Fix non-transparent background in handwritten

Author: Belval

.travis.yml

@@ -2,7 +2,7 @@ language: python
 python:
   - "3.5"
 install:
-  - pip install -r requirements.txt
+  - pip install -r requirements-hw.txt
   - pip install codecov
 script:
   - python3 tests.py

TextRecognitionDataGenerator/background_generator.py

@@ -19,15 +19,15 @@ def gaussian_noise(cls, height, width):
         # We add gaussian noise
         cv2.randn(image, 235, 10)
 
-        return Image.fromarray(image).convert('RGB')
+        return Image.fromarray(image).convert('RGBA')
 
     @classmethod
     def plain_white(cls, height, width):
         """
             Create a plain white background
         """
 
-        return Image.new("L", (width, height), 255).convert('RGB')
+        return Image.new("L", (width, height), 255).convert('RGBA')
 
     @classmethod
     def quasicrystal(cls, height, width):
@@ -53,7 +53,7 @@ def quasicrystal(cls, height, width):
                     z += math.cos(r * math.sin(a) * frequency + phase)
                 c = int(255 - round(255 * z / rotation_count))
                 pixels[kw, kh] = c # grayscale
-        return image.convert('RGB')
+        return image.convert('RGBA')
 
     @classmethod
     def picture(cls, height, width):

TextRecognitionDataGenerator/handwritten_text_generator.py

@@ -90,11 +90,11 @@ def __sample_text(cls, sess, args_text, translation):
     @classmethod
     def __crop_white_borders(cls, image):
         image_data = np.asarray(image)
-        non_empty_columns = np.where(image_data.min(axis=0)<255)[0]
-        non_empty_rows = np.where(image_data.min(axis=1)<255)[0]
+        grey_image_data = np.asarray(image.convert('L'))
+        non_empty_columns = np.where(grey_image_data.min(axis=0) < 255)[0]
+        non_empty_rows = np.where(grey_image_data.min(axis=1) < 255)[0]
         cropBox = (min(non_empty_rows), max(non_empty_rows), min(non_empty_columns), max(non_empty_columns))
-
-        image_data_new = image_data[cropBox[0]:cropBox[1]+1, cropBox[2]:cropBox[3]+1]
+        image_data_new = image_data[cropBox[0]:cropBox[1]+1, cropBox[2]:cropBox[3]+1, :]
 
         return Image.fromarray(image_data_new)
 
@@ -105,7 +105,7 @@ def __join_images(cls, images):
         total_width = sum(widths) - 35 * len(images)
         max_height = max(heights)
 
-        compound_image = Image.new('L', (total_width, max_height))
+        compound_image = Image.new('RGBA', (total_width, max_height))
 
         x_offset = 0
         for im in images:
@@ -142,10 +142,13 @@ def generate(cls, text):
                 for stroke in cls.__split_strokes(cls.__cumsum(np.array(coords))):
                     plt.plot(stroke[:, 0], -stroke[:, 1], color='#080808')
 
+                fig.patch.set_alpha(0)
+                fig.patch.set_facecolor('none')
+
                 canvas = plt.get_current_fig_manager().canvas
                 canvas.draw()
 
-                image = Image.frombytes('RGB', canvas.get_width_height(), canvas.tostring_rgb()).convert('L')
+                image = Image.frombytes('RGBA', canvas.get_width_height(), canvas.buffer_rgba())
                 images.append(cls.__crop_white_borders(image))
 
                 plt.close()

requirements-hw.txt

@@ -0,0 +1,9 @@
+beautifulsoup4==4.6.0
+numpy==1.15.1
+opencv-python==4.0.0.21
+tqdm==4.23.4
+Pillow==5.1.0
+requests==2.20.0
+tensorflow==1.13.1
+matplotlib==3.0.2
+seaborn==0.9.0

tests.py

@@ -512,7 +512,7 @@ def test_generate_chinese_string(self):
         )
 
     def test_generate_data_with_white_background(self):
-        BackgroundGenerator.plain_white(64, 128).save('tests/out/white_background.jpg')
+        BackgroundGenerator.plain_white(64, 128).convert('RGB').save('tests/out/white_background.jpg')
 
         self.assertTrue(
             md5('tests/out/white_background.jpg') == md5('tests/expected_results/white_background.jpg')
@@ -521,7 +521,7 @@ def test_generate_data_with_white_background(self):
         os.remove('tests/out/white_background.jpg')
 
     def test_generate_data_with_gaussian_background(self):
-        BackgroundGenerator.gaussian_noise(64, 128).save('tests/out/gaussian_background.jpg')
+        BackgroundGenerator.gaussian_noise(64, 128).convert('RGB').save('tests/out/gaussian_background.jpg')
 
         self.assertTrue(
             md5('tests/out/gaussian_background.jpg') == md5('tests/expected_results/gaussian_background.jpg')
@@ -598,6 +598,12 @@ def test_random_sequences_symbols_only(self):
             self.assertTrue(all([c in "!\"#$%&'()*+,-./:;?@[\\]^_`{|}~" for c in f.readline().split(' ')[1][:-1]]))
         empty_directory('tests/out/')
 
+    def test_handwritten(self):
+        args = ['python3', 'run.py', '-c', '1', '--output_dir', '../tests/out/']
+        subprocess.Popen(args, cwd="TextRecognitionDataGenerator/").wait()
+        self.assertTrue(len(os.listdir('tests/out/')) == 1)
+        empty_directory('tests/out/')
+
 #    def test_word_count(self):
 #        args = ['python3', 'run.py', '-c', '1', '-w', '5']
 #        subprocess.Popen(args, cwd="TextRecognitionDataGenerator/").wait()