Can this model be used to extract features of anime pictures？

[OysterQAQ]

first, thx for your awesome project. Can this model be used to extract features of anime pictures？ If it can, I would like to use it to extract image features and build reverse image search system.

[KichangKim]

DeepDanbooru model returns 8k of (0.0~1.0) scores, so you can use this vector as feature vector. But it may be noisy and has too high dimension, so you may need something post-processing like PCA.

[OysterQAQ]

Yes, I noticed that it seems to output features vector of 7722 dimensions. Will there be any problems if i use keras vgg16(imagenet) model(Exclude fc layer) to extract the features of anime pictures? It will output 512 dimensions feature but it seems that the imagenet dataset only contains real photos

[KichangKim]

It doesn't matter what model you use, but I recommend to try DeepDanbooru model (resnet based) first because it is already trained and released so you can freely use it. Also DeepDanbooru is trained by using anime-pictures, unlike popular real-photo based models.

[OysterQAQ]

Thanks for your answer. I prefer to use the DeepDanbooru model, but I understand that dimensionality reduction methods such as pca need to load the entire matrix, it is expensive. Is there some better dimensionality reduction method? The pre-trained VGG16 model still needs dimensionality reduction, because it is expensive to calculate the similarity of 512-dimensional vectors in tens of millions of pictures.

[KichangKim]

Autoencoder may be the answer I think. https://www.tensorflow.org/tutorials/generative/autoencoder#first_example_basic_autoencoder

[OysterQAQ]

Thank you for your help. I will research it, Forgive me I am not good at deep learning