|
|
|
|
|
from cProfile import label
|
|
import gradio as gr
|
|
import numpy as np
|
|
|
|
|
|
import tensorflow as tf
|
|
import cv2
|
|
|
|
tf.config.experimental.set_visible_devices([], 'GPU')
|
|
|
|
|
|
COLOR = np.array([163, 23, 252])/255.0
|
|
ALPHA = 0.8
|
|
|
|
|
|
def parse_image(image):
|
|
image = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
|
|
image = cv2.resize(image, (100, 100))
|
|
image = image.astype(np.float32)
|
|
image = image / 255.0
|
|
image = np.expand_dims(image, axis=0)
|
|
image = np.expand_dims(image, axis=-1)
|
|
return image
|
|
|
|
|
|
|
|
def cnn(input_shape, output_shape):
|
|
num_classes = output_shape[0]
|
|
dropout_seed = 708090
|
|
kernel_seed = 42
|
|
|
|
model = tf.keras.models.Sequential([
|
|
tf.keras.layers.Conv2D(16, 3, activation='relu', input_shape=input_shape, kernel_initializer=tf.keras.initializers.GlorotUniform(seed=kernel_seed)),
|
|
tf.keras.layers.MaxPooling2D(),
|
|
tf.keras.layers.Dropout(0.1, seed=dropout_seed),
|
|
tf.keras.layers.Conv2D(32, 5, activation='relu', kernel_initializer=tf.keras.initializers.GlorotUniform(seed=kernel_seed)),
|
|
tf.keras.layers.MaxPooling2D(),
|
|
tf.keras.layers.Dropout(0.1, seed=dropout_seed),
|
|
tf.keras.layers.Conv2D(64, 10, activation='relu', kernel_initializer=tf.keras.initializers.GlorotUniform(seed=kernel_seed)),
|
|
tf.keras.layers.MaxPooling2D(),
|
|
tf.keras.layers.Dropout(0.1, seed=dropout_seed),
|
|
tf.keras.layers.Flatten(),
|
|
tf.keras.layers.Dense(128, activation='relu', kernel_regularizer='l2', kernel_initializer=tf.keras.initializers.GlorotUniform(seed=kernel_seed)),
|
|
tf.keras.layers.Dropout(0.2, seed=dropout_seed),
|
|
tf.keras.layers.Dense(16, activation='relu', kernel_regularizer='l2', kernel_initializer=tf.keras.initializers.GlorotUniform(seed=kernel_seed)),
|
|
tf.keras.layers.Dropout(0.2, seed=dropout_seed),
|
|
tf.keras.layers.Dense(num_classes, activation='sigmoid', kernel_initializer=tf.keras.initializers.GlorotUniform(seed=kernel_seed))
|
|
])
|
|
|
|
return model
|
|
|
|
|
|
model = cnn((100, 100, 1), (1,))
|
|
model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=False), optimizer='Adam', metrics='accuracy')
|
|
|
|
model.load_weights('weights.h5')
|
|
|
|
|
|
|
|
def saliency_map(img):
|
|
"""
|
|
return the normalized gradients overs the image, and also the prediction of the model
|
|
"""
|
|
inp = tf.convert_to_tensor(
|
|
img[None, :, :, None],
|
|
dtype = tf.float32
|
|
)
|
|
inp_var = tf.Variable(inp)
|
|
|
|
with tf.GradientTape() as tape:
|
|
pred = model(inp_var, training=False)
|
|
loss = pred[0][0]
|
|
grads = tape.gradient(loss, inp_var)
|
|
grads = tf.math.abs(grads) / (tf.math.reduce_max(tf.math.abs(grads))+1e-14)
|
|
return grads, round(float(model(inp_var, training = False)))
|
|
|
|
|
|
def segment(image):
|
|
|
|
print(image.shape)
|
|
image = parse_image(image)
|
|
print(image.shape)
|
|
output = model.predict(image)
|
|
|
|
labels = {
|
|
"Farsi" : 1-float(output),
|
|
"Ruqaa" : float(output)
|
|
}
|
|
grads, _ = saliency_map(image[0, :, :, 0])
|
|
s_map = grads.numpy()[0, :, :, 0]
|
|
reconstructed_image = cv2.cvtColor(image.squeeze(0), cv2.COLOR_GRAY2RGB)
|
|
for i in range(reconstructed_image.shape[0]):
|
|
for j in range(reconstructed_image.shape[1]):
|
|
reconstructed_image[i, j, :] = reconstructed_image[i, j, :] * (1-ALPHA) + s_map[i, j]* COLOR * ALPHA
|
|
|
|
V = reconstructed_image
|
|
|
|
|
|
return labels, reconstructed_image
|
|
|
|
iface = gr.Interface(fn=segment,
|
|
description="""
|
|
This is an Arab Calligraphy Style Recognition.
|
|
This model predicts the style (binary classification) of the image.
|
|
The model also outputs the Saliency map.
|
|
""",
|
|
inputs="image",
|
|
outputs=[
|
|
gr.outputs.Label(num_top_classes=2, label="Style"),
|
|
gr.outputs.Image(label = "Saliency map")
|
|
],
|
|
examples=[["images/Farsi_1.jpg"],
|
|
["images/Farsi_2.jpg"],
|
|
["images/real_Farsi.jpg"],
|
|
["images/Ruqaa_1.jpg"],
|
|
["images/Ruqaa_2.jpg"],
|
|
["images/Ruqaa_3.jpg"],
|
|
["images/real_Ruqaa.jpg"],
|
|
]).launch()
|
|
|
|
|
