for a school project I have been given a dataset containing images of plants and weeds. The goal is to detect when there is a weed in the pictures. The training and validation sets have already been created by our teachers, however they probably didn't have enough images for both so they "photoshopped" some weeds in some of the training pictures.
Here are examples of images with the weed label in the training set:
In some cases, the "photoshopped" weed is hard to detect, and no shape resembling a weed is clearly visible like in this picture (weed at the very bottom, near the middle):
And here is an example of an image with the weed label in the validation set:
How would I go about preprocessing the training set so that a CNN trained on it would perform well on the validation set? I was thinking of applying a low-pass filter to the rough edges of the photoshopped images so that the network doesn't act as an edge detector, but it doesn't seem very robust. Should I manually select the best images from the training set? Thank you!