CascadedGaze: Efficiency in Global Context Extraction for Image Restoration This paper introduces CascadedGaze Network (CGNet), an encoder-decoder architecture that efficiently captures global context for image restoration. CGNet employs a Global Context Extractor (GCE) module, which uses small kernels across convolutional layers to learn global dependencies without requiring self-attention. The GCE module is composed of three depthwise convolutions followed by pointwise convolutions and GELU. The GCE module is designed to progressively capture local and global information from the input data, with each layer operating on patches generated by the previous layer. The GCE module is combined with a Range Fuser module to aggregate local and global context. The GCE module is compared to self-attention mechanisms in Vision Transformers, showing that it can achieve competitive performance with lower computational overhead. The GCE module is tested on various image restoration tasks, including real and synthetic image denoising and single image deblurring. The results show that CGNet achieves state-of-the-art performance on these tasks while being computationally efficient. The GCE module is also compared to other efficient attention mechanisms, such as NAF blocks, showing that it can achieve similar performance with lower computational cost. The paper also discusses the computational efficiency of the GCE module, showing that it can be merged with similar channels to reduce computational overhead. The results show that CGNet outperforms previous methods on real image denoising tasks and achieves competitive performance on synthetic image denoising and single image deblurring tasks. The paper concludes that the GCE module is an effective way to capture global context for image restoration tasks while maintaining computational efficiency.CascadedGaze: Efficiency in Global Context Extraction for Image Restoration This paper introduces CascadedGaze Network (CGNet), an encoder-decoder architecture that efficiently captures global context for image restoration. CGNet employs a Global Context Extractor (GCE) module, which uses small kernels across convolutional layers to learn global dependencies without requiring self-attention. The GCE module is composed of three depthwise convolutions followed by pointwise convolutions and GELU. The GCE module is designed to progressively capture local and global information from the input data, with each layer operating on patches generated by the previous layer. The GCE module is combined with a Range Fuser module to aggregate local and global context. The GCE module is compared to self-attention mechanisms in Vision Transformers, showing that it can achieve competitive performance with lower computational overhead. The GCE module is tested on various image restoration tasks, including real and synthetic image denoising and single image deblurring. The results show that CGNet achieves state-of-the-art performance on these tasks while being computationally efficient. The GCE module is also compared to other efficient attention mechanisms, such as NAF blocks, showing that it can achieve similar performance with lower computational cost. The paper also discusses the computational efficiency of the GCE module, showing that it can be merged with similar channels to reduce computational overhead. The results show that CGNet outperforms previous methods on real image denoising tasks and achieves competitive performance on synthetic image denoising and single image deblurring tasks. The paper concludes that the GCE module is an effective way to capture global context for image restoration tasks while maintaining computational efficiency.