However, the result is rather disappointing -- its accuracy rate is as low as 27.8%, so it does not seem to "guess". The source is a Japanese magazine "Interface" April 2017 issue by CQ Publishing (Japan). Anyway, this is a joyful experiment.
Unfortunately, this article is written in Japanese without English title, I will introduce the overview.
Image data are divided into 128 x 128 areas (pixels), so the resolution is comparatively low (e. g. that of HDTV is 1920 x 1080 pixels).
Disappointed by the poor result, researchers focused on various features such as nose, correcting face images to the front direction, but things did not get better.
I think it is too early to conclude that "there is no correlation between face photos and blood types" in the experiment alone ...
<Reference data>
[Image data] Convert 6000 photos to grayscale 128 x 128 pixels
[Equipments for data learning] i7-4770 / 3.4 GHz / 16 GB (DDR 3) + Amazon cloud GPU (AWS EC 2 p 2. Xlarge parg) + machine learning library dlib
[Equipments for blood type determination] Raspberry Pi 3 / PiCamera (960 × 540 pixels)
[Processing time] About 20 seconds for face detection, about 30 seconds for blood type determination
[Results] The accuracy rate of test data is as low as 27.8%! Thereafter, various measures are taken, but accuracy rates showed little improvement. Researchers finally concluded that "there is no correlation between face photos and blood types."
<Source>
Hitoaki NAKAMURA, Satoru IWASADA (Bee Co., Ltd, Osaka Japan)
Cloud is good for studying image deep learning! A challenge to make up Raspberry Pi AI to guess blood type from face photo (Featured specials for beginners! RasPi x cloud AI: GPUs from 100 yen!)
Interface magazine (April 2017 issue) by CQ Publishing Co., Limited (Tokyo Japan).
中村 仁昭 , 岩貞 智
画像ディープ・ラーニングの学習はクラウドが良し! 顔写真から血液型を当てるラズパイ人工知能に挑戦してみた (注目特集 ビギナ向け! ラズパイ×クラウド人工知能 : GPUも100円から!) [in Japanese]
Research and development activities are on the company's site. (in Japanese)
No comments:
Post a Comment