Search code examples
image-processingdeep-learningartificial-intelligenceclassificationconv-neural-network

Transparent background on Image for training CNNs

I was wondering if having a transparent background for training a CNN affect it learning the features in any way? For example, for training a model to recognize cats, if we were to use only the polygon crop of a cat with a transparent background. Would it learn to recognize cats better?

enter image description here

Solution

  • Image-capturing devices are becoming very common, along with object classification software, to such an extent even a simple cellphone can realize state-of-the-art image processing methods. However, transparent objects do not offer features easy to identify: instead of hiding the background like opaque objects, they merely distort it. Therefore, their appearance drastically change regarding to their environment. You can think of it as a human would think of it. Is it easier for a human to recognize an object against an opaque background? I don't think so. Also, it's not really natural. Unless, you see an object embedded in a pile of snow, perhaps.