An efficient multi-scaled super resolution framework for image enhancement

In recent years, the super-resolution (SR) technique achieved impressive results by using a deep convolutional neural network (CNN). Generally, deep CNNs networks can extract better feature maps and reconstruct high-resolution (HR) images. However, a wider and deeper network increases complexity and computation time for any real-time applications. It is essential to develop a lightweight network for useful SR images that have less computation time and complexity. In this paper, we proposed multi-scale residual blocks based network for single image super-resolution. The multi-scale layers are effectively utilized in the model for feature extraction and reconstruct high-resolution(HR) images with less computation time and complexity. We try to find out an optimal point for image quality and computational time. In comparison to others, the optimal point of our model is (100.7) greater than other state-of-the-art CNN models.

I. Introduction

Super-resolution (SR) has become one of the most active research areas in the field of image processing. The equipment for high-resolution(HR) images is expensive and can be difficult to acquire, thus we need another way to increase the resolution of the low-resolution(LR) image i.e SR approach. The reconstructed HR image from the LR image by SR approach is extensively used in many real-world applications such as object detection, satellite imaging, and medical imaging, where the analysis of LR images can be arduous to interpret or diagnose [1]. The factors such as subsampling, blurring, brightness, complex background, distinct weather conditions, darkness, and appearance distortions affect the quality of the LR image. Due to the influence of these factors, the LR image is useless in practical applications, especially in traffic crime control.