I am working on a project where we need to detect/count the presence of specific icon/logo on packages images.
My first idea is to label many packages images with the icons/logos to detect and use an object detection model (e.g. YOLO).
Sometimes the logos/icons are pretty similar to another and can become quite small.
I guess data augmentation might help improving the robustness against those.
I wanted to know if there was other possible/recommended approaches for such a task?
The searched icons/logos might be of different scale, rotation and appear multiple times. There won’t be any noise or other object masking a part of those.
More traditional template matching approach are not very robust regarding scale/rotation transformation or when there are multiple occurrence appearing.
Looking forward to discuss about this topics with all the folks here!