I am trying to build a model that takes in very noisy 200x200 greyscale images of spatially sparse objects and attempts to localise them with bounding boxes. The objects are very thin streaks (data of particle tracks from a particle accelerator) which form triangular patterns, and the background environment is swarmed with gaussian noise so that the patterns are very faint. There is only one such pattern that I need to detect per positive example.

I was wondering what good preprocessing steps and model architectures for such a problem would be. Things I have tried so far are:


  1. Gaussian blurring to reduce the effect of noise: this also blurs out the patterns and features to be detected and makes the data useless

  2. Canny edge detection: because the background noise is high, this does not pick out to the edges very well and has a similar blurring effect


I tried using a denoising Autoencoder trained on noisy images with the ground truth image as the label with 128, 64, 32 filter convolutions with max pooling and batch norm (and upsampling similarly on the decoder side) to remove some of the noise first. This just produces black images and kills all the features as well with the noise. I have tried different variants of this architecture but they all seem to have the same effect.

I was wondering what a good starting point architecture and preprocessing for such a sparse object localisation problem would be, and whether the denoising approach is worth pursuing at all given the nature of the data.


Your Answer

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

Browse other questions tagged or ask your own question.