Gradient-Infused Multi-Scale Attention Based CNN for Robust Image Denoising

In the field of digital image processing, noise removal is a critical preprocessing step for tasks such as object detection, segmentation, and tracking. Convolutional Neural Networks (CNNs) have gained popularity for image denoising due to their adaptable architecture. Despite advancements in deep learning, several fundamental challenges remain. One key issue is the often-neglected role of gradient information in the denoising process. Additionally, traditional CNNs face difficulties in managing varying noise levels and distributions commonly encountered in real-world images, which hampers their ability to produce high-quality denoised results across different conditions. Our study emphasizes the negative impact of disregarding gradient information in real-world scenarios, leading to a notable drop in denoising performance. To overcome these challenges, we introduce the Gradient-Infused Multi-Scale Attention Driven CNN for Robust Image Denoising (GIMSANet). This innovative model is tailored to handle both synthetic additive white Gaussian noise (AWGN) and real-world image denoising tasks. Through extensive testing, we show that GIMSANet outperforms current methods, and our thorough ablation studies validate the efficacy of our approach.