A system that distinguishes images of ramen Jiro by machine learning with more than 90% accuracy is developed using Google's Cloud AutoML Vision

Google Cloud announced in January 2018 "Cloud AutoML VisionIs a service that can easily use machine learning using image recognition even for those who have little knowledge about machine learning but using this Cloud AutoML Vision, Japanese engineers can use Ramen Jiro's ramen with high precision We succeeded in identifying with.

Noodle on this: Machine learning that can identify ramen by shop
https://www.blog.google/topics/machine-learning/noodle-machine-learning-can-identify-ramen-shop/

AutoML Vision in action: from ramen to branded goods | Google Cloud Big Data and Machine Learning Blog | Google Cloud
https://cloud.google.com/blog/big-data/2018/03/automl-vision-in-action-from-ramen-to-branded-goods

The Cloud AutoML Vision announced by the Google Cloud AI team is a service that makes it easy to create a custom machine learning model for image recognition by loading labels and labeled images with even those without expert knowledge.

On the other hand, Ramen Jiro is a popular ramen shop with a head office in Tokyo · Mita, with more than 40 branches, mainly in the metropolitan area. The ramen with mountain vegetables, thick chishews, chopped garlic, plenty of back fat brings enthusiastic fans, also known as "girolian". Even though there are some differences in the taste and flavor of ramen made at 41 stores, the rough style is the same, so unless you are a fairly core fan, you can not first distinguish it.


Kenji DoiIn January 2018, I learned that Google released the alpha version of Cloud AutoML Vision. Mr. Doi made use of a deep learning framework called "Apache MXNet" to build a neural network for ramen Jiro's ramen identification and succeeded in identifying ramen images with an accuracy of 87%.

Mr. Doi reads the label of the shop and the image of about 48,000 ramen to Cloud AutoML Vision and as a result of testing "Can you distinguish the store from the ramen image", it distinguishes the ramen with an accuracy of 94.5% Especially succeeded. The following image is a mixing matrix representing the prediction accuracy of the ramen classifier output by Cloud AutoML Vision. The ramen images of each shop are loaded 50 times at a time, the shop name in the vertical direction, the shop name predicted sideways, the blue number part in which the vertical and horizontal shop name matched is the predictive number.


Even if the table and the bowl color and shape are the same, the classifier using Cloud AutoML Vision was able to distinguish the store with high precision. For reasons that could be identified with high precision so far, Mr. Doi considers the differences in how to arrange for each store and how to cut meat.

In addition, second-hand goods sold application "Mercari"Engineers used about 50,000 images and tried identifying brand products by Cloud AutoML Vision, it said that it got an accuracy of 91.3%. Given that Mercury's proprietary machine learning model had an accuracy of 75%, you can see that identification by Cloud AutoML Vision is quite accurate.

Cloud AutoML Vision,NASNetIt uses a new architecture called "machine learning" and it is a model that optimizes machine learning. With the system that updates the machine learning system itself according to the result obtained by machine learning, Cloud AutoML Vision, which has become easy to use highly accurate machine learning model by anyone, Democratization of the world "as a symbol of technology.


In addition, Mr. Doi independently developed "Ramen Jiro All Store Classifier"@ Jirou_deepIt is made into Twitter bot as, and it is said that if you actually reply with the image of ramen Jiro, it will tell you the store that you predicted.