Researchers Apply Machine Learning to Analyze Books Based On Their Covers
You shouldn’t judge a book by its cover, but a supercomputer can. Specifically, researchers at Japan’s Kyushu University developed a deep neural network that determines book categories by analyzing their covers.
Brian Kenji Iwana and Seiichi Uchida, the researchers behind this project, built their dataset from over 130,000 book covers and genres downloaded from Amazon. The neural network they trained is multi-layered, making it easier for the program to recognize correlations between covers and genre. If a book was listed as more than one genre, Iwana and Uchida used the first listing.
After initial programming and validation, the program’s ability to link book cover to genre was tested on new covers, and it identified the correct genre in its top 3 choices with an over 40% success rate, which is nothing to sneeze at.
The algorithm had the easiest time identifying travel books and computer/technology books, and had the most difficult time with biographies and memoirs. It also had trouble discerning between children’s books and comics or graphic novels, suggesting that subtle or varied cover design choices are an obstacle for programs like this.
As of this writing, Iwana and Uchida have not compared their neural network’s performance to a human’s ability to connect book covers and genre. Click here to read their paper discussing their findings.