def preprocess(self, image):
image = resize(image,
(self.inp_shape[1],
self.inp_shape[2])) # rescale to match the input size
# change to grayscale if the model is in grayscale
if self.inp_shape[3] == 1 and image.shape[2] != 1:
if image.shape[2] == 3:
image = rgb_to_grayscale(image)
elif image.shape[2] == 4:
image = rgb_to_grayscale(image[:, :, :3])
image = img_to_array(image)
image = np.expand_dims(image, axis=0)
return image
def __random_circles_image(fig, circle_num, get_radius=__get_radius):
ax = fig.add_axes([0, 0, 1, 1], frameon=False)
ax.set_xlim(0, SIDE_LIMIT)
ax.set_ylim(0, SIDE_LIMIT)
ax.axis('off')
circle_params = []
for _ in range(circle_num):
radius = get_radius()
center = __random_center(circle_params, radius)
circle_param = {'r': radius, 'c': center}
circle_params.append(circle_param)
circle = ptchs.Circle(center, radius, fill=False)
ax.add_artist(circle)
canvas = fig.canvas
canvas.draw()
fig_size = fig.get_size_inches()
fig_dpi = fig.get_dpi()
width, height = fig_size * fig_dpi
data = np.fromstring(canvas.tostring_rgb(), np.uint8).reshape(
(int(height), int(width), 3))
data = tfimg.rgb_to_grayscale(data)
data = np.squeeze(data)
fig.clf()
return data
def grayscale_and_resize(PIL, shape=(256, 256), padding=False, grayscale=True):
"""
This is the preprocessing function that will take the raw jpeg, gray scale it, resize it and
turn it into an array
PIL --> PIL object
shape --> tuple: size of the final array
padding --> bool: if True, will use tf.resize_with_pad
"""
if padding:
gray_image = rgb_to_grayscale(PIL)
resized_image_arr = resize_with_pad(gray_image,
target_height=shape[0],
target_width=shape[1])
else:
if grayscale:
resized_image_arr = img_to_array(
PIL.convert(mode='L').resize(shape))
else:
resized_image_arr = img_to_array(PIL.resize(shape))
return resized_image_arr
hsv_image = cv2.cvtColor(frame3, cv2.COLOR_BGR2HSV)
hsv_image[:, :, 2] = cv2.equalizeHist(hsv_image[:, :, 2])
histEqualized_image = cv2.cvtColor(hsv_image, cv2.COLOR_HSV2BGR)
cv2.putText(histEqualized_image, 'Histogram equalization', (10, 30),
font, 1, (0, 0, 255), 1)
result3 = histEqualized_image
###########Gray#############
frame4 = cv2.resize(frame4, (width // 2, height // 2))
enable_eager_execution()
image_data = rgb_to_grayscale(frame4)
image_data_array = image_data.numpy()
result4 = cv2.cvtColor(image_data_array, cv2.COLOR_GRAY2BGR)
cv2.putText(result4, 'Gray', (10, 30), font, 1, (0, 0, 255), 1)
###########################
up_frame = np.hstack((frame1, result2))
down_frame = np.hstack((result3, result4))
result_video = np.vstack((up_frame, down_frame))
out.write(result_video)
else:
break
from tflearn import fully_connected, regression, input_data, dropout
from tensorflow import image
FOLDER = '/home/jer/Workspace/cs5600/project1/trained_nets/'
NAME = 'ANN_Bee2_1S_5Layer_relu_grayscale'
SAVE_POINT = f'{FOLDER}' + f'{NAME}' + '.tfl'
CHECK_POINT = SAVE_POINT + '.meta'
beeX, beeY = tflearn.data_utils.image_preloader('BEE2_1S/class_labels.txt',
image_shape=(90, 90),
mode='file',
categorical_labels=True,
normalize=True)
input_layer = input_data(shape=[None, 90, 90, 3])
input_layer = image.rgb_to_grayscale(input_layer)
fc_layer_1 = fully_connected(input_layer,
8100,
activation='relu',
name='fc_layer_1')
dropout_1 = dropout(fc_layer_1, 0.5)
fc_layer_2 = fully_connected(dropout_1,
4000,
activation='relu',
name='fc_layer_2')
dropout_2 = dropout(fc_layer_2, 0.5)
fc_layer_3 = fully_connected(dropout_2,
500,
activation='relu',
name='fc_layer_3')
dropout_3 = dropout(fc_layer_3, 0.5)
def rgb255_to_obj_net(rgb):
input_x_ = ContrastNoise(0.25)(rgb)
input_x_ = ContrastNoiseSingle(0.25)(input_x_)
input_x_ = GaussianNoise(.08 * 128)(input_x_)
input_x_ = BrightnessNoise(0.2 * 128)(input_x_)
rgb = Lambda(lambda x: rgb_to_grayscale(x) / 255.)(input_x_)
x = Conv2D(8, 5, padding='same', activation='selu')(rgb)
tier0 = x
x = Conv2D(16, 5, strides=2, padding='same', activation='selu')(x)
x = dense_block(x, 32, 3)
tier1 = x
x = Conv2D(64, 5, strides=2, padding='same', activation='selu')(x)
x = dense_block(x, 64, 6)
x = dense_block(x, 64, 6)
tier2 = x
x = Conv2D(128, 5, strides=2, padding='same', activation='selu')(x)
x = dense_block(x, 64, 12)
x = dense_block(x, 64, 12)
tier3 = x
x = Conv2D(128, 5, strides=2, padding='same', activation='selu')(x)
x = dense_block(x, 128, 12)
def up_path(x):
x = UpSampling2D()(x)
x = Concatenate()([x, tier3])
x = Conv2D(64, 3, padding='same', activation='selu')(x)
x = dense_block(x, 32, 12)
x = UpSampling2D()(x)
x = Concatenate()([x, tier2])
x = Conv2D(32, 3, padding='same', activation='selu')(x)
x = dense_block(x, 16, 6)
x = UpSampling2D()(x)
x = Concatenate()([x, tier1])
x = Conv2D(24, 3, padding='same', activation='selu')(x)
x = dense_block(x, 12, 4)
return x
star_x = up_path(x)
star = Conv2D(3, 1, padding='same', activation=None, name='star')(star_x)
dash_x = up_path(x)
dash = Conv2D(3, 1, padding='same', activation=None, name='dash')(dash_x)
w_px_x = up_path(x)
w_px = Conv2D(4, 1, padding='same', activation=None, name='wpx')(w_px_x)
w_d_x = up_path(x)
w_d = Conv2D(1, 1, padding='same', activation=None, name='wd')(w_d_x)
seg_x = up_path(x)
seg = Conv2D(1, 1, padding='same', activation='sigmoid')(seg_x)
return star, dash, w_px, w_d, seg