Our project aims to revolutionize image synthesis and classification by combining Deep Convolutional Generative Adversarial Networks (DCGANs) with Convolutional Neural Networks (CNNs). We seek to demonstrate the potential of DCGAN techniques in producing highly realistic images and achieving strong performance in image classification by utilizing these synthetic images during training. For this purpose, we selected the Intel Natural Scene dataset. After processing with the DCGANs model, we generated natural scene images that are extremely authentic. To assess the performance of our method, we fed the synthetic images into a pre-trained ResNet-18 model and conducted the classification task using actual testing data. Our approach yielded impressive results, achieving a 78.1% 6-class classification accuracy.
data.py: creates dataloaderfake_image.py: generates fake imagesmain.py: performs image classification taskmodel.py: implements the DCGANs modelresnet.py: implements the ResNet modelzip.py: zips the images
The dataset and associated structures for this project can be downloaded from here and here.
To generate fake images, run:
python3 fake_image.pyTo perform the classification task, run:
python3 main.pyBoth files can take arguments. For more information, see the ArgumentParser within the files.
Contributors (alphabetical order):
- Li, Weiyue - wel019@ucsd.edu
- Li, Yi - yil115@ucsd.edu