I am implementing a vanilla neural network (MLP) to do image classification in python using tensorflow on images of honey bees to detect their health status. The images in my dataset are of different shapes and sizes, so I decided to do image resize using cv2. All my images are now of the same size (64 by 64) but some of them have been stretched/shrieked due to resizing. Does this have an effect on the low prediction accuracy I am getting from my MLP?
Generally speaking, it highly depends on your data. If you have images of numbers for each image, it may not be that much bad but for images of cats or dogs, you completely put your information away by resizing to that size.
To answer your question, yes. The reason is that it leads to high Bayes error. It simply means that you as an expert can not say what they are. Consequently, it is not possible for the network to learn them. You can easily see the images and figure out that there is not anything to be learned. For instance, in that case, what is the difference between the sky and sea? Can a $64\times64$ image represent them? Can you as an expert find it out without any previous knowledge?