AI Applied to Business - Fashion Recognition

The AI Competence Cluster of Arvato Systems (Niels Pothmann and his team) recently implemented a new showcase that can detect and classify various pieces of clothing. This showcase was initially planned to aid e-commerce customers in fashion recommendations and to optimize web content based on the users type of clothing style. 

How does it work?

The algorithm is based on trained neural networks that are capable of recognizing fashion live and in real-time. The user has to place the desired piece of clothing in front of the webcam of the computer. The algorithm will detect the category (t-shirt, sweater, blazer, etc.), pattern (e.g., logo detected), and dominant color of the particular fashion piece. As a result of this, an intelligent interplay between the neural networks guarantees a precise and fast classification of the above-mentioned clothing properties.

The algorithm is hosted in the Microsoft Azure cloud. Hereby, Azure functions and cloud storage accounts are used to handle the algorithm as well as the data-base for storing the training images. Furthermore, a web app provides the connection between the users' webcam and required Azure functions for classification.

What makes this showcase unique?

The advantage of this showcase compared to more straightforward, previous applications are that the present algorithm is pretty robust against confounding factors. That is, to be classified piece of clothing does not necessarily have to be held clean and sterile into the webcam. Instead, the user may still wear the fashion on his/her body. Even if a person wears multiple pieces, the algorithm will detect the most dominant one (e.g., a jacket will be recognized even if it is worn above a t-shirt). Even obstacles in the background usually do not directly lead to miss-classifications.

Another huge advantage of the present algorithm is its high flexibility and dynamics. The above describes the intelligent interplay of neural networks, and new data can be added with only a few efforts and less training data than usually needed in machine learning. Thus, e-commerce customers might add new clothing categories or patterns individually at any time.