Anime Classification with CNN

This project classifies anime posters by genre using 5 different approaches. First, we employed a custom CNN model that begins with a convolutional layer equipped with 32 filters of 3x3 size and a ReLU activation function, immediately followed by a 2x2 max pooling layer, which helps in reducing the spatial size of the output and controlling overfitting. This pattern is repeated twice more: the second convolutional layer has 64 filters, and the third one has 128 filters, each followed by a respective 2x2 max pooling layer. These layers are instrumental in feature extraction and capturing spatial hierarchies. After the convolutional and pooling layers, a flattening layer is introduced to transform the 2D feature maps into a 1D feature vector. This vector is then fed into a dense layer with 128 neurons, again using a ReLU activation function, serving the purpose of learning non-linear combinations of features. The architecture culminates in an output layer, which is a dense layer with a number of neurons equal to the number of classes in the dataset, using a sigmoid activation function for multi-label classification.

Next, we applied transfer learning techniques with a ResNet50 Model trained on ImageNet10k. We then applied semi-supervised learning techniques using the unlabeled data in our dataset. We also employed AutoML systems using AutoKeras to compare the performance of our custom model to the AutoML model as well as the transfer learning model