The project shows implementation based on a variety of applications of convolutional networks. DCGAN structure is used for model construction on GAN to learn deeply with multiple convolutional layers. The accuracy for classification task is about doubled after being supported by GAN-based data augmentation. This quantity is even significantly increased with the suitable selection of batch size, which is distinctly differentiated with 2 given label categories. Particularly, by being provided by more high-quality synthesized images, CNN for cow can obtain up to 95% accuracy with the adequate batch size for the training process on GAN. Lastly, the application of surface-feature extraction of PatchGAN trained along with CycleGAN considering the cycle consistency loss of image reconstruction from a distribution to the target one helps grasping the mapping between them and creates generators able to synthesize the transferred version containing the style feature of the objective distribution from the image of the original one without requiring any paired supportive similarity.
Keywords—GAN, CNN, convolutional layers.
The report of the project can be viewed in this Google Drive Link: https://drive.google.com/file/d/1hQLZC2AHNlFqR6J_Ej6855lZJp99WQld/view?usp=sharing
The data used for the project is cow-horse dataset.