CNN working principle

#1

@prashant_ml
Sir I have a query related to working of CNN. Actually I was playing with TensorSpace. Here I am attaching some images where the model failed to detect the digits.

image
image357×675 13.8 KB
image
image318×707 18 KB
image
image346×697 19.6 KB
image
image320×686 16.4 KB

These are 4 images where I have drawn the digits near to the edges and each and every time model failed to detect it correctly.
But we as a human can detect them easily because for us edges or vertices doesn’t matter. Our predictions is based upon CURVATURE or GESTURE. But when we are using CNN model, the dataset contains all the digits which a centric i.e. all those digits are drawn at the center. As a result model learnt that all the digits are CENTRIC in nature.

Now the point of all this observation is I want to develop an algorithm that can rectify this problem but how do I work upon the CURVATURE ?

#2

hey @Joy-Gupta-2763139277246091,
To do this ,
there are some ways, as i think.

  1. Get some of such images in the dataset , thats the best.
    But as we dont have them , then we got to the other one.

  2. Do data augmentation and increase the values of horizontal_shift, so that you can provide such images which are shifted somewhat to either left or right.

  3. You can also to do manually and store those images in a folder and use that for training.

I hope this helps