Black border around image

Hello, i'm working in Convolutional Neural Network project right now. The training data i prepared is slightly different when i input it on Matlab. It has black border around it, but it shouldn't be there because it can affect the accuracy of my network.

This is the original image

And this is after being input

is this affect my network's accuracy? and if it is how can i get rid of it?

Answer 1

[Javier Pinzón]

Hello Mellysa

Sometimes, when an img is in PNG format, it have some properties that matlab change, such as transparencies. I recommend you to change it to .JPG format, and perhaps re-colour your backgound to avoid problems when reading the image.

About accuracy, if black-borders are touching your kanjis, as is happening in the first image, the CNN may contemplate it as part of your symbol. but, as i can see, every single image have exact the same black-border, so, i think the network may "erase" this border when using ReLU layer. Also, you can train your network to test if borders affect, and perhaps check the activations.

[Mellysa Susilo]

I see, so image format is also important. The network show very low accuracy when i use those images. So i decide to change the images format to JPG.

Thank you for your answer :)