# MIT License # # Copyright (c) 2019-2024 Iván de Paz Centeno # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. # pylint: disable=duplicate-code import tensorflow as tf L = tf.keras.layers class ONet(tf.keras.Model): """ Definition of ONet (Output Network) for MTCNN. This network takes as input an image of size 48x48 with 3 channels, and outputs: * The regression of the bounding boxes (x1, y1, x2, y2) with a linear activation. * The regression of 5 facial landmarks (10 points total). * The classification of the area as a softmax operation ([1, 0] -> Not face; [0, 1] -> Face). """ def __init__(self, **kwargs): super(ONet, self).__init__(**kwargs) # Defining the layers according to the provided architecture self.conv1 = L.Conv2D(32, kernel_size=(3, 3), strides=(1, 1), padding="valid", activation="linear", name="conv1") self.prelu1 = L.PReLU(shared_axes=[1, 2], name="prelu1") self.maxpool1 = L.MaxPooling2D(pool_size=(3, 3), strides=(2, 2), padding="same", name="maxpooling1") self.conv2 = L.Conv2D(64, kernel_size=(3, 3), strides=(1, 1), padding="valid", activation="linear", name="conv2") self.prelu2 = L.PReLU(shared_axes=[1, 2], name="prelu2") self.maxpool2 = L.MaxPooling2D(pool_size=(3, 3), strides=(2, 2), padding="valid", name="maxpooling2") self.conv3 = L.Conv2D(64, kernel_size=(3, 3), strides=(1, 1), padding="valid", activation="linear", name="conv3") self.prelu3 = L.PReLU(shared_axes=[1, 2], name="prelu3") self.maxpool3 = L.MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding="same", name="maxpooling3") self.conv4 = L.Conv2D(128, kernel_size=(2, 2), strides=(1, 1), padding="valid", activation="linear", name="conv4") self.prelu4 = L.PReLU(shared_axes=[1, 2], name="prelu4") self.permute = L.Permute((2, 1, 3), name="permute") self.flatten = L.Flatten(name="flatten4") self.fc5 = L.Dense(256, activation="linear", name="fc5") self.prelu5 = L.PReLU(name="prelu5") self.fc6_1 = L.Dense(4, activation="linear", name="fc6-1") # Bounding box regression self.fc6_2 = L.Dense(10, activation="linear", name="fc6-2") # Landmark regression (5 landmarks, 10 points total) self.fc6_3 = L.Dense(2, activation="softmax", name="fc6-3") # Classification (face or not) def build(self, input_shape=(None, 48, 48, 3)): """ Build the network by defining the input and manually creating each layer step by step, computing output shapes. This method mirrors the layer initialization in the functional API. """ # Build conv1 block self.conv1.build(input_shape) output_shape = self.conv1.compute_output_shape(input_shape) self.prelu1.build(output_shape) output_shape = self.prelu1.compute_output_shape(output_shape) self.maxpool1.build(output_shape) output_shape = self.maxpool1.compute_output_shape(output_shape) # Build conv2 block self.conv2.build(output_shape) output_shape = self.conv2.compute_output_shape(output_shape) self.prelu2.build(output_shape) output_shape = self.prelu2.compute_output_shape(output_shape) self.maxpool2.build(output_shape) output_shape = self.maxpool2.compute_output_shape(output_shape) # Build conv3 block self.conv3.build(output_shape) output_shape = self.conv3.compute_output_shape(output_shape) self.prelu3.build(output_shape) output_shape = self.prelu3.compute_output_shape(output_shape) self.maxpool3.build(output_shape) output_shape = self.maxpool3.compute_output_shape(output_shape) # Build conv4 block self.conv4.build(output_shape) output_shape = self.conv4.compute_output_shape(output_shape) self.prelu4.build(output_shape) output_shape = self.prelu4.compute_output_shape(output_shape) # Permute and flatten self.permute.build(output_shape) output_shape = self.permute.compute_output_shape(output_shape) self.flatten.build(output_shape) output_shape = self.flatten.compute_output_shape(output_shape) # Fully connected layers self.fc5.build(output_shape) output_shape = self.fc5.compute_output_shape(output_shape) self.prelu5.build(output_shape) output_shape = self.prelu5.compute_output_shape(output_shape) # Outputs (classification, bounding box regression, and landmark regression) self.fc6_1.build(output_shape) self.fc6_2.build(output_shape) self.fc6_3.build(output_shape) # Call the super build to finalize the model building super(ONet, self).build(input_shape) def call(self, inputs, *args, **kwargs): x = inputs # First conv block x = self.conv1(x) x = self.prelu1(x) x = self.maxpool1(x) # Second conv block x = self.conv2(x) x = self.prelu2(x) x = self.maxpool2(x) # Third conv block x = self.conv3(x) x = self.prelu3(x) x = self.maxpool3(x) # Fourth conv block x = self.conv4(x) x = self.prelu4(x) # Permute, flatten, and fully connected layers x = self.permute(x) x = self.flatten(x) x = self.fc5(x) x = self.prelu5(x) # Outputs bbox_reg = self.fc6_1(x) # Regression of bounding boxes landmarks = self.fc6_2(x) # Regression of facial landmarks bbox_class = self.fc6_3(x) # Classification (face or not) return [bbox_reg, landmarks, bbox_class]