Different result for two similar image convolutional neural network

Question

I've trained simple convolutional neural network for image classification. it reached 97% accuracy for validation set. my dataset includes images taken by camera from paper book and also screenshot from pdf. but when i want to predict image taken from book by webcam, the model accuracy and confidence is very bad. these are images:

image taken by mobile, high predict accuracy

very bad result.

for better generalization, I also added random noise to my dataset images but it's not good yet.

Answer 1

Even i face this kind of problem in my computer vision tasks.
i think there is some image quality difference b/w your webcam image and mobile phone image...
what i do is mix both types (can also include images taken from other cameras just to increase the variety) of images in your training data and validation data.
This would generalize your model on any data