Anime-Character-Detector

Anime or sometimes known as Japanimation is hand-drawn or computer animation associated with Japan. Anime Characters are specially designed according to the requirements of the shows and are extremely difficult to draw. Animators spend most of their time in designing the eyes of the character, each anime character has different key features that make them special.

I have used VGG19 model training on ImageNet. ImageNet is a dataset of nearly 1M Images divided into nearly ~1000 Categories. One such category is Humans (Person) and as Anime Character resembles human it gets really hard to differentiate them from one another just using the Images. To tackle this I used different approaches, one of them was to take the pre-trained vgg19 model and remove all the layers after the last max-pooling layers and add GlobalAveragePooling and Dense Layers.

Dataset

I collected a total of 820 medium size images using google and other resources to train the model.

• Categories:

  • anime
    Contained different images of anime characters, some were full body images, some only faces and some only the upper half of the body.
  • not_anime
    Contained images of cartoon characters and humans to force the model to classify humans as an anime character

Model

Pre-Trained VGG19 Model with ImageNet Weights. I freezed the VGG19 Model till block5_pool layer and added GlobalAveragePooling, Dropout and Dense Layers.

Layer (type)                 Output Shape              Param #   
=================================================================
global_avg_pooling (GlobalAv (None, 512)               0         
_________________________________________________________________
dropout_1 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 256)               131328    
_________________________________________________________________
dropout_2 (Dropout)          (None, 256)               0         
_________________________________________________________________
dense_2 (Dense)              (None, 128)               32896     
_________________________________________________________________
dropout_3 (Dropout)          (None, 128)               0         
_________________________________________________________________
output (Dense)               (None, 2)                 258       
=================================================================
Total params: 20,188,866
Trainable params: 164,482
Non-trainable params: 20,024,384
                          _________________________________________________________________

Download the Trainined Model

Training & Testing:

Loss Function: Binary CrossEntropy
Optimizer: Adam
Metrics: Accuracy

Batch Size: 32
Epochs: 10

Training:

• Data Size: 697 (224 x 224 x 3) Images
• Labels: 2 - Anime & Not Anime

Training

Testing:

• Data Size: 123 (224 x 224 x 3) Images
• Labels: 2 - Anime & Not Anime

Testing

Evaluate

I create a function to make predictions, it takes image as an input and returns image and label (anime or not_anime) along with probabilities.