CNN model design recommendation

[inverse_kinematics]

I am training a traffic sign dataset that contains 39,000 images which are predominantly blurry images. Each are all of 96x96 pixels and categorized into 43 classes e.g. “slow down”, “stop” etc. The model will then be tested on a test dataset which are clear traffic signs unlike the train dataset.

I can’t seem to generate correct predictions on the test dataset. I’ve tried different configurations like adding drop out layers, normalization and augmentation but seem to be stuck at 4% in correct predictions of the test dataset.

These are some examples of the train dataset

Some examples of the test dataset

My model:

I could be overfitting, but reducing the parameters didn’t seem to make much difference. If anyone can recommend any changes or suggest a better model that would be awesome. I’m still new to DL so i apologize if this was painful to read

[chunduriv]

@inverse_kinematics,

Welcome to the Tensorflow Forum!

It looks like the training images are of low resolutions and hence the model might be struggling to learn the features to make the correct predictions.

It is suggested to improve the quality of training images so that they generalised to the test images and other images from the same distribution.

You can apply image enhancement techniques to improve the quality of blurry images like denoising, sharpening or contrast adjustment might help.

Thank you!