possible.append([pos, class_id])
characters.add(dg.int_to_char(class_id))
# WHEN NO FILTERING!!!
# rect = patches.Rectangle(pos, window_width, window_height, linewidth=1, edgecolor=cmap(norm(class_id)),
# facecolor='none')
# ax.add_patch(rect)
# WHEN FILTERING
used = []
filtered = set()
for i in range(len(possible)):
if i in used:
continue
pos = possible[i]
# avg_x = pos[0][0]
# avg_y = pos[0][1]
same = 1
diff = 0
for j in range(len(possible)):
if i != j:
pos2 = possible[j]
pos2_x = pos2[0][0]
pos2_y = pos2[0][1]
if abs(pos[0][0] - pos2_x) <= 10 and abs(pos[0][1] -
pos2_y) <= 10:
if pos[1] == pos2[1]:
same += 1
else:
diff += 1
if diff / same > 1.0:
continue
filtered.add(dg.int_to_char(pos[1]))
handles = []
for letter in filtered: # FILTERING
# for letter in characters: # NO FILTERING
handles.append(
patches.Patch(color=cmap(norm(dg.char_to_int(letter))),
label=letter))
# red_patch = patches.Patch(color='red', label='The red data')
plt.legend(handles=handles)
plt.show()
print("Characters found with threshold " +
str(window_character_threshold) + ":" + str(characters))
