基于深度学习神经网络MTCNN的人脸检测系统源码

AI街潜水的八角

已于 2024-01-19 15:21:45 修改

阅读量678

点赞数 9

分类专栏：目标检测实战文章标签：深度学习神经网络人工智能

于 2024-01-19 11:36:39 首次发布

本文链接：https://blog.csdn.net/m0_59023219/article/details/135693271

版权

目标检测实战专栏收录该内容

16 篇文章 0 订阅

订阅专栏

一. mtcnn概述

MTCNN，英文全称是Multi-task convolutional neural network，中文全称是多任务卷积神经网络，该神经网络将人脸区域检测与人脸关键点检测放在了一起。

二. mtcnn的网络结构

mtcnn从整体上划分分为P-Net、R-Net、和O-Net三层网络结构。各层的作用直观上感受如下图所示：

其网络结构：

三. mtcnn的网络结构代码

import tensorflow as tf


class PNet(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.conv1 = tf.keras.layers.Conv2D(10, 3, 1, name='conv1')
        self.prelu1 = tf.keras.layers.PReLU(shared_axes=[1,2], name="PReLU1")

        self.conv2 = tf.keras.layers.Conv2D(16, 3, 1, name='conv2')
        self.prelu2 = tf.keras.layers.PReLU(shared_axes=[1,2], name="PReLU2")

        self.conv3 = tf.keras.layers.Conv2D(32, 3, 1, name='conv3')
        self.prelu3 = tf.keras.layers.PReLU(shared_axes=[1,2], name="PReLU3")

        self.conv4_1 = tf.keras.layers.Conv2D(2, 1, 1, name='conv4-1')
        self.conv4_2 = tf.keras.layers.Conv2D(4, 1, 1, name='conv4-2')

    def call(self, x, training=False):
        out = self.prelu1(self.conv1(x))
        out = tf.nn.max_pool2d(out, 2, 2, padding="SAME")
        out = self.prelu2(self.conv2(out))
        out = self.prelu3(self.conv3(out))
        score = tf.nn.softmax(self.conv4_1(out), axis=-1)
        boxes = self.conv4_2(out)
        return boxes, score

class RNet(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.conv1 = tf.keras.layers.Conv2D(28, 3, 1, name='conv1')
        self.prelu1 = tf.keras.layers.PReLU(shared_axes=[1,2], name="prelu1")

        self.conv2 = tf.keras.layers.Conv2D(48, 3, 1, name='conv2')
        self.prelu2 = tf.keras.layers.PReLU(shared_axes=[1,2], name="prelu2")

        self.conv3 = tf.keras.layers.Conv2D(64, 2, 1, name='conv3')
        self.prelu3 = tf.keras.layers.PReLU(shared_axes=[1,2], name="prelu3")

        self.dense4 = tf.keras.layers.Dense(128, name='conv4')
        self.prelu4 = tf.keras.layers.PReLU(shared_axes=None, name="prelu4")

        self.dense5_1 = tf.keras.layers.Dense(2, name="conv5-1")
        self.dense5_2 = tf.keras.layers.Dense(4, name="conv5-2")

        self.flatten = tf.keras.layers.Flatten()

    def call(self, x, training=False):
        out = self.prelu1(self.conv1(x))
        out = tf.nn.max_pool2d(out, 3, 2, padding="SAME")
        out = self.prelu2(self.conv2(out))
        out = tf.nn.max_pool2d(out, 3, 2, padding="VALID")
        out = self.prelu3(self.conv3(out))
        out = self.flatten(out)
        out = self.prelu4(self.dense4(out))
        score = tf.nn.softmax(self.dense5_1(out), -1)
        boxes = self.dense5_2(out)
        return boxes, score

class ONet(tf.keras.Model):
    def __init__(self):
        super().__init__()

        self.conv1 = tf.keras.layers.Conv2D(32, 3, 1, name="conv1")
        self.prelu1 = tf.keras.layers.PReLU(shared_axes=[1,2], name="prelu1")

        self.conv2 = tf.keras.layers.Conv2D(64, 3, 1, name="conv2")
        self.prelu2 = tf.keras.layers.PReLU(shared_axes=[1,2], name="prelu2")

        self.conv3 = tf.keras.layers.Conv2D(64, 3, 1, name="conv3")
        self.prelu3 = tf.keras.layers.PReLU(shared_axes=[1,2], name="prelu3")

        self.conv4 = tf.keras.layers.Conv2D(128, 2, 1, name="conv4")
        self.prelu4 = tf.keras.layers.PReLU(shared_axes=[1,2], name="prelu4")

        self.dense5 = tf.keras.layers.Dense(256, name="conv5")
        self.prelu5 = tf.keras.layers.PReLU(shared_axes=None, name="prelu5")

        self.dense6_1 = tf.keras.layers.Dense(2  , name="conv6-1")
        self.dense6_2 = tf.keras.layers.Dense(4  , name="conv6-2")
        self.dense6_3 = tf.keras.layers.Dense(10 , name="conv6-3")

        self.flatten = tf.keras.layers.Flatten()

    def call(self, x, training=False):
        out = self.prelu1(self.conv1(x))
        out = tf.nn.max_pool2d(out, 3, 2, padding="SAME")
        out = self.prelu2(self.conv2(out))
        out = tf.nn.max_pool2d(out, 3, 2, padding="VALID")
        out = self.prelu3(self.conv3(out))
        out = tf.nn.max_pool2d(out, 2, 2, padding="SAME")
        out = self.prelu4(self.conv4(out))


        out = self.dense5(self.flatten(out))
        out = self.prelu5(out)
        score = tf.nn.softmax(self.dense6_1(out))
        boxes = self.dense6_2(out)
        lamks = self.dense6_3(out)
        return boxes, lamks, score

四. mtcnn的演示效果