Click Here to read the full experiment of DC-GAN
In order to get larger set, I manually labeled hundreds of images and used them to train a simple classification network. I designed a five layer convolutional network with softmax classifier, like LeNet. And then combine anime headshot cropper with the classifier, similar to R-CNN, but it is more efficient than R-CNN because I change its segmentation method to specific anime headshot detector with margin detection.
I downloaded some sample clips of To Aru Kagaku No Rail Gun and extracted all the keyframes with FFMPEG. With the procedure that used in last post, I manually generated nearly 1000 headshots of Misaka Mikoto. Then, it was the show time of deep learning. With the use of ImageLoader function in torchvision library, all the labeled data was easily loaded into neural networks. After hundreds of training epochs, the test accuracy reached approximately 96%.
Here’s the code implemented in pytorch and opencv. I use it as a dataset generator for GAN tests.Its test accuracy reached 98% in Asuna dataset after several epochs of training.