1
$\begingroup$

I am trying to train a model that detects logos in documents. Since I am not really interested in what kind of logo there is, but simply if there is a logo, does it make sense to combine all logos into 1 logo class?

Or are "logos" too diverse to group them together (like some logos are round, some are rectangular, some are even text based etc.) and the diversity of features will just make it hard for the neural network to learn? Or doesn't it matter?

(I am currently trying out the YOLOv3 architecture to begin with. Any other suggestions better suited are also welcome)

$\endgroup$
2
$\begingroup$

I think there is no absolute answer for this. Often its kind of trial and error. In general the CNN tries to generalize the problem, so using all logos with different augmentations and ground truths can maybe lead to some feature maps, which are so general that the CNN can find logos.

But if your logos are so various, and embedded in colorful websites, the tasks seems quite difficult, also if they vary in shape and form.Like you said, I think you definitely need an FPN (Feature Pyrimad network) to get the different sizes, scales and so on combined with an RPN (region-proposal network) to find the logos multiple times in the websites (if thats nessecary). For that you can use Mask-RCNN (https://github.com/matterport/Mask_RCNN). You can try to transfer-train it on Imagenet-Backbone for example to reduce training time. Just tried to use Mask-RCNN to segment colored cells from medical images. Worked out quite good.My own approach

I used ImageLabel for labeling, worked out quite intiutive and saved all in a JSON file.Imglabel

How large is your dataset and how complex are the logos ? Do you have examples which you can demonstrate?

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.