CASE STUDY

Automatic Merchandise Image Segmentation
(Removing Background)

CATEGORY

E-Commerce

DATE

July 7th, 2018

AUTHOR

Xingwei Liu

PROGRAMMING LANGUAGE

Python (Tensorflow), MATLAB

Challenge

In online shopping environment, sellers have to put up images of their products for customers to see. In order to get a clean shot of a product, sellers usually have to ask designers to process a product image to look professional by replacing the background with a uniform scene. Different from normal landscape image, merchandises have complex and various color scheme and are difficult to be seperated from background by traditional method solely depending on color statistical information. Also, because human vision is sensitive to non-smooth color change, any edge error in detecting foreground will be noticeable by human.

Solution

We present a two-step approach addressing this problem

01

Saliency Detection

We use a multi-scale convolutional neural network to learn how human pays attention to certain area on an image. Which is called saliency detection.

02

Refinement & Postprocessing

Based on another convolutional network with auto-encoder structure, which tries to break image into pixel-level information and calculate a pixel-wise transparency mask that naturally blend the foreground with any background

Result

See images below


Want to learn more?

Let’s Chat

Location

3200 Park Center Dr, Suite 1400
Costa Mesa, CA 92626
United States