*4.1. Dataset*

The training set consisting of 1500 images was chosen from a global road damage detection challenge dataset [47]. Rain streaks of different angles and intensities have been added to those images using Photoshop to create a synthesized rainy image set. Corresponding clean images become the target ground truth image set for the synthesized rainy image set. The test set consists of both synthesized and real-world rainy images. Three hundred synthesized images were chosen from the global road damage detection challenge dataset and pre-processed similarly when preparing the training set. Test dataset outputs are shown in Figure 2 as a comparison between the proposed CGANet model and the state-of-the-art de-raining methods. Real-world rainy images were taken from the internet, and they were considered only for demonstrating the effectiveness of the CGANet model. Since ground truth images were not available for the real-world rainy images, they were not taken into the account when training the model. Test results of real-world images are shown in Figure 3.

#### *4.2. Evaluation Matrix and Results*

The peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) [48] were used to evaluate and compare the performance of the model. PSNR measures how far the de-rained image is distorted from its real ground truth image by using the mean squared error at the pixel level. As shown in Table 1, the proposed CGANet model obtained the best PSNR value compared to the other two methods. The structural similarity index (SSIM) is a perception-based index that evaluates image degradation as the perceived difference in structural information while also incorporating both luminance masking and contrast masking terms. Table 3 shows the SSIM value comparison between the proposed CGANet model and the other two state-of-the art methods. By referring to this comparison, we could verify that the proposed method performed well compared to other de-raining mechanisms, and this is also visually verifiable in Figures 2 and 3.

**Figure 2.** Qualitative comparison between GMM, pyramid networks, and proposed CGANet methods.

**Figure 3.** CGANet on real-world dataset ((**Left**) Input image; (**Right**) de-rained output).


**Table 3.** Quantitative comparison between different de-raining methods (mean ± STD).

## *4.3. Parameter Settings*

To optimize the proposed model, we followed the findings provided in the original GAN paper [14]. Instead of training the generator to minimize *log(1* − *D(x; G(x; z))*, we trained it to maximize *log D(x; G(x; z)).* Since the discriminator could be trained much faster compared to the generator, we divided the discriminator loss by 2 while optimizing the discriminator. As such, the discriminator training speed slowed down compared to the generator. Both the discriminator and generator models were trained with an Adam optimizer [49] with a learning rate of 0.0002 and a momentum parameter β1 of 0.5 [15]. The model was trained using 150 epochs and updated after each image, and as such, the batch size was 1.
