def predict():
if request.method == 'POST':
# read image file string data
if 'file' not in request.files:
return redirect(url_for('home'))
else:
filestr = request.files['file'].read()
# convert string data to numpy array
npimg = np.fromstring(filestr, np.uint8)
# convert numpy array to image
img = cv2.imdecode(npimg, cv2.IMREAD_GRAYSCALE)
ctx = ctx = mx.gpu(0) if mx.context.num_gpus() > 0 else mx.cpu()
# Models
paragraph_segmentation_net = SegmentationNetwork(ctx=ctx)
paragraph_segmentation_net.cnn.load_parameters(paragraph_segmentation_model, ctx)
word_segmentation_net = WordSegmentationNet(2, ctx=ctx)
word_segmentation_net.load_parameters(word_segmentation_model, ctx)
handwriting_line_recognition_net = HandwritingRecognitionNet(rnn_hidden_states=rnn_hidden_states,
rnn_layers=rnn_layers,
max_seq_len=max_seq_len,
ctx=ctx)
handwriting_line_recognition_net.load_parameters(recognition_model, ctx)
MAX_IMAGE_SIZE_FORM = (1120, 800)
img_arr = np.asarray(img)
resized_image = paragraph_segmentation_transform(img_arr, image_size=MAX_IMAGE_SIZE_FORM)
paragraph_bb = paragraph_segmentation_net(resized_image.as_in_context(ctx))
paragraph_segmented_image = crop_handwriting_page(img_arr, paragraph_bb[0].asnumpy(),
image_size=segmented_paragraph_size)
word_bb = predict_bounding_boxes(word_segmentation_net, paragraph_segmented_image, min_c, overlap_thres, topk,
ctx)
line_bbs = sort_bbs_line_by_line(word_bb)
line_images = crop_line_images(paragraph_segmented_image, line_bbs)
predicted_text = []
for line_image in line_images:
line_image = handwriting_recognition_transform(line_image, line_image_size)
character_probabilities = handwriting_line_recognition_net(line_image.as_in_context(ctx))
decoded_text = denoise_func(character_probabilities)
predicted_text.append(decoded_text)
text = ' '.join(predicted_text)
print(text)
translated = gc_translate(text)
sentiment = gc_sentiment(text)
print(translated)
print(sentiment)
text_dict = {"text": text,
"translated": translated['translatedText'],
"polarity": sentiment[0],
"magnitude": sentiment[1]}
a = text_dict
return render_template('result.html', prediction=a)
def get_ns(is_train):
network = BiLSTMNetwork(rnn_hidden_states=512,
rnn_layers=2,
max_seq_len=160,
ctx=ctx)
network.load_parameters("models/handwriting_line_sl_160_a_512_o_2.params",
ctx=ctx)
def noise_source_transform(image, text):
image = handwriting_recognition_transform(image)
output = network(image)
predict_probs = output.softmax().asnumpy()
return predict_probs
ns = Noisy_forms_dataset(noise_source_transform,
train=is_train,
name="OCR_noise2",
topK_decode=topK_decode)
return ns
#%%
train_dataset = FontDataset()
batch_size = 64
train_data = gluon.data.DataLoader(train_dataset.transform_first(gluon.data.vision.transforms.ToTensor()), batch_size=batch_size, last_batch='rollover', shuffle=True, num_workers=8)
#%%
plt.imshow(train_dataset._get_img('m"u'), cmap='gray')
#%%
handwriting_line_recognition_net = HandwritingRecognitionNet(rnn_hidden_states=512,
rnn_layers=2, ctx=ctx, max_seq_len=160)
handwriting_line_recognition_net.load_parameters("models/handwriting_line8.params", ctx=ctx)
#%%
im = np.array(train_dataset._get_img('cow'))
plt.imshow(im)
#%%
line_image_size = (60, 800)
line_image = handwriting_recognition_transform(np.array(im).mean(axis=2), line_image_size)
line_character_prob = handwriting_line_recognition_net(line_image.as_in_context(ctx))