Facial Attribute Recognition from face images

This is a Keras implementation of a CNN for facial attribute recognition. I trained Visual Transformer and facenet for facial attribute extraction.

Star History

Dependencies

• Python3.6+

Tested on

• Ubuntu 16.04, Python 3.6.9, Tensorflow 2.3.0, CUDA 10.01, cuDNN 7.6

Dataset Preparation Script

You can find the dataset preparation script below. Feel free to copy and make changes as needed:

Dataset Preparation Script on Kaggle

Dataset

I trained the face attribute extraction models using the CelebFaces Attributes (CelebA) Dataset.

You can download the preprocessed dataset from the link below. The faces have been cropped and converted to RGB format. The dataset contains 202,599 images with facial attributes, split into 100,000 for training and 50,000 for validation. You can download it from the output section of the Kaggle notebook:

Download Preprocessed Dataset

Train

python train.py --imagepath=/data/imageFile100000.npz --labelpath=/data/labelFile100000.npz

Testing

python demo.py

Testing in real time using the webcamera

python realtime_testing.py

Pre_trained weights

Please use the below weights for testing. https://drive.google.com/drive/folders/1NWz2E3b75mO_Ox8tb9d77vBi8dNHUv1T?usp=sharing

Model results:

Model	Train Accuracy	Validation Accuracy	Test Accuracy
VIT	81.2	82	81.28
FaceNet	84.5	85.71	86.25
InclusiveFaceNet			90.96

Validation Dataset results

Test sample

I used the bigbangtheory cast image as a testing image. Please find the person's result.

Output Video

Docker Installation Steps:

• Step 1: Download the Pretrained facenet model and create the new folder inside the DockerFiles place it. i.e /DockerFiles/models/

/DockerFiles/models/model_inception_facial_keypoints.h5

• Step 2: Use the below command for docker compose.

docker-compose up -d

• Step 3: Run the following the command to build the docker image.

 docker build -t facial_attribute .

• Step 4: Start the detection service.

 docker run -it facial_attribute

• Step 5: Pass the image for testing.

curl -X POST -F 'file=@/home/DSN/Desktop/1.jpg' http://172.17.0.2:5000/

• Step 6: JSON results format:

{
  "result": [
    {
      "coord": [
        607,
        65,
        711,
        169
      ],
      "face": [
        {
          "label": "Attractive",
          "prob": "0.5497437"
        },
        {
          "label": "Male",
          "prob": "0.8896191"
        },
        {
          "label": "No_Beard",
          "prob": "0.92911637"
        },
        {
          "label": "Young",
          "prob": "0.92061347"
        }
      ]
    },
    {
      "coord": [
        1149,
        131,
        1235,
        218
      ],
      "face": [
        {
          "label": "Big_Nose",
          "prob": "0.5611748"
        },
        {
          "label": "Male",
          "prob": "0.96252704"
        },
        {
          "label": "Mouth_Slightly_Open",
          "prob": "0.78494644"
        },
        {
          "label": "No_Beard",
          "prob": "0.5100374"
        },
        {
          "label": "Smiling",
          "prob": "0.7040582"
        },
        {
          "label": "Young",
          "prob": "0.8379371"
        }
      ]
    },
    {
      "coord": [
        803,
        150,
        889,
        237
      ],
      "face": [
        {
          "label": "Attractive",
          "prob": "0.59744525"
        },
        {
          "label": "Heavy_Makeup",
          "prob": "0.552807"
        },
        {
          "label": "No_Beard",
          "prob": "0.986242"
        },
        {
          "label": "Wearing_Lipstick",
          "prob": "0.692116"
        },
        {
          "label": "Young",
          "prob": "0.93902016"
        }
      ]
    },
    {
      "coord": [
        976,
        141,
        1062,
        227
      ],
      "face": [
        {
          "label": "Eyeglasses",
          "prob": "0.6799438"
        },
        {
          "label": "Male",
          "prob": "0.7749488"
        },
        {
          "label": "No_Beard",
          "prob": "0.9461371"
        },
        {
          "label": "Young",
          "prob": "0.6490406"
        }
      ]
    },
    {
      "coord": [
        179,
        150,
        266,
        237
      ],
      "face": [
        {
          "label": "Male",
          "prob": "0.9068607"
        },
        {
          "label": "Mouth_Slightly_Open",
          "prob": "0.91096807"
        },
        {
          "label": "No_Beard",
          "prob": "0.623013"
        },
        {
          "label": "Smiling",
          "prob": "0.8010901"
        },
        {
          "label": "Wearing_Hat",
          "prob": "0.57096326"
        },
        {
          "label": "Young",
          "prob": "0.88812125"
        }
      ]
    },
    {
      "coord": [
        446,
        158,
        549,
        261
      ],
      "face": [
        {
          "label": "Big_Nose",
          "prob": "0.7039994"
        },
        {
          "label": "Eyeglasses",
          "prob": "0.87806904"
        },
        {
          "label": "High_Cheekbones",
          "prob": "0.596"
        },
        {
          "label": "Male",
          "prob": "0.9493711"
        },
        {
          "label": "Mouth_Slightly_Open",
          "prob": "0.82170117"
        },
        {
          "label": "No_Beard",
          "prob": "0.86256987"
        },
        {
          "label": "Smiling",
          "prob": "0.88448894"
        }
      ]
    },
    {
      "coord": [
        304,
        170,
        390,
        256
      ],
      "face": [
        {
          "label": "Bangs",
          "prob": "0.56573707"
        },
        {
          "label": "Eyeglasses",
          "prob": "0.65550697"
        },
        {
          "label": "High_Cheekbones",
          "prob": "0.67516124"
        },
        {
          "label": "Mouth_Slightly_Open",
          "prob": "0.8242004"
        },
        {
          "label": "No_Beard",
          "prob": "0.9694848"
        },
        {
          "label": "Smiling",
          "prob": "0.8470793"
        },
        {
          "label": "Young",
          "prob": "0.6668907"
        }
      ]
    }
  ]
}

References:

FaceNet Tensorflow

Vision Transformer (ViT)

CelebFaces Attributes (CelebA) Dataset