DeSmoke-LAP: Improved Unpaired Image-to-Image Translation for Desmoking in Laproscopic Surgery

Authors: Yirou Pan (✉), Sophia Bano, Vasconcelos Francisco, Hyun Park, Taikyeong Ted. Jeong and Danail Stoyanov

Model Architecture

The model is built based on the architecture of CycleGAN network, and two additional loss functions are added for inter-channel discrepancies and dark channel prior.

Sample Clips

Result comparisons on clips can be assess in the 'video clips' folder.

Dataset

The dataset is available to download here.

Prerequisites

• Linux or macOS
• Python 3
• CPU or NVIDIA GPU + CUDA CuDNN

Model train/test

• You can follow the steps in our provided Colab Notebook to train with your own data or use our pretrained models.

Acknowledgments

Our code is inspired by pytorch-CycleGAN-and-pix2pix.