Rain Streak Removal Using Improved Generative Adversarial Network with Loss Function Optimization

Artificial Intelligence Computer Science Information Science
Rain Streak Removal Generative Adversarial Network Loss Function Optimization Self-Attention Mechanism Cross-Guided Bilateral Filter

Academic Background

In the field of computer vision, rain streaks are a common interference factor, especially in outdoor surveillance, autonomous driving, and intelligent transportation systems. Rain streaks significantly degrade image quality, affecting the recognition and analysis capabilities of visual systems. Traditional rain streak removal methods typically rely on processing single images. However, due to the complexity and diversity of rain streaks, these methods have limited effectiveness in handling distant rain streaks or complex scenes. In recent years, deep learning techniques, particularly Generative Adversarial Networks (GANs), have shown great potential in the field of image processing. Nevertheless, existing GAN-based rain streak removal methods still face challenges in dealing with rain streaks of different directions, shapes, and transparencies. Therefore, this study aims to propose an improved GAN framework to more effectively remove rain streaks and enhance image quality.

Source of the Paper

This paper was co-authored by Prabha R, Suma R, Suresh Babu D, and S Saila, who are affiliated with T. John Institute of Technology, PES University RR Campus, and Jyothy Institute of Technology in India, respectively. The paper was published in March 2025 in the journal Cognitive Computation, titled Rain Streak Removal Using Improved Generative Adversarial Network with Loss Function Optimization.

Research Process

1. Image Preprocessing

The study first preprocesses the input rain-streaked images using a cross-guided bilateral filter to extract the detail layer of the image. This filter combines spatial and intensity information to preserve edges and details in the image while removing noise. The specific steps are as follows: - Bilateral Filtering: Apply bilateral filtering to the input image to retain edge information. - Residual Image Calculation: Subtract the bilateral filtering result from the original image to obtain the residual image, which contains high-frequency details. - Guided Bilateral Filtering: Use a guided image to perform guided bilateral filtering on the residual image, further extracting the detail layer.

2. Rain Streak Removal

After preprocessing, the study employs an improved De-Rain Generative Adversarial Network (DR_GAN) for rain streak removal. The generator module of DR_GAN is replaced with a Dense Bidirectional Network (Attn_DBNet), which integrates DenseNet-121, Bidirectional Gated Recurrent Unit (BiGRU), and a Self-Attention Mechanism to enhance the effectiveness of rain streak removal. - Generator: Attn_DBNet extracts image features through DenseNet-121, models sequential dependencies with BiGRU, and focuses on key regions in the image using the self-attention mechanism to generate rain-free images. - Discriminator: The discriminator distinguishes between generated rain-free images and real rain-free images, improving the generator’s performance through adversarial training.

3. Loss Function Optimization

To further optimize the loss function, the study proposes the Chaotic Logistic Gazelle Optimization (CL-G) algorithm. This algorithm enhances randomness by introducing Chaotic Logistic Mapping, avoiding local optima. The specific steps include: - Initialization: Randomly initialize the positions of gazelles in the search space, with each position representing a potential solution to the loss function. - Exploration Phase: Disperse gazelles in the search space to ensure diverse starting points and prevent premature convergence. - Exploitation Phase: The optimal gazelle guides the group to focus on the most promising solutions.

Research Results

1. Image Preprocessing Results

The detail layer extracted by the cross-guided bilateral filter significantly preserves high-frequency details in the image, providing high-quality input for subsequent rain streak removal.

2. Rain Streak Removal Results

DR_GAN performs excellently in removing rain streaks, with the generated images outperforming existing methods in metrics such as Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Visual Information Fidelity (VIF). Specific data are as follows: - SSIM: DR_GAN achieves an SSIM value of 0.98, which is 6.59% to 13.42% higher than existing methods. - PSNR: DR_GAN achieves a PSNR value of 38.58, which is 16.68% to 28.64% higher than existing methods. - VIF: DR_GAN achieves a VIF value of 0.76, which is 8.31% to 14.60% higher than existing methods.

3. Loss Function Optimization Results

The CL-G algorithm performs excellently in optimizing the loss function, effectively avoiding local optima and improving the model’s convergence speed and performance.

Conclusion and Significance

This study proposes a rain streak removal method based on an improved generative adversarial network, significantly enhancing the effectiveness of rain streak removal by integrating DenseNet-121, BiGRU, and a self-attention mechanism. Additionally, the introduction of the Chaotic Logistic Gazelle Optimization algorithm further optimizes the loss function, enhancing the model’s robustness. This research not only theoretically advances the application of deep learning in image processing but also provides high-quality image processing solutions for practical applications such as outdoor surveillance and autonomous driving.

Research Highlights

  1. Innovative Network Structure: Proposes Attn_DBNet, which integrates DenseNet-121, BiGRU, and a self-attention mechanism, significantly improving rain streak removal effectiveness.
  2. Optimization Algorithm: Introduces the Chaotic Logistic Gazelle Optimization algorithm, effectively avoiding local optima and improving the model’s convergence speed and performance.
  3. Significant Performance Improvement: DR_GAN outperforms existing methods in key metrics such as SSIM, PSNR, and VIF, demonstrating its potential in practical applications.

Other Valuable Information

This study also explores the impact of different training data volumes and K-fold cross-validation on model performance, showing that as training data volume increases and K-fold values rise, model performance significantly improves. Additionally, the study compares the performance of different optimization algorithms, proving the superiority of the CL-G algorithm in optimizing the loss function.

Through this research, we not only propose an efficient rain streak removal method but also provide new ideas and approaches for the application of deep learning in image processing.