目录
1. 作者介绍
2. 算法介绍
3.调用阿里云API进行证件照生成实例
3.1 准备工作
3.2 实验代码
3.3 实验结果与分析
参考（可供参考的链接和引用文献）
1. 作者介绍 王逸腾，男，西安工程大学电子信息学院，2022级硕士研究生
路治东，男，西安工程大学电子信息学院，2022级研究生，张宏伟人工智能课题组
2. 算法介绍 2.1 阿里云介绍
2.2 证件照生成背景 传统做法：通常是人工进行P图，不仅费时费力，而且效果也很难保障，容易有瑕疵。深度学习 做法：通常使用分割算法进行人物分割。例如U-Net网络 。
2.3 图像分割算法 《BiHand: Recovering Hand Mesh with Multi-stage Bisected Hourglass Networks》里的SeedNet网络是很经典的网络，它把分割任务转变成多个任务。作者的思想是：尽可能的通过多任务学习 收拢语义，这样或许会分割的更好或姿态估计的更好。其实这个模型就是多阶段学习网络的一部分，作者想通过中间监督来提高网络的性能。
下面为修改后的net_seedd代码：# Copyright (c) Lixin YANG. All Rights Reserved.
r"""
Networks for heatmap estimation from RGB images using Hourglass Network
"Stacked Hourglass Networks for Human Pose Estimation", Alejandro Newell, Kaiyu Yang, Jia Deng, ECCV 2016
"""
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from skimage import io,transform,util
from termcolor import colored, cprintfrom bihand.models.bases.bottleneck import BottleneckBlock
from bihand.models.bases.hourglass import HourglassBisected
import bihand.utils.func as func
import matplotlib.pyplot as pltfrom bihand.utils import misc
import matplotlib.cm as cm
def color_mask(output_ok):# 颜色映射cmap = plt.cm.get_cmap('jet')# 将张量转换为numpy数组mask_array = output_ok.detach().numpy()# 创建彩色图像cmap = cm.get_cmap('jet')colored_mask = cmap(mask_array)return colored_mask# 可视化# plt.imshow(colored_mask, cmap='jet')# plt.axis('off')# plt.show()
def two_color(mask_tensor):# 将张量转换为numpy数组mask_array = mask_tensor.detach().numpy()# 将0到1之间的值转换为二值化掩码threshold = 0.5 # 阈值，大于阈值的为白色，小于等于阈值的为黑色binary_mask = np.where(mask_array > threshold, 1, 0)return binary_mask# 可视化# plt.imshow(binary_mask, cmap='gray')# plt.axis('off')# plt.show()
class SeedNet(nn.Module):def __init__(self,nstacks=2,nblocks=1,njoints=21,block=BottleneckBlock,):super(SeedNet, self).__init__()self.njoints = njointsself.nstacks = nstacksself.in_planes = 64self.conv1 = nn.Conv2d(3, self.in_planes, kernel_size=7, stride=2, padding=3, bias=True)self.bn1 = nn.BatchNorm2d(self.in_planes)self.relu = nn.ReLU(inplace=True)self.maxpool = nn.MaxPool2d(2, stride=2)self.layer1 = self._make_residual(block, nblocks, self.in_planes, 2*self.in_planes)# current self.in_planes is 64 * 2 = 128self.layer2 = self._make_residual(block, nblocks, self.in_planes, 2*self.in_planes)# current self.in_planes is 128 * 2 = 256self.layer3 = self._make_residual(block, nblocks, self.in_planes, self.in_planes)ch = self.in_planes # 256hg2b, res1, res2, fc1, _fc1, fc2, _fc2= [],[],[],[],[],[],[]hm, _hm, mask, _mask = [], [], [], []for i in range(nstacks): # 2hg2b.append(HourglassBisected(block, nblocks, ch, depth=4))res1.append(self._make_residual(block, nblocks, ch, ch))res2.append(self._make_residual(block, nblocks, ch, ch))fc1.append(self._make_fc(ch, ch))fc2.append(self._make_fc(ch, ch))hm.append(nn.Conv2d(ch, njoints, kernel_size=1, bias=True))mask.append(nn.Conv2d(ch, 1, kernel_size=1, bias=True))if i < nstacks-1:_fc1.append(nn.Conv2d(ch, ch, kernel_size=1, bias=False))_fc2.append(nn.Conv2d(ch, ch, kernel_size=1, bias=False))_hm.append(nn.Conv2d(njoints, ch, kernel_size=1, bias=False))_mask.append(nn.Conv2d(1, ch, kernel_size=1, bias=False))self.hg2b = nn.ModuleList(hg2b) # hgs: hourglass stackself.res1 = nn.ModuleList(res1)self.fc1 = nn.ModuleList(fc1)self._fc1 = nn.ModuleList(_fc1)self.res2 = nn.ModuleList(res2)self.fc2 = nn.ModuleList(fc2)self._fc2 = nn.ModuleList(_fc2)self.hm = nn.ModuleList(hm)self._hm = nn.ModuleList(_hm)self.mask = nn.ModuleList(mask)self._mask = nn.ModuleList(_mask)def _make_fc(self, in_planes, out_planes):bn = nn.BatchNorm2d(in_planes)conv = nn.Conv2d(in_planes, out_planes, kernel_size=1, bias=False)return nn.Sequential(conv, bn, self.relu)def _make_residual(self, block, nblocks, in_planes, out_planes):layers = []layers.append( block( in_planes, out_planes) )self.in_planes = out_planesfor i in range(1, nblocks):layers.append(block( self.in_planes, out_planes))return nn.Sequential(*layers)def forward(self, x):l_hm, l_mask, l_enc = [], [], []x = self.conv1(x) # x: (N,64,128,128)x = self.bn1(x)x = self.relu(x)x = self.layer1(x)x = self.maxpool(x) # x: (N,128,64,64)x = self.layer2(x)x = self.layer3(x)for i in range(self.nstacks): #2y_1, y_2, _ = self.hg2b[i](x)y_1 = self.res1[i](y_1)y_1 = self.fc1[i](y_1)est_hm = self.hm[i](y_1)l_hm.append(est_hm)y_2 = self.res2[i](y_2)y_2 = self.fc2[i](y_2)est_mask = self.mask[i](y_2)l_mask.append(est_mask)if i < self.nstacks-1:_fc1 = self._fc1[i](y_1)_hm = self._hm[i](est_hm)_fc2 = self._fc2[i](y_2)_mask = self._mask[i](est_mask)x = x + _fc1 + _fc2 + _hm + _maskl_enc.append(x)else:l_enc.append(x + y_1 + y_2)assert len(l_hm) == self.nstacksreturn l_hm, l_mask, l_encif __name__ == '__main__':#a = torch.randn(10, 3, 256, 256)# SeedNetmodel = SeedNet()# output1,output2,output3 = SeedNetmodel(a)# print(output1,output2,output3)#total_params = sum(p.numel() for p in SeedNetmodel.parameters())/1000000#print("Total parameters: ", total_params)pretrained_weights_path = 'E:/bihand/released_checkpoints/ckp_seednet_all.pth.tar'img_rgb_path=r"E:\FreiHAND\training\rgb\00000153.jpg"img=io.imread(img_rgb_path)resized_img = transform.resize(img, (256, 256), anti_aliasing=True)img256=util.img_as_ubyte(resized_img)#plt.imshow(resized_img)#plt.axis('off') # 关闭坐标轴#plt.show()''' implicit HWC -> CHW, 255 -> 1 '''img1 = func.to_tensor(img256).float() #转换为张量并且进行标准化处理''' 0-mean, 1 std, [0,1] -> [-0.5, 0.5] '''img2 = func.normalize(img1, [0.5, 0.5, 0.5], [1, 1, 1])img3 = torch.unsqueeze(img2, 0)ok=img3print(img.shape)SeedNetmodel = SeedNet()misc.load_checkpoint(SeedNetmodel, pretrained_weights_path)#加载权重output1, output2, output3 = SeedNetmodel(img3)#mask_tensor = torch.rand(1, 64, 64)output=output2[1] # 1,1,64,64output_1=output[0]# 1,64,64output_ok=torch.sigmoid(output_1[0])output_real=output_1[0].detach().numpy()#直接产生的张量图color_mask=color_mask(output_ok) #显示彩色分割图two_color=two_color(output_ok)#显示黑白分割图see=output_ok.detach().numpy() # 使用Matplotlib库显示分割掩码# plt.imshow(see, cmap='gray')# plt.axis('off')# plt.show()# print(output1, output2, output3)images = [resized_img, color_mask, two_color,output_real,see,see]rows = 1cols = 4# 创建子图并展示图像fig, axes = plt.subplots(1, 6, figsize=(30, 5))# 遍历图像列表，并在每个子图中显示图像for i, image in enumerate(images):ax = axes[i] if cols > 1 else axes # 如果只有一列，则直接使用axesif i ==5:ax.imshow(image, cmap='gray')else:ax.imshow(image)ax.imshowax.axis('off')# 调整子图之间的间距plt.subplots_adjust(wspace=0.1, hspace=0.1)# 展示图像plt.show()
上述的代码文件是在bihand/models/net_seed.py中，全部代码链接在https://github.com/lixiny/bihand。
3.调用阿里云API进行证件照生成实例 3.1 准备工作 1.找到接口 购买试用包 查看APPcode 下载代码 参数说明
3.2 实验代码 # !/usr/bin/python
# encoding: utf-8
"""
===========================证件照制作接口===========================
"""import requests
import json
import base64
import hashlibclass Idphoto:def __init__(self, appcode, timeout=7):self.appcode = appcodeself.timeout = timeoutself.make_idphoto_url = 'https://idp2.market.alicloudapi.com/idphoto/make'self.headers = {'Authorization': 'APPCODE ' + appcode,}def get_md5_data(self, body):"""md5加密:param body_json::return:"""md5lib = hashlib.md5()md5lib.update(body.encode("utf-8"))body_md5 = md5lib.digest()body_md5 = base64.b64encode(body_md5)return body_md5def get_photo_base64(self, file_path):with open(file_path, 'rb') as fp:photo_base64 = base64.b64encode(fp.read())photo_base64 = photo_base64.decode('utf8')return photo_base64def aiseg_request(self, url, data, headers):resp = requests.post(url=url, data=data, headers=headers, timeout=self.timeout)res = {"status_code": resp.status_code}try:res["data"] = json.loads(resp.text)return resexcept Exception as e:print(e)def make_idphoto(self, file_path, bk, spec="2"):"""证件照制作接口:param file_path::param bk::param spec::return:"""photo_base64 = self.get_photo_base64(file_path)body_json = {"photo": photo_base64,"bk": bk,"with_photo_key": 1,"spec": spec,"type": "jpg"}body = json.dumps(body_json)body_md5 = self.get_md5_data(body=body)self.headers.update({'Content-MD5': body_md5})data = self.aiseg_request(url=self.make_idphoto_url, data=body, headers=self.headers)return dataif __name__ == "__main__":file_path = "图片地址"idphoto = Idphoto(appcode="你的appcode")d = idphoto.make_idphoto(file_path, "red", "2")print(d)
3.3 实验结果与分析
参考（可供参考的链接和引用文献）
1.参考：BiHand: Recovering Hand Mesh with Multi-stage Bisected Hourglass Networks（BMVC2020）
