A Fast Super-Resolution Convolutional Neural Network Model Based on Module Enhancement

The Fast Super-Resolution Convolutional Neural Network (FSRCNN) attains commendable image super-resolution outcomes while maintaining an extremely low parameter count. However, it still has certain limitations, such as the need to enhance its capacity in processing image details. Taking into comprehensive consideration both efficiency and super-resolution performance, this paper presents a fast super-resolution convolutional neural network model based on module enhancement. The core concept of this model is to make full use of the strong feature representation capabilities of non-linear mapping and the image detail information contained in the lower layers of the network. By incorporating dense connection modules and residual learning modules into the original model, the model's ability to capture image details and textures is enhanced. Specifically, the dense connection module receives and synthesizes the image features and textures with narrow fields of view (receptive fields) from all relatively shallow network layers. The residual learning module, on the other hand, further facilitates the flow of information between the shallow and deep network layers. Additionally, an innovative activation function and a novel activation function structure are proposed, which significantly accelerate the model's convergence speed. Experimental results demonstrate that the FSRCNN model with module enhancement achieves remarkable performance improvements on multiple standard datasets. These improvements are evident in both quantitative metrics, such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), and in the intuitive visual quality.