Comparative Evaluation of Modified Wasserstein GAN-GP and State-of-the-Art GAN Models for Synthesizing Agricultural Weed Images in RGB and Infrared Domain

Abstract

This study investigates the application of modified Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) to generate synthetic RGB and infrared (IR) datasets to meet the annotation requirements for wild radish (Raphanus raphanistrum). The RafanoSet dataset was used for evaluation. Traditional WGAN models struggle with vanishing gradients and poor convergence, affecting data quality. Customizations in WGAN-GP improved synthetic image quality, especially in maintaining SSIM for RGB datasets. However, generating high-quality IR images remains challenging due to spectral complexities, with lower SSIM scores. Architectural enhancements including transposed convolutions, dropout, and selective batch normalization improved SSIM scores from 0.5364 to 0.6615 for RGB and from 0.3306 to 0.4154 for IR images. This study highlights the customized model's key features:

• •
  Produces a 128 × 7 × 7 tensor, optimizes feature map size for subsequent layers, with two layers using 4 × 4 kernels and 128 and 64 filters for upsampling.
• •
  Uses 3 × 3 kernels in all convolutional layers to capture fine-grained spatial features, incorporates batch normalization for training stability, and applies dropout to reduce overfitting and improve generalization.