update

README.md

@@ -1,23 +1,51 @@
-In this repository you will learn to build your own neural-network from scratch. To get started make sure you have `git`, a C/C++ compiler, and `make` installed. Then run:
+## 人脸识别
+--------------------
+这个仓库是基于 arcface 论文上完成的，其中主要分为四大块：人脸检测、人脸矫正、提取特征和特征比对。各个模块的大小和在 17 款 macbook-pro 的 CPU 上跑耗时如下：
 
+- 人脸检测：使用的是 mtcnn 网络，模型大小约 1.9MB，耗时约 70ms；
+- 人脸矫正：OpenCV 的仿射变换，耗时约 0.83ms；
+- 提取特征：使用 MobileFaceNet 和 IResNet 网络；
+- 特征比对：使用曼哈顿距离，单次搜索和完成比对耗时约 0.011 ms；
 
-⏳ Contents
---------------------
 
-- [x] support to load image，more details see [example01.cpp](https://github.com/YunYang1994/yynet/blob/master/examples/example01.cpp)
-- [x] support to resize, copy and gray in image processing，more details see [example02.cpp](https://github.com/YunYang1994/yynet/blob/master/examples/example02.cpp)
-- [x] implement a fully connected layer to classify mnist data
+## 注册人脸
 
-⚙️ Useage
---------------------
+注册人脸的方式有两种，分别是:
+
+1. 打开相机注册:
 
-- download mnist data
 ```bashrc
-$ wget https://pjreddie.com/media/files/mnist.tar.gz
-$ tar xvzf mnist.tar.gz
+$ python register_face.py -person Sam -camera
 ```
-- make your example
+
+按 `s` 键保存图片，需要在不同距离和角度拍摄 10 张图片或者按 `q` 退出。
+
+2. 导入人脸图片:
+
+保证文件的名字与注册人名相同，并且每张图片只能出现一张这个 ID 的人脸。
+
+
 ```bashrc
-$ make example01
-$ ./example01 images/sample.png 3
+$ python register_face.py -person Jay
 ```
+
+## 识别人脸
+
+|Method | LFW(%) | CFP-FP(%) | AgeDB-30(%) | MegaFace(%)| TensorFlow | 权重链接 |
+|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| MobileFaceNet | 99.50 | 88.94 | 95.91 | --- | 35ms | [提取码: xgmo](https://pan.baidu.com/s/1QIYpHYazaPMTI0E15WRGug)
+| MobileFaceNet | 99.77 | 98.27 | 98.28 | 98.47 | 435ms | --
+
+
+
+识别模型用的是 `MobileFaceNet` 网络，这里直接使用了 [insightface](https://github.com/deepinsight/insightface) 在 ms1m-refine-v1 三百万多万张人脸数据集上训练的模型。这部分工作在 `mxnet` 分支上，你可以通过 `git checkout mxnet` 进行切换。
+
+由于该模型是 mxnet 格式，因此使用了 [mmdnn](https://github.com/microsoft/MMdnn) 导出了其模型权重 `mobilefacenet.npy`。接着使用了 `TF2` 自己手写了一个 `MobileFaceNet` 网络并导入权重，预测精度没有任何损失。这部分工作在 `master` 分支上。
+
+最后，如果你要识别人脸，可以执行：
+
+```bashrc
+$ python main.py
+```
+
+

deploy/convert_to_ir.sh

@@ -0,0 +1,15 @@
+#! /bin/bash
+
+#================================================================
+#   Copyright (C) 2020 * Ltd. All rights reserved.
+#   
+#   Editor      : VIM 
+#   File name   : convert_to_ir.sh
+#   Author      : YunYang1994
+#   Created date: 2020-02-28 16:06:54
+#   Description : 
+#
+#================================================================
+
+python3 -m mmdnn.conversion._script.convertToIR -f mxnet -n model-symbol.json -w model-0000.params --inputShape 3,112,112 -o mobilefacenet
+python3 -m mmdnn.conversion._script.IRToCode -f tensorflow --IRModelPath mobilefacenet.pb --IRWeightPath mobilefacenet.npy --dstModelPath tf_mobilefacenet.py 

deploy/convert_to_tflite.py

@@ -0,0 +1,61 @@
+#! /usr/bin/env python
+# coding=utf-8
+#================================================================
+#   Copyright (C) 2020 * Ltd. All rights reserved.
+#
+#   Editor      : VIM
+#   File name   : convert_to_tflite.py
+#   Author      : YunYang1994
+#   Created date: 2020-02-15 14:47:28
+#   Description :
+#
+#================================================================
+
+import sys
+sys.path.append("../")
+
+import numpy as np
+import tensorflow as tf
+from mtcnn import PNet, RNet, ONet
+
+def load_weights(model, weights_file):
+    weights_dict = np.load(weights_file, encoding='latin1').item()
+    for layer_name in weights_dict.keys():
+        layer = model.get_layer(layer_name)
+        if "conv" in layer_name:
+            layer.set_weights([weights_dict[layer_name]["weights"], weights_dict[layer_name]["biases"]])
+        else:
+            prelu_weight = weights_dict[layer_name]['alpha']
+            try:
+                layer.set_weights([prelu_weight])
+            except:
+                layer.set_weights([prelu_weight[np.newaxis, np.newaxis, :]])
+    return True
+
+pnet, rnet, onet = PNet(), RNet(), ONet()
+pnet(tf.ones(shape=[1,  12,  12, 3]))
+rnet(tf.ones(shape=[1,  24,  24 ,3]))
+onet(tf.ones(shape=[1,  48,  48, 3]))
+load_weights(pnet, "./det1.npy"), load_weights(rnet, "./det2.npy"), load_weights(onet, "./det3.npy")
+
+pnet.predict(tf.ones(shape=[1,  12,  12, 3]))
+pnet_converter = tf.lite.TFLiteConverter.from_keras_model(pnet)
+pnet_converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
+with open("pnet.tflite", "wb") as f:
+    pnet_tflite_model = pnet_converter.convert()
+    f.write(pnet_tflite_model)
+
+rnet.predict(tf.ones(shape=[1,  24,  24, 3]))
+rnet_converter = tf.lite.TFLiteConverter.from_keras_model(rnet)
+rnet_converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
+with open("rnet.tflite", "wb") as f:
+    rnet_tflite_model = rnet_converter.convert()
+    f.write(rnet_tflite_model)
+
+onet.predict(tf.ones(shape=[1,  48,  48, 3]))
+onet_converter = tf.lite.TFLiteConverter.from_keras_model(onet)
+onet_converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
+with open("onet.tflite", "wb") as f:
+    onet_tflite_model = onet_converter.convert()
+    f.write(onet_tflite_model)
+