Deep Ranking/Best way to classify book covers?

Question

I recently came across a paper on Deep Ranking. I was wondering whether this could be used to classify book covers as book titles. (For example, if I had a picture for the cover of the second HP book, the classifier would return Harry Potter and the Chamber of Secrets.)

For example, say I have a dataset of book covers along with the book titles in text. Could that data set be used for this deep ranking algorithm, or is there a much better way to approach my problem? I'm quite new to this whole thing, and this one of my first projects in this field.

What I'm trying to create is a mobile app where people can take a picture of a book cover, have an algorithm/neural net classify the title of the book, and then have some other algorithm connect that to the book's Goodreads page.

Thanks for the help!

Answer 1

I read some papers talked about it you can take a look on them maybe help you.

Deep Neural Network Learns to Judge Books by Their Covers

Classification of Book Genres By Cover and Title

Answer 2

Could that data set be used for this deep ranking algorithm?

Yes you can! I think there are at least two approaches for this task:

1. First, solve using Image Classification. If you want to use Deep Learning, you can use Deep Convolutional Neural Network to create classifier that decide the image is HP book or not. You can read papers mentioned on Mahmoud's answer. But the problem is you need a very very large dataset, to make a good classifier you can't just provide one image for one book so if you have thousand title of books (or more) you need train your model with very hugh dataset.

2. Using image similarity or content based image retrieval (CBIR), there is a good discussion in Stackoverflow about this topic, there are a lot of techniques include Deep Ranking, perceptual hash, and others. One of their difference is Deep Ranking use fewer feature engineering technique than others. In my opinion using image similarity technique is better approach than using image classification (it's also compared in Deep Ranking paper) because some methods will be faster and didn't required a lot of dataset.

You also can read another simple reference of image similarity.