IMPORTANT: If you want to mark this code, please use Google Drive. The reason is that this code needs to place shape_predictor_68_face_landmarks.dat and main.py in the same directory, but this file is larger than 25 MB so that it cannot be uploaded to GitHub

The development of biometric recognition technology has greatly promoted modern security, administration and business systems. Facial recognition, the most important part of biometric recognition, is defined as a technology that can recognize people based on their facial features. It is based on sophisticated mathematical AI and machine learning algorithms that can capture, store and analyze facial features to match individual images in pre-existing databases. The aim of this project is to further develop programming skill and comprehensive understanding of machine learning systems studied during this course. In this project, we downloaded the face data set which included 5000 celebrities and 10000 face of anime characters, and the test code can be downloaded here. Four tasks are included in this project:

• A1 Gender detection: male or female.
  
• A2 Emotion detection: smiling or not smiling.
  
• B1 Face shape recognition: 5 types of face shapes.
  
• B2 Eye color recognition: 5 types of eye colors.
  

In task A, traditional machine learning methods SVM and RF are adopted. In task B, we adopted and compared traditional machine learning methods (RF) with Neural Networks (MLP and CNN).

• Python 3.3+
  
• macOS, Linux or Windows

Go check them out if you do not have them locally installed.

import numpy
import pandas
import sklearn
import cv2
import dlib
import matplotlib
import tensorflow

Make sure your tensorflow version is higher than 2.0

print(tf.__version__)

• main.py can be excuted to run the whole project.
• shape_predictor_68_face_landmarks.dat is a face recognition database used for detecting 68 feature points.
• Datasets folder: Please use your own dataset.
• The file folder A1, A2, B1, B2 contain code fles for each task. Basically, they cannot run separately but need to rely on the support of main.py
• Dlib_load_data.py is done independently by us for importing 68 feature points of images and labels.
  

In this file, we shall only look at and call

def pre_processing(num, label, root_path, point, split = True)

num is the number of images to be imported, label is the index of the labels, root_path is the location path of the filea called, point is the number of feature points needed, split determines whether split the dataset into training set and validation set.

• CV2_load_data.py is done independently by us for importing image pixel matrix sets and labels.

In this file, we shall only look at and call

def pre_processing(num, label, root_path, split = True)

num is the number of images to be imported, label is the index of the labels, root_path is the location path of the filea called, split determines whether split the dataset into training set and validation set. The imported pixels will be compressed to 50*50. If you are not satisfied with this compression,

def load_cv2_data(num, root_path):
    img_path = 'img/'
    X = np.zeros((num,50,50,3))
    for file in range(num):
        image = cv2.imread(root_path + img_path + '{}'.format(file) +'.png', 1)
        X = cv2.resize(image,(50,50))
   return X

you can modify line 3 and line 6

• A1.py contains a lot of defined function which can be called in main.py.

Specifically, it includes hyper-parameter selection by using GridSearchCV A1_SVM_ParameterTuning() A1_RF_ParameterTuning(), model construction A1_SVM() A1_RF(), accuracy report A1_acc(), and learning curve plotting plot_learning_curve()

• A2.py contains a lot of defined function which can be called in main.py.

Specifically, it includes hyper-parameter selection by using GridSearchCV A2_SVM_ParameterTuning() A2_RF_ParameterTuning(), model construction A2_SVM() A2_RF(), accuracy report A2_acc(), and learning curve plotting plot_learning_curve()

• B1 B1_CNN.py is the most accurate model and will be called in main.py.

Specifically, it includes model construction B1_CNN(), accuracy report B1_acc(), confusion matrix plotting plot_confusion_matrix(), loss curve plotting plot_loss_curve(), accuracy curve plotting plot_accuracy_curve().

• B1_MLP.py is not called in main.py and can be executed separately.

Specifically, it includes On hot coding transformation On_Hot_Coding(), model construction allocate_weights_and_biases()+multilayer_perceptron(), hyper-parameter setting, accuracy report
To ensure that it can be run separately, please import Dlib_load_data in the correct path and copy shape_predictor_68_face_landmarks.dat

• B1_RF.py is not called in main.py and cannot be executed separately.

Specifically, it includes hyper-parameter selection by using GridSearchCV B1_RF_ParameterTuning(), model construction B1_RF(), accuracy report B1_acc(), and learning curve plotting plot_learning_curve()

• B2 B2_CNN.py will be called in main.py.

Specifically, it includes model construction B2_CNN(), accuracy report B2_acc(), confusion matrix plotting plot_confusion_matrix(), loss curve plotting plot_loss_curve(), accuracy curve plotting plot_accuracy_curve().

Make sure your own datasets are in the same directory with main.py and have the following structure (subfiles)

• Datasets

• cartoon_set
• cartoon_set_test
• celeba
• celeba_test
  

Remember, if datasets are placed in the wrong path or missing subfiles, the main fuction will not run. If you really cannot unify datasets due to some specific reasons, then you are recommended to read the description of Dlib_load_data.pre_processing() and CV2_load_data.pre_processing() in the file description and try to modify the imported parameters in main.py.

Next, just enter the following code in the command window

python main.py

Then, you can see the accuracy socre for training set and test set in each task

Hint: It will takes few minutes to executed. (The testing computer is configured as CPU: i7-10750H, RAM: 16GB, GPU: Nvidia GeForce RTX 2060, and it taks about 8 min to run this code)

This project exists thanks to all the people who contribute.