Solving an Image Classification Problem (Overfitting)

Working on Using ResNet, Inception V3, VGG19. While Building an image classifier, I have trained in batch-wise. After training, I am getting accuracy on the dataset but getting poor accuracy on different images. On my dataset validation accuracy is good and validation is also decreasing. But when I use the same model to classify images captured from a mobile phone, the accuracy is poor (3-5 out of 12 are correct). I acquired the validation accuracy of 88%. In the Inception V3, there's no fluctuation of accuracy and it's decreasing smoothly. I have used Nvidia GPU. Sample size = 4000, Training dataset = 3200 I have used 10 classes, 400 each class size. Training loss is decreasing but the validation error is increasing. I have performed scaling (data augmentation). I could not perform the rest of data augmentation techniques like flip, etc. Because I'm getting a memory error. I have considered 40 epochs. Accuracy differs while using fit() & fit_generator(). So, I have switched to shuffle = True in get_batches(). I am getting a memory error after using shuffle in fit(). I'm using keras data generator. When I increase class size from 400 to 500, I'm facing memory error. I want to build a classifier, no detection. Can you solve this issue?