Image-to-image translation

CycleGAN with attention mechanism

I used Tensorflow-Keras to implement Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks by Jun-Yan Zhu et al., and add attention mechanism in generator network. You can find the original CycleGAN implementation from Keras official website in A_K_Nain. (https://keras.io/examples/generative/cyclegan/)

• As in the style transfer case, I use a CycleGAN model to transform the Banana Ripeness, and use an attention mechanism to keep the background style from being transferred as much as possible.
• For another cases, you may use this skill to do data augmentation, to improve accuracy rate of classification model when hard to collect defect image in early new product released stage.