Automatic Image Enhancement

---

Description

For detailed information, see this link.

Basically, google shopping ads doesn't approve product images that contain call to actions text, promotional overlay and other distracting patches. The goal of this project is to create a machine learning model that automatically crop out relevant region of the product within an image.

Installation

    git clone https://github.com/Tranquangdai/Auto-Image-Improvements.git
    cd Auto-Image-Improvements/
    python setup.py install

Usage

Please refer to this Google Collab for a thorough guide on how to install, prepare data and train the model. The notebook also provide way to load a pretrained model and use it for inference.

Method:

• We make use of self-supervised learning to artificially generate bounding box and segmentations for product regions, and overlay that region to a designed background.
• Then we use detectron2 to train a Masked-RCNN model for Object-Detection and Instance-Segmentation.
• For a more general approach, see this link

Pretrained models

The pretrained model was trained on nearly 11500 data samples over various products and background for 2400 iterations. Visit this link to download the model.