Classifying images into 2 classes

[Вадим Дрозд]

Hello! I have different two big sets of images, and my task is classify them into 2 classed: good and bad.

For the start i tried to train Imagenet, but it takes a lot of time, and accuracy don’t increases, maybe, there is a more simple architecture to this type of task?

Size of images near 256x256.

Thanks.

[Steve Schmugge]

I would try fine tuning the pre-trained model first.  I have had good results doing this. . Here is the flickr finetuning example. http://caffe.berkeleyvision.org/gathered/examples/finetune_flickr_style.html

[Andriy Lysak]

I had a similar project:

1. Use the image resizer that comes with caffe to get the images to be 256X256
2. I used imagenet as an example and modeled my CNN from that, i didn't change much, just a few settings and the output should not be a 1000 but how ever many classes you have, you said 2 so set it to 2.

Test a few variations of the model so you then will see what needs to be changed.

Best of luck!
Andriy