A light weight model to detect website logo form website screenshots

Discussion
,
1

Hello,

I want to create a light weight model which can preferably run in the browser, and detect website logo from the website screenshots.

Example:

Screenshot 2024-12-19 at 6.15.21 PM
Screenshot 2024-12-19 at 6.15.21 PM1822×886 79.5 KB

Here are the things that I have tried so far. Would be great if I can get some feedbacks on whether I am approaching this right.

I am exploring variants of YOLO for my use case. Since the model is trained on COCO dataset, I won’t be able to infer using zero shot. I will have to fine tune the model. I am using ultralytics APIs to train the model.

For dataset, I could not find similar dataset on internet, which has website screenshot with annotated logos, so I am thinking of creating one myself for top 100 websites. I am not sure if this data will be sufficient, but I can try to get started and see how the performance looks like.

I am using roboflow to annotate the images and the download the dataset so that I can train my YOLO model.

My questions are: Is this the right approach or are there better approach to this problem?

Thanks!

2

Have you tried YOLO World?

You can also fine-tune it.

1 Like
3

@Toxite - No, I have not. Thank you for your suggestion. I will give it a try. :slight_smile:

4

There are also lots of existing logo datasets you can start from. They might not have precisely what you’re looking for, but it could be helpful to include as a start. Additionally, if you find logos that have an alpha channel (transparent background) you can generate a lot of example images by placing them on various images (same with ones that have an opaque background, but I guess that depends on the brand).

An example dataset from searching “website logo dataset” https://weblogo2m.github.io/

1 Like