Awesome

Deep Data Augmentation for Weed Recognition Enhancement: A Diffusion Models and Transfer Learning Based Approach

Generated Samples

Performance comparison between diffusion models and GANs.

<img src="docs/figs/performance.png" alt="output_example" width="30%" height="30%"> Fig.1 Performance comparison between diffusion models and state-of-the-art GANs. Best values are in bold.

Diffusion Models

<img src="docs/figs/samples_cottonweed.png" alt="output_example" width="80%" height="40%"> Fig.2 Real (upper) and generated (lower) weed samples for CottonWeedID15 dataset. Each column represents one weed class. <img src="docs/figs/samples_deepweeds.png" alt="output_example" width="80%" height="40%"> Fig.3 Real (upper) and generated (lower) weed samples for DeepWeeds dataset. Each column represents one weed class.

GANs

BigGAN

<img src="docs/figs/samples_CottonWeeds_BigGAN.png" alt="output_example" width="80%" height="40%"> Fig.4 Generated weed samples for CottonWeedID15 dataset by BigGAN. Each column represents one weed class. <img src="docs/figs/samples_deepweeds_BigGAN.png" alt="output_example" width="80%" height="40%"> Fig.5 Generated weed samples for DeepWeeds dataset by BigGAN. Each column represents one weed class.

StyleGAN2

<img src="docs/figs/samples_CottonWeeds_StyleGAN2.png" alt="output_example" width="80%" height="40%"> Fig.6 Generated weed samples for CottonWeedID15 dataset by StyleGAN2. Each column represents one weed class. <img src="docs/figs/samples_deepweeds_StyleGAN2.png" alt="output_example" width="80%" height="40%"> Fig.7 Generated weed samples for DeepWeeds dataset by StyleGAN2. Each column represents one weed class.

StyleGAN3

<img src="docs/figs/samples_CottonWeeds_StyleGAN3.png" alt="output_example" width="80%" height="40%"> Fig.8 Generated weed samples for CottonWeedID15 dataset by StyleGAN3. Each column represents one weed class.