关于使用 CNN 进行图像分类

Question

I built a image classification model using convolutional neural network for 5 classes.It is giving the training accuracy as 100% and testing accuracy as 82%. But when I gave an image manually for predicting, the model is not able to classify them correctly.For 10 images the model is classifying only 3-4 images correctly.What is the mistake? What should I do??

Answer 1

It is the symptoms of overfitting. You can try tweaking the model's parameters and number of layers. Please also try using Dropouts. If using Tensorflow, learn more tensorflow documentation

Answer 2

This means that you are clearly overfitting. Your model is not able to generalize well to the images that you enter manually. There are a few things that you could do.

1. You could try regularization so that your model does not overfit.
2. Try getting more data
3. Data augmentation

I hope that by introducing regularization, your model would work better on manually entered images.

Answer 3

Your model is not able to generalize properly. the training has gone very quickly through all the images in the training set and even started making predictions on the same images.

1. Try adding images and of different sizes.

2. Try flipping the original images horizontally to create transformed images that might also help in accurately training your model.

Answer 4

Besides that, probably your training set has way easier images than your test set (those image that you give manually). That can happen even if you are feeding the network from the same dataset.

To have better results on unseen images, you must have to reach a good generalisation level. Try what is said previously above...