0
$\begingroup$

I am trying to recognize doodles in noisy images like in this one below. My dataset consists of only 10 000 images and 30 categories I've implemented a CNN but it is giving me a 6% accuracy. I am thinking about removing the noise before feeding the images to my CNN, but I have no idea which methods to use to remove this type of noise and I am not even sure that removing noise improves the NNs performance. Do you have any suggestions about methods to remove the noise or other tips to improve the performance of the model?

enter image description here

$\endgroup$
0
$\begingroup$

There are some techniques that you can research and apply. My first advice would be decrease the number of the neurons that you implemented already on your hidden layers. You can use some regularization and dropout functions. There are some denoising techniques as well if the problem is exactly this denoising data. I would recommend you to detect the real problem because usually there is a method calls "data augmentation" to overcome overfitting problems which increase the dataset with noisy data in a logical way.

As I mentioned above there are some techniques for denoising. Here is the link that you can look for more details about them. Also there are further techniques to approach this denoising but I highly recommend to detect your exact problem and improve your question.

Maybe it is related to your network modelling because 6% is very low. You should get more useful data at the end with this much dataset.

$\endgroup$
1
  • $\begingroup$ Thanks for the tips, I actually had the problem of overfitting but i solved it with data augmentation. I may be wrong but i don't think it's related to the model, because I have similar results with xgboost ans SVM. I will try the denoising techniques described in the article $\endgroup$
    – patelR
    Nov 25 '20 at 8:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.