This paper proposes a lightweight and accurate super-resolution (SR) method called Cascading Residual Network (CARN) and its variant CARN-M. The proposed method addresses the challenge of applying deep learning methods to real-world applications due to their high computational requirements. CARN is designed to achieve high performance in SR tasks while reducing the number of parameters and operations. It incorporates a cascading mechanism based on the ResNet architecture, which allows for multi-level representation and multiple shortcut connections to enhance performance. CARN-M further improves efficiency by using an efficient residual block (residual-E) and recursive network architecture. The experiments show that CARN and CARN-M achieve performance comparable to state-of-the-art methods with significantly fewer parameters and operations. The models are evaluated on standard benchmark datasets and demonstrate superior performance in terms of both accuracy and computational efficiency. The cascading mechanism enables the model to effectively propagate information across multiple layers, improving the reconstruction of details and contexts in images. The results show that CARN-M, being a more lightweight model, achieves comparable results to other methods with much fewer operations. The paper also discusses the trade-off between performance and computational efficiency, highlighting the effectiveness of the proposed methods in real-world applications.This paper proposes a lightweight and accurate super-resolution (SR) method called Cascading Residual Network (CARN) and its variant CARN-M. The proposed method addresses the challenge of applying deep learning methods to real-world applications due to their high computational requirements. CARN is designed to achieve high performance in SR tasks while reducing the number of parameters and operations. It incorporates a cascading mechanism based on the ResNet architecture, which allows for multi-level representation and multiple shortcut connections to enhance performance. CARN-M further improves efficiency by using an efficient residual block (residual-E) and recursive network architecture. The experiments show that CARN and CARN-M achieve performance comparable to state-of-the-art methods with significantly fewer parameters and operations. The models are evaluated on standard benchmark datasets and demonstrate superior performance in terms of both accuracy and computational efficiency. The cascading mechanism enables the model to effectively propagate information across multiple layers, improving the reconstruction of details and contexts in images. The results show that CARN-M, being a more lightweight model, achieves comparable results to other methods with much fewer operations. The paper also discusses the trade-off between performance and computational efficiency, highlighting the effectiveness of the proposed methods in real-world applications.