Image analysis of coordinates using artificial intelligence

Image analysis using artificial intelligence has appeared one after another new cases. Amazon Go is trying to create a store with no cash register using image analysis, and some companies try to estimate the traffic volume and pedestrian age and sex by installing cameras near the Shibuya scramble intersection.

By the way, I think that opportunities to be "recommended" are increasing in everyday life. Amazon recommends goods that you like from the purchase history and Spotify recommends the music I want to hear by analyzing the viewing data.

Even fashion is decided to be saved if you really recommend your clothes and brands you like. But since fashion has many variables and its appearance changes with seasons and fashion, I feel that accuracy of recommendation is not so high yet.

Companies that are trying to open windows in the recommendation situation of such fashionNew rope. He is good at image analysis of fashion items and provides systems of image analysis system and item recommendation to EC and media site. This time I talked to Mr. Sakai, CEO and CEO.(Interview: January 2018)

Media that can see fashionista's snaps "# CBK"

New rope has developed several projects, but this time we would like to focus on "#CBK (Kabuki)" and "#CBK scnnr (Kabuki scanner)".

First#CBKIs a fashion media where snaps such as instagrams and reader models can be seen. The affiliated fashionista upgraded their coordination photos with Instagram and blogs. Reprint it to #CBK. Each coordination is tagged, for example, "Knit Red V Neck" or "Glenn Check Wide Pants" if the following image coordinates. (I hope that this tagging will come up to "# CBK scnnr" later, so please remember)

Tags are not managed by fashionistas who posted pictures, but managed by staff, they are done manually with all posts, which improves searchability. There are many people posting # CBK who are collecting fans with Instagram Grammar and Ameblo etc.

The user searches for coordination on #CBK, and "# CBK information" is described together as "# CBK magazine (Kabuki magazine)You can check the latest trends and dresses by watching. It is also possible to search dresses that match events and TPOs, such as second party, dating, commuting clothes and so on.

"# CBK scnnr" image analysis of fashion snaps

continue"#CBK scnnrThat's it. Image analysis using deep learning on snap posts as posted in #CBK. It is a system that judges information such as what kind of clothes, style, color item are worn by each item.

# CBK One of applications using scnnr "Uncle of Fashion"

New rope uses this technology, in terms of test service "fashion snap analysis API"Uncle of fashion"" Artificial intelligence shop clerk "Mika (Micah)"Has been developed.

Uncle fashion is a snap analysis system running on LINE. When snap (Whole body coordination is preferable) is submitted to LINE with fashion uncle, a page which analyzes the snap is generated. In analysis, areas such as tops, pants and bags are first recognized. And what kind of item is it, for example, if it is a top, "Riders Jacket" "Jacket" "Blouson" is analyzed. Since the analysis itself ends in about 2 seconds, there is no sense to wait for the analysis result. The generated page is linked with #CBK, and the point that the similar item and coordination worn by fashionista is posted from #CBK is also a point.

The other MIKA.This is the same until you submit a snap, but analyze the image and recommend items like 'Coordinate to this item is here'. If you adjust the taste such as "girly" "natural" "boyish" etc., higher precision recommendation is also possible. Since MIKA also runs on LINE, it is, so to speak, a LINE version of a real store clerk.

Develop # CBK scnnr with # CBK image as teacher data

Now, in order to analyze images using deep learning, a large amount of teacher data (original data) is required. Moreover, as fashion trends and new styles appear one after another, once you have accumulated data, it is not the end, you must collect data continuously. In this regard, the teacher data of the new rope is based on the #CBK data described above. In other words, # CBK continuously collects images of fashionistas such as instagrams, tags it with human power, and lets fashion learn artificial intelligence as teacher data.

We launched #CBK in 2014, prepared 100,000 tagged images and let them learn artificial intelligence. Tuning is also serious and I think that this can not be easily managed easily.

Thus, the New Rope developed artificial intelligence that can analyze every snap with high accuracy. We are offering this to partner companies such as EC, fashion media, SPA brand. For example, an application on the EC site provides an image retrieval system so that when retrieving an image, it is possible to search similar items in the application. It is also recommended for this item to recommend this item, such as to provide a recommendation function, it is possible to customize according to what you want to do for the site or app.

What I am working on now is to analyze snaps on SNS. For example, if you can analyze all # hash tags called coordination, you can see that this year's fashions, more particularly influencers, are popular with such things. Since they adopt such style earlier than fashion, they will be able to predict the next popular style and items if they know it.

In recent years it is said that fashion icons are dispersed. There is not a celebrity who influences 100,000 people alone, there are 100 influencers that affect 1000 people. Therefore, we must pursue the trends of influencers by using various methods, such as apparel and media that must follow the trend. Naturally the power of technology is necessary, the means taken by the new rope may be one of the answers.

There is much room for introducing technology in the fashion industry. For example, in fashion web content pictures have important meanings. There may be logs of the extent to which images of each company are seen, but it is in the future stage how to mean that log and how to apply it to business. #CBK I think that applying image analysis technology of scnnr makes it possible to use images more successfully.

Analysis of apparel as well as analysis of cosmetics and hairstyle

New rope plans to introduce the system to various ECs, brands, media etc. in the future. How do you think about the development after that?

I am thinking how to grow using data. If the number of installers increases, this data will also increase, and you can also increase the section of items and introductions. I would like to connect this to advertising and marketing automation this time.

For example, with Facebook ads it is possible to target ads by age, living area, interests, etc. But there is no fashion directionality included. New rope is trying to add fashion information there. If this becomes possible, it becomes possible to target people who like a fashion of the same taste as a certain brand, and it becomes possible to advertise more efficiently.

I am specializing in the narrow sense of fashion called apparel now but from the user's perspective there is no meaning specialized only so I would like to advance into cosmetics and hair style as well.

Analysis of # CBK scnnr is theoretically possible to analyze cosmetics and hairstyles. If it becomes possible, it will be possible to search and recommend across traversing the genre, such as "This style fits this style, this is recommended for cosmetics".

Furthermore, it seems that there is also a demand that apparel companies can not identify materials. If this is realized it will be possible to say "this material could not be worn as last time, so suggest other materials". Although knitting etc can be recognized now, analysis by units such as synthetic fiber and cashmere is obviously difficult. But depending on the resolution of the camera and the progress of the technology it may be possible.

New rope trying to provide items that match people's tastes from image analysis using deep learning. Please check fashion and EC business operators by all means.

The contents of the interview are also distributed by podcast。

The latest information is LINE@!

We announce you in LINE@ at article delivery. If you would like to be notified of the update of the article, please follow LINE@!