【Keras实战】SegNet

一、描述

本文使用 Keras 实现了 SegNet 网络的整体架构
论文地址：SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation
代码地址：GitHub

二、keras 实现 SegNet

首先导入所需的函数包

# -*- coding: utf-8 -*-
"""
Created on Fri Apr 19 16:33:17 2019

@author: zhangyonghui
"""
from keras.models import Model
from keras.layers import Conv2D
from keras.layers import BatchNormalization, Activation, Input
from keras.optimizers import Adam
from keras import backend as K
from keras.layers.core import Layer
from keras.layers.convolutional import Convolution2D

1、定义 MaxPoolingWithArgmax：

通过此类记住最大池化时每个窗口的最大值的位置

class MaxPoolingWithArgmax2D(Layer):
    def __init__(self, pool_size=(2, 2), strides=(2, 2), padding='same', **kwargs):
        super(MaxPoolingWithArgmax2D, self).__init__(**kwargs)
        self.padding = padding
        self.pool_size = pool_size
        self.strides = strides

    def call(self, inputs, **kwargs):
        padding = self.padding
        pool_size = self.pool_size
        strides = self.strides
        if K.backend() == 'tensorflow':
            ksize = [1, pool_size[0], pool_size[1], 1]
            padding = padding.upper()
            strides = [1, strides[0], strides[1], 1]
            output, argmax = K.tf.nn.max_pool_with_argmax(inputs, ksize=ksize, strides=strides, padding=padding)
        else:
            errmsg = '{} backend is not supported for layer {}'.format(K.backend(), type(self).__name__)
            raise NotImplementedError(errmsg)
        argmax = K.cast(argmax, K.floatx())
        return [output, argmax]

    def compute_output_shape(self, input_shape):
        ratio = (1, 2, 2, 1)
        output_shape = [dim // ratio[idx] if dim is not None else None for idx, dim in enumerate(input_shape)]
        output_shape = tuple(output_shape)
        return [output_shape, output_shape]

    def compute_mask(self, inputs, mask=None):
        return 2 * [None]

2、定义 MaxUnpooling：

在还原特征大小时，在记忆的最大值的位置上还原该值，窗口其余的视野内均为0

class MaxUnpooling2D(Layer):
    def __init__(self, size=(2, 2), **kwargs):
        super(MaxUnpooling2D, self).__init__(**kwargs)
        self.size = size

    def call(self, inputs, output_shape=None):
        updates, mask = inputs[0], inputs[1]
        with K.tf.variable_scope(self.name):
            mask = K.cast(mask, 'int32')
            input_shape = K.tf.shape(updates, out_type='int32')
            #  calculation new shape
            if output_shape is None:
                output_shape = (input_shape[0], input_shape[1] * self.size[0], input_shape[2] * self.size[1], input_shape[3])
                self.output_shape1 = output_shape

        # calculation indices for batch, height, width and feature maps
        one_like_mask = K.ones_like(mask, dtype='int32')
        batch_shape = K.concatenate([[input_shape[0]], [1], [1], [1]], axis=0)
        batch_range = K.reshape(K.tf.range(output_shape[0], dtype='int32'), shape=batch_shape)
        b = one_like_mask * batch_range
        y = mask // (output_shape[2] * output_shape[3])
        x = (mask // output_shape[3]) % output_shape[2]
        feature_range = K.tf.range(output_shape[3], dtype='int32')
        f = one_like_mask * feature_range

        # transpose indices & reshape update values to one dimension
        updates_size = K.tf.size(updates)
        indices = K.transpose(K.reshape(K.stack([b, y, x, f]), [4, updates_size]))
        values = K.reshape(updates, [updates_size])
        ret = K.tf.scatter_nd(indices, values, output_shape)
        return ret

    def compute_output_shape(self, input_shape):
        mask_shape = input_shape[1]
        return mask_shape[0], mask_shape[1] * self.size[0], mask_shape[2] * self.size[1], mask_shape[3]

3、构建模型

def SegNet(input_shape, n_labels=2, kernel=3, pool_size=(2, 2), output_mode="softmax", model_summary=None):
    # encoder
    inputs = Input(shape=input_shape)

    conv_1 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(inputs)
    conv_1 = BatchNormalization()(conv_1)
    conv_1 = Activation("relu")(conv_1)
    conv_2 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_1)
    conv_2 = BatchNormalization()(conv_2)
    conv_2 = Activation("relu")(conv_2)

    pool_1, mask_1 = MaxPoolingWithArgmax2D(pool_size)(conv_2)

    conv_3 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(pool_1)
    conv_3 = BatchNormalization()(conv_3)
    conv_3 = Activation("relu")(conv_3)
    conv_4 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_3)
    conv_4 = BatchNormalization()(conv_4)
    conv_4 = Activation("relu")(conv_4)

    pool_2, mask_2 = MaxPoolingWithArgmax2D(pool_size)(conv_4)

    conv_5 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(pool_2)
    conv_5 = BatchNormalization()(conv_5)
    conv_5 = Activation("relu")(conv_5)
    conv_6 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_5)
    conv_6 = BatchNormalization()(conv_6)
    conv_6 = Activation("relu")(conv_6)
    conv_7 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_6)
    conv_7 = BatchNormalization()(conv_7)
    conv_7 = Activation("relu")(conv_7)

    pool_3, mask_3 = MaxPoolingWithArgmax2D(pool_size)(conv_7)

    conv_8 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(pool_3)
    conv_8 = BatchNormalization()(conv_8)
    conv_8 = Activation("relu")(conv_8)
    conv_9 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_8)
    conv_9 = BatchNormalization()(conv_9)
    conv_9 = Activation("relu")(conv_9)
    conv_10 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_9)
    conv_10 = BatchNormalization()(conv_10)
    conv_10 = Activation("relu")(conv_10)

    pool_4, mask_4 = MaxPoolingWithArgmax2D(pool_size)(conv_10)

    conv_11 = Convolution2D(2160, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(pool_4)
    conv_11 = BatchNormalization()(conv_11)
    conv_11 = Activation("relu")(conv_11)
    conv_12 = Convolution2D(2160, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_11)
    conv_12 = BatchNormalization()(conv_12)
    conv_12 = Activation("relu")(conv_12)
    conv_13 = Convolution2D(2160, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_12)
    conv_13 = BatchNormalization()(conv_13)
    conv_13 = Activation("relu")(conv_13)

    pool_5, mask_5 = MaxPoolingWithArgmax2D(pool_size)(conv_13)
    print("Build enceder done..")

    # decoder
    unpool_1 = MaxUnpooling2D(pool_size)([pool_5, mask_5])

    conv_14 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(unpool_1)
    conv_14 = BatchNormalization()(conv_14)
    conv_14 = Activation("relu")(conv_14)
    conv_15 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_14)
    conv_15 = BatchNormalization()(conv_15)
    conv_15 = Activation("relu")(conv_15)
    conv_16 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_15)
    conv_16 = BatchNormalization()(conv_16)
    conv_16 = Activation("relu")(conv_16)

    unpool_2 = MaxUnpooling2D(pool_size)([conv_16, mask_4])

    conv_17 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(unpool_2)
    conv_17 = BatchNormalization()(conv_17)
    conv_17 = Activation("relu")(conv_17)
    conv_18 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_17)
    conv_18 = BatchNormalization()(conv_18)
    conv_18 = Activation("relu")(conv_18)
    conv_19 = Convolution2D(1080, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_18)
    conv_19 = BatchNormalization()(conv_19)
    conv_19 = Activation("relu")(conv_19)

    unpool_3 = MaxUnpooling2D(pool_size)([conv_19, mask_3])

    conv_20 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(unpool_3)
    conv_20 = BatchNormalization()(conv_20)
    conv_20 = Activation("relu")(conv_20)
    conv_21 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_20)
    conv_21 = BatchNormalization()(conv_21)
    conv_21 = Activation("relu")(conv_21)
    conv_22 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_21)
    conv_22 = BatchNormalization()(conv_22)
    conv_22 = Activation("relu")(conv_22)

    unpool_4 = MaxUnpooling2D(pool_size)([conv_22, mask_2])

    conv_23 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(unpool_4)
    conv_23 = BatchNormalization()(conv_23)
    conv_23 = Activation("relu")(conv_23)
    conv_24 = Convolution2D(540, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(conv_23)
    conv_24 = BatchNormalization()(conv_24)
    conv_24 = Activation("relu")(conv_24)

    unpool_5 = MaxUnpooling2D(pool_size)([conv_24, mask_1])

    conv_25 = Convolution2D(270, (kernel, kernel), padding="same", kernel_initializer = 'he_normal')(unpool_5)
    conv_25 = BatchNormalization()(conv_25)
    conv_25 = Activation("relu")(conv_25)

    out=Conv2D(n_labels,1, activation = output_mode)(conv_25)
    print("Build decoder done..")

    model = Model(inputs=inputs, outputs=out, name="SegNet")
    model.compile(optimizer = Adam(lr = 1e-4), loss = 'categorical_crossentropy', metrics = ['categorical_accuracy'])

    if model_summary is True:
        model.summary()

    return model

4、输出模型的整体结构

if __name__ == '__main__':
    SegNet(input_shape=(64, 64, 270), n_labels=4, model_summary=True)

运行结果：

祝您浏览愉快！