Data Science Stack Exchange is a question and answer site for Data science professionals, Machine Learning specialists, and those interested in learning more about the field. It only takes a minute to sign up.
Sign up to join this community
I have 30% blurred images in each classes. I have a total of 10 classes. I'm not allowed to drop these blurred images. How do I train the model to get better accuracy for both blurred and nonblurred training dataset ? Currently, I'm at 11% accuracy.
The images were blurred using a Gaussian blur.
I have used a Wiener filter, but not able to restore the image from blurred images.
Please can anyone suggest a good way to train the model.
I will suggest using data augmentation approaches to even out your data distributions. It will make your blurred images more usable to the model
The data distribution of 30% of your images deviates from the rest because they are blurred. Experiment with training using random blur with appropriate min-max ranges in the data augmentation pipeline (on the images that aren't blurred). This will help the model to smoothly generalize across blurred images. If you don't have labels of which images are blurred, use blur detection algorithms to determine a threshold after which you want to augment.
After doing this, it may be important that you do test-time data augmentation as well.
It would be helpful if you could add a bit more information, e.g. what problem you are solving. Assuming you are solving a classification problem, do these blurred images have any significance, e.g. are they uniformly distributed across classes? One thing that might help is adding some labels, for example in the input in the model. Normal images will have label $0$, blurred ones label $1$. In general, blurring is a legitimate way of transforming/augmenting the data. Another suggestion is that you can transform other images in the training data (not necessarily with Gaussian blur), to add variation.
