Leveraging GANs for Synthetic Data Generation in ResNet-50 Pattern Recognition

Leveraging GANs for Synthetic Data Generation in ResNet-50 Pattern Recognition

Authors

  • Anshuman Singh
  • Ansh Kapoor

Dataset

The dataset used in our experiments is PatternNet. It is a large-scale high-resolution remote sensing dataset collected for remote sensing image retrieval. There are 38 classes and each class has 800 images of size 256×256 pixels. The images in PatternNet are collected from Google Earth imagery.

Model(s)

The final model used is GAN Model 1.7 employs a Conditional Generative Adversarial Network (CGAN) architecture, enabling the generation of labeled images by conditioning both the generator and the discriminator on class labels. additional noise was introduced into the training data to increase the robustness of the generator. A dropout layer was added to the generator, helping to mitigate overfitting. Training was conducted for 300 epochs, and the model exhibited the best performance of all trained GANs in this project. Then ResNet 50 model was used train the model and test the model. For training we used 400 original images by PatternNet and 300 images generated by GAN1.7 per class. For testing we used 400 original images of PatternNet dataset.

Results

The model is compiled with the Adam optimizer, categorical cross-entropy loss, and accuracy as a metric. Since the classification task involves multiple categories, categorical cross-entropy is the appropriate loss function. Model Evaluation: The model is evaluated using the test data, resulting in a test accuracy of 64.59 percent and a test loss of 2.3822 percent. The test accuracy indicates that the model correctly classified approximately 64.59 percent of the test images.