【DFL学习日记#3】S3FDExtractor.py 检测人脸边界框

LeoSasion · 2025-10-3 11:49:54

【学习日记】Extract 切脸 - Deep 换脸 - BBS_Monster
Extractor.py 包含了本文的S3FD（边界框）和FANExtractor（关键点）
本文只讲S3FD

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
S3FD人脸检测器实现
该模块实现了Single Shot Scale-invariant Face Detector (S3FD)人脸检测算法，
用于检测图像中的人脸并返回人脸边界框坐标。
"""
import operator
from pathlib import Path
import cv2
import numpy as np
from core.leras import nn
class S3FDExtractor(object):
"""
S3FD人脸检测器提取器类
该类封装了S3FD人脸检测模型，提供了加载模型和从图像中提取人脸边界框的功能。
"""
def __init__(self, place_model_on_cpu=False):
"""
初始化S3FD人脸检测器
参数:
place_model_on_cpu (bool): 是否将模型放在CPU上运行，默认为False（使用GPU）
"""
# 初始化神经网络环境，使用NHWC数据格式
nn.initialize(data_format="NHWC")
tf = nn.tf
# 模型路径设置
model_path = Path(__file__).parent / "S3FD.npy"
if not model_path.exists():
raise Exception("Unable to load S3FD.npy")
class L2Norm(nn.LayerBase):
"""
L2归一化层
对输入特征图进行L2归一化处理，并应用可学习的缩放参数
"""
def __init__(self, n_channels, **kwargs):
self.n_channels = n_channels
super().__init__(**kwargs)
def build_weights(self):
# 创建可学习的缩放权重参数
self.weight = tf.get_variable("weight", (1, 1, 1, self.n_channels),
dtype=nn.floatx, initializer=tf.initializers.ones)
def get_weights(self):
return [self.weight]
def __call__(self, inputs):
"""
执行L2归一化操作
参数:
inputs: 输入特征图
返回:
归一化并缩放后的特征图
"""
x = inputs
# L2归一化: x / sqrt(sum(x^2) + 1e-10) * weight
x = x / (tf.sqrt(tf.reduce_sum(tf.pow(x, 2), axis=-1, keepdims=True)) + 1e-10) * self.weight
return x
class S3FD(nn.ModelBase):
"""
S3FD人脸检测模型类
实现了S3FD网络架构，包含特征提取和多尺度检测头
"""
def __init__(self):
super().__init__(name='S3FD')
def on_build(self):
"""
构建网络层结构
定义VGG风格的特征提取网络和多尺度检测头
"""
# 图像预处理：均值减法
self.minus = tf.constant([104, 117, 123], dtype=nn.floatx)
# VGG风格的卷积层 - 第一阶段
self.conv1_1 = nn.Conv2D(3, 64, kernel_size=3, strides=1, padding='SAME')
self.conv1_2 = nn.Conv2D(64, 64, kernel_size=3, strides=1, padding='SAME')
# 第二阶段
self.conv2_1 = nn.Conv2D(64, 128, kernel_size=3, strides=1, padding='SAME')
self.conv2_2 = nn.Conv2D(128, 128, kernel_size=3, strides=1, padding='SAME')
# 第三阶段
self.conv3_1 = nn.Conv2D(128, 256, kernel_size=3, strides=1, padding='SAME')
self.conv3_2 = nn.Conv2D(256, 256, kernel_size=3, strides=1, padding='SAME')
self.conv3_3 = nn.Conv2D(256, 256, kernel_size=3, strides=1, padding='SAME')
# 第四阶段
self.conv4_1 = nn.Conv2D(256, 512, kernel_size=3, strides=1, padding='SAME')
self.conv4_2 = nn.Conv2D(512, 512, kernel_size=3, strides=1, padding='SAME')
self.conv4_3 = nn.Conv2D(512, 512, kernel_size=3, strides=1, padding='SAME')
# 第五阶段
self.conv5_1 = nn.Conv2D(512, 512, kernel_size=3, strides=1, padding='SAME')
self.conv5_2 = nn.Conv2D(512, 512, kernel_size=3, strides=1, padding='SAME')
self.conv5_3 = nn.Conv2D(512, 512, kernel_size=3, strides=1, padding='SAME')
# 第六阶段 (卷积替代全连接层)
self.fc6 = nn.Conv2D(512, 1024, kernel_size=3, strides=1, padding=3)
self.fc7 = nn.Conv2D(1024, 1024, kernel_size=1, strides=1, padding='SAME')
# 第七阶段
self.conv6_1 = nn.Conv2D(1024, 256, kernel_size=1, strides=1, padding='SAME')
self.conv6_2 = nn.Conv2D(256, 512, kernel_size=3, strides=2, padding='SAME')
# 第八阶段
self.conv7_1 = nn.Conv2D(512, 128, kernel_size=1, strides=1, padding='SAME')
self.conv7_2 = nn.Conv2D(128, 256, kernel_size=3, strides=2, padding='SAME')
# L2归一化层用于不同尺度特征
self.conv3_3_norm = L2Norm(256)
self.conv4_3_norm = L2Norm(512)
self.conv5_3_norm = L2Norm(512)
# 多尺度检测头 - 用于检测不同大小的人脸
# conv3_3层检测头 (小尺度人脸)
self.conv3_3_norm_mbox_conf = nn.Conv2D(256, 4, kernel_size=3, strides=1, padding='SAME')
self.conv3_3_norm_mbox_loc = nn.Conv2D(256, 4, kernel_size=3, strides=1, padding='SAME')
# conv4_3层检测头
self.conv4_3_norm_mbox_conf = nn.Conv2D(512, 2, kernel_size=3, strides=1, padding='SAME')
self.conv4_3_norm_mbox_loc = nn.Conv2D(512, 4, kernel_size=3, strides=1, padding='SAME')
# conv5_3层检测头
self.conv5_3_norm_mbox_conf = nn.Conv2D(512, 2, kernel_size=3, strides=1, padding='SAME')
self.conv5_3_norm_mbox_loc = nn.Conv2D(512, 4, kernel_size=3, strides=1, padding='SAME')
# fc7层检测头
self.fc7_mbox_conf = nn.Conv2D(1024, 2, kernel_size=3, strides=1, padding='SAME')
self.fc7_mbox_loc = nn.Conv2D(1024, 4, kernel_size=3, strides=1, padding='SAME')
# conv6_2层检测头
self.conv6_2_mbox_conf = nn.Conv2D(512, 2, kernel_size=3, strides=1, padding='SAME')
self.conv6_2_mbox_loc = nn.Conv2D(512, 4, kernel_size=3, strides=1, padding='SAME')
# conv7_2层检测头 (大尺度人脸)
self.conv7_2_mbox_conf = nn.Conv2D(256, 2, kernel_size=3, strides=1, padding='SAME')
self.conv7_2_mbox_loc = nn.Conv2D(256, 4, kernel_size=3, strides=1, padding='SAME')
def forward(self, inp):
"""
前向传播
参数:
inp: 输入数据，包含一个图像张量
返回:
各尺度的分类和回归结果列表
"""
x, = inp
# 图像预处理：减去均值
x = x - self.minus
# 第一阶段卷积和池化
x = tf.nn.relu(self.conv1_1(x))
x = tf.nn.relu(self.conv1_2(x))
x = tf.nn.max_pool(x, [1, 2, 2, 1], [1, 2, 2, 1], "VALID")
# 第二阶段卷积和池化
x = tf.nn.relu(self.conv2_1(x))
x = tf.nn.relu(self.conv2_2(x))
x = tf.nn.max_pool(x, [1, 2, 2, 1], [1, 2, 2, 1], "VALID")
# 第三阶段卷积和池化
x = tf.nn.relu(self.conv3_1(x))
x = tf.nn.relu(self.conv3_2(x))
x = tf.nn.relu(self.conv3_3(x))
f3_3 = x # 保存特征图用于检测
x = tf.nn.max_pool(x, [1, 2, 2, 1], [1, 2, 2, 1], "VALID")
# 第四阶段卷积和池化
x = tf.nn.relu(self.conv4_1(x))
x = tf.nn.relu(self.conv4_2(x))
x = tf.nn.relu(self.conv4_3(x))
f4_3 = x # 保存特征图用于检测
x = tf.nn.max_pool(x, [1, 2, 2, 1], [1, 2, 2, 1], "VALID")
# 第五阶段卷积和池化
x = tf.nn.relu(self.conv5_1(x))
x = tf.nn.relu(self.conv5_2(x))
x = tf.nn.relu(self.conv5_3(x))
f5_3 = x # 保存特征图用于检测
x = tf.nn.max_pool(x, [1, 2, 2, 1], [1, 2, 2, 1], "VALID")
# 第六阶段卷积
x = tf.nn.relu(self.fc6(x))
x = tf.nn.relu(self.fc7(x))
ffc7 = x # 保存特征图用于检测
# 第七阶段卷积
x = tf.nn.relu(self.conv6_1(x))
x = tf.nn.relu(self.conv6_2(x))
f6_2 = x # 保存特征图用于检测
# 第八阶段卷积
x = tf.nn.relu(self.conv7_1(x))
x = tf.nn.relu(self.conv7_2(x))
f7_2 = x # 保存特征图用于检测
# 特征归一化
f3_3 = self.conv3_3_norm(f3_3)
f4_3 = self.conv4_3_norm(f4_3)
f5_3 = self.conv5_3_norm(f5_3)
# 在各尺度特征图上进行检测
# 分类分支(confidence)和回归分支(location)
cls1 = self.conv3_3_norm_mbox_conf(f3_3)
reg1 = self.conv3_3_norm_mbox_loc(f3_3)
cls2 = tf.nn.softmax(self.conv4_3_norm_mbox_conf(f4_3))
reg2 = self.conv4_3_norm_mbox_loc(f4_3)
cls3 = tf.nn.softmax(self.conv5_3_norm_mbox_conf(f5_3))
reg3 = self.conv5_3_norm_mbox_loc(f5_3)
cls4 = tf.nn.softmax(self.fc7_mbox_conf(ffc7))
reg4 = self.fc7_mbox_loc(ffc7)
cls5 = tf.nn.softmax(self.conv6_2_mbox_conf(f6_2))
reg5 = self.conv6_2_mbox_loc(f6_2)
cls6 = tf.nn.softmax(self.conv7_2_mbox_conf(f7_2))
reg6 = self.conv7_2_mbox_loc(f7_2)
# 对conv3_3层的分类结果进行特殊处理：max-out背景标签
# 这是S3FD算法的一个关键创新，用于提高小尺度人脸检测性能
bmax = tf.maximum(tf.maximum(cls1[:, :, :, 0:1], cls1[:, :, :, 1:2]), cls1[:, :, :, 2:3])
cls1 = tf.concat([bmax, cls1[:, :, :, 3:4]], axis=-1)
cls1 = tf.nn.softmax(cls1)
# 返回所有尺度的分类和回归结果
return [cls1, reg1, cls2, reg2, cls3, reg3, cls4, reg4, cls5, reg5, cls6, reg6]
# 根据配置决定在CPU还是GPU上加载模型
e = None
if place_model_on_cpu:
e = tf.device("/CPU:0")
if e is not None: e.__enter__()
# 创建并加载模型
self.model = S3FD()
self.model.load_weights(model_path)
if e is not None: e.__exit__(None, None, None)
# 构建模型运行环境，输入为任意大小的RGB图像
self.model.build_for_run([(tf.float32, nn.get4Dshape(None, None, 3))])
def __enter__(self):
"""
上下文管理器入口
返回:
S3FDExtractor实例自身
"""
return self
def __exit__(self, exc_type=None, exc_value=None, traceback=None):
"""
上下文管理器出口
参数:
exc_type: 异常类型
exc_value: 异常值
traceback: 堆栈跟踪
返回:
False - 将异常传递到外层
"""
return False # 传递异常到外层
def extract(self, input_image, is_bgr=True, is_remove_intersects=False):
"""
从图像中提取人脸边界框
参数:
input_image: 输入图像数组
is_bgr: 输入图像是否为BGR格式，默认为True
is_remove_intersects: 是否移除相交的人脸框，默认为False
返回:
list: 人脸边界框列表，每个边界框为[l, t, r, b]格式（左、上、右、下坐标）
"""
# 颜色空间转换：如果是BGR格式，转换为RGB
if is_bgr:
input_image = input_image[:, :, ::-1]
is_bgr = False
# 获取图像尺寸
(h, w, ch) = input_image.shape
# 计算缩放因子，保持图像纵横比
d = max(w, h)
scale_to = 640 if d >= 1280 else d / 2
scale_to = max(64, scale_to) # 确保最小尺寸为64
# 计算输入缩放比例
input_scale = d / scale_to
# 缩放图像以提高检测效率
input_image = cv2.resize(input_image, (int(w/input_scale), int(h/input_scale)),
interpolation=cv2.INTER_LINEAR)
# 运行模型进行检测
olist = self.model.run([input_image[None, ...]])
# 处理检测结果
detected_faces = []
# 对模型输出进行后处理，获取人脸边界框
for ltrb in self.refine(olist):
# 将边界框坐标缩放到原始图像尺寸
l, t, r, b = [x * input_scale for x in ltrb]
bt = b - t
# 过滤过小的人脸（小于40像素的边界框）
if min(r - l, bt) < 40: # 过滤任何边长小于40像素的人脸
continue
# 稍微扩大底部边界，以更好地包含下巴区域
# 这是为了与2DFAN-4关键点检测器更好地配合
b += bt * 0.1
detected_faces.append([int(x) for x in (l, t, r, b)])
# 按面积大小对人脸框进行排序，面积大的排在前面
detected_faces = [[(l, t, r, b), (r - l) * (b - t)] for (l, t, r, b) in detected_faces]
detected_faces = sorted(detected_faces, key=operator.itemgetter(1), reverse=True)
detected_faces = [x[0] for x in detected_faces]
# 移除相交的人脸框（保留面积较大的）
if is_remove_intersects:
for i in range(len(detected_faces) - 1, 0, -1):
l1, t1, r1, b1 = detected_faces[i]
l0, t0, r0, b0 = detected_faces[i - 1]
# 计算两个边界框的交集
dx = min(r0, r1) - max(l0, l1)
dy = min(b0, b1) - max(t0, t1)
# 如果有交集，则移除当前人脸框
if (dx >= 0) and (dy >= 0):
detected_faces.pop(i)
return detected_faces
def refine(self, olist):
"""
处理模型输出，生成边界框列表
参数:
olist: 模型输出的分类和回归结果列表
返回:
list: 处理后的边界框列表
"""
bboxlist = []
# 遍历所有尺度的输出结果
for i, ((ocls,), (oreg,)) in enumerate(zip(olist[::2], olist[1::2])):
# 计算当前层的步长（每个特征图单元对应原图的像素数）
stride = 2 ** (i + 2) # 4, 8, 16, 32, 64, 128
s_d2 = stride / 2
s_m4 = stride * 4
# 遍历所有置信度大于阈值的位置
for hindex, windex in zip(*np.where(ocls[..., 1] > 0.05)):
# 获取置信度分数
score = ocls[hindex, windex, 1]
# 获取回归偏移量
loc = oreg[hindex, windex, :]
# 计算先验框的中心和大小
priors = np.array([windex * stride + s_d2, hindex * stride + s_d2, s_m4, s_m4])
priors_2p = priors[2:]
# 根据回归偏移量调整边界框
box = np.concatenate((priors[:2] + loc[:2] * 0.1 * priors_2p,
priors_2p * np.exp(loc[2:] * 0.2)))
# 将中心点+大小格式转换为左上角+右下角格式
box[:2] -= box[2:] / 2
box[2:] += box[:2]
# 添加边界框和置信度到列表
bboxlist.append([*box, score])
# 转换为numpy数组
bboxlist = np.array(bboxlist)
# 如果没有检测到人脸，返回空边界框
if len(bboxlist) == 0:
bboxlist = np.zeros((1, 5))
# 应用非极大值抑制，过滤重叠边界框
bboxlist = bboxlist[self.refine_nms(bboxlist, 0.3), :]
# 过滤低置信度边界框，并转换为整数坐标
bboxlist = [x[:-1].astype(np.int) for x in bboxlist if x[-1] >= 0.5]
return bboxlist
def refine_nms(self, dets, thresh):
"""
非极大值抑制（NMS）算法实现
参数:
dets: 检测边界框数组，格式为[x1, y1, x2, y2, score]
thresh: 重叠阈值，大于此值的重叠边界框将被抑制
返回:
list: 保留的边界框索引列表
"""
keep = list()
# 处理空输入情况
if len(dets) == 0:
return keep
# 提取边界框坐标和置信度
x_1, y_1, x_2, y_2, scores = dets[:, 0], dets[:, 1], dets[:, 2], dets[:, 3], dets[:, 4]
# 计算每个边界框的面积
areas = (x_2 - x_1 + 1) * (y_2 - y_1 + 1)
# 按置信度降序排列
order = scores.argsort()[::-1]
keep = []
# 迭代执行非极大值抑制
while order.size > 0:
# 保留置信度最高的边界框
i = order[0]
keep.append(i)
# 计算当前边界框与其他边界框的交集
xx_1, yy_1 = np.maximum(x_1[i], x_1[order[1:]]), np.maximum(y_1[i], y_1[order[1:]])
xx_2, yy_2 = np.minimum(x_2[i], x_2[order[1:]]), np.minimum(y_2[i], y_2[order[1:]])
# 计算交集区域的宽度和高度
width, height = np.maximum(0.0, xx_2 - xx_1 + 1), np.maximum(0.0, yy_2 - yy_1 + 1)
# 计算交并比(IoU)
ovr = width * height / (areas[i] + areas[order[1:]] - width * height)
# 保留重叠度小于阈值的边界框
inds = np.where(ovr <= thresh)[0]
order = order[inds + 1]
return keep

复制代码

【DFL学习日记#3】S3FDExtractor.py 检测人脸边界框

浏览过的版块

曼加塔第一期

曼加塔第二期