Inquiry About Accelerating Defog and Stabilization on Hailo

General
1

Hi Everyone,
I am currently developing a program on Raspberry Pi that enhances images from a camera. Among the implemented features, Defog and Stabilization consume significant computational resources, resulting in slow processing speeds.

I would like to inquire whether these two functions—Defog and Stabilization—can be accelerated using Hailo. If so, could you kindly provide guidance on how to achieve this?

I sincerely appreciate your time and support.

2

Hey @inpegvision.nav,

Welcome to the Hailo Community!
Yes, you can implement these features by training CNN models (either separate models for each task or a combined solution).

Recommended Defogging Models:

  • AOD-Net: Lightweight, real-time dehazing with encoder-decoder architecture
  • DehazeNet: Directly estimates transmission maps for haze removal
  • GridDehazeNet: Uses attention mechanisms, efficient for low-power accelerators

Image Stabilization Models:

  • STN (Spatial Transformer Networks): Learns geometric transformations
  • DeepStab: Predicts smooth camera trajectories
  • PWAN: Uses CNN with optical flow for real-time stabilization

You can deploy these on Hailo-8/8L by converting your trained TensorFlow/ONNX models to Hailo format (.hef) using DFC and using HailoRT for inference.

Best Regrads,
Omria