Overview / Usage

Image dehazing refers to procedures that attempt to remove the haze amount in a hazy image and grant the degraded image an overall sharpened appearance to obtain clearer visibility and a smooth image.

Most of the handy techniques include one-pass image filters. However, the result isn't pleasing to the human eye and degrades images upon tested on nontrivial real-life images. On the other hand, deep learning-based techniques that do not depend upon prior yield good performance in terms of aesthetics and PSNR metric. However, due to high parameters and FLOPs count, tend to be slow during inference.

To find the right balance between good metric score & runtime performance, we resort to techniques like Knowledge Distillation, Feature Fusion Attention(FFA), and OpenVino's model optimization to get the best of both worlds!

Guided by Prof. Mohammad Bohara & Dr. Amit Ganatra (Principal DEPSTAR)

Methodology / Approach

High-level Overview:

1. We train a heavy teacher model which achieves SOTA.
2. Experimenting with multiple Knowledge Distillation techniques we train a lightweight student model. The model learns by itself + from its teacher + from their mutual findings.
3. Convert the ONNX model to IR model using OpenVino's Model Optimizer.

TODO:

1. Experiment with Model Optimizer Extensions, specifically Graph Transformations.
2. Try out Vision transformer architecture.
3. Deep mutual learning.

Technologies Used

• Python
• PyTorch
• OpenVino
• OneAPI

Repository

https://github.com/manncodes/dehazing-openvino

Other links

• Research paper for the base model