def _pre_process_image(img_in, _parse_method):
im = cv2.imread(img_in, cv2.IMREAD_GRAYSCALE)
if np.size(im) == 1: # skip if the image data is corrupt.
return None
# reduce the size of form images so that it can fit in memory.
if _parse_method in ["form", "form_bb"]:
im, _ = resize_image(im, MAX_IMAGE_SIZE_FORM)
if _parse_method == "line":
im, _ = resize_image(im, MAX_IMAGE_SIZE_LINE)
if _parse_method == "word":
im, _ = resize_image(im, MAX_IMAGE_SIZE_WORD)
img_arr = np.asarray(im)
return img_arr
def handwriting_recognition_transform(image):
image, _ = resize_image(image, line_image_size)
image = mx.nd.array(image) / 255.
image = (image - 0.942532484060557) / 0.15926149044640417
image = image.as_in_context(ctx)
image = image.expand_dims(0).expand_dims(0)
return image
def __getitem__(self, idx):
filename, text = self.data[idx]
image = cv2.imread(filename, 0)
return resize_image(np.expand_dims(image, axis=2), (60, 200))[0], text