An Intelligent Deep Learning Architecture Using Multi-scale Residual Network Model for Image Interpolation

Abstract

—Image interpolation techniques based on learning have been shown to be efficient in recent days, due to their promising results. Deep neural networks can considerably enhance the quality of image super-resolution, according to recent studies. Convolutional neural networks with deeper layers are commonly used in current research to improve the performance of image interpolation. As the network’s depth grows, more issues with training arise. This research intends to implement an advanced deep learning mechanism called Deep Multi-Scaled Residual Network (DMResNet) for effective image interpolation. A network cannot be substantially improved by merely increasing the depth of the network. New training strategies are required for improving the accuracy of interpolated images. By using the proposed framework, the Low Resolution (LR) images are reconstructed to the High Resolution (HR) images with low computational burden and time complexity. In order to dynamically discover the image features at multiple scales, convolution kernels of various sizes based on the residual blocks have been utilized in this work. In the meantime, the multi-scaled residual architecture is formulated to allow these characteristics to interact with one another for obtaining the most accurate image data. The interpolation performance and image reconstruction efficiency of the proposed model have been validated by using a variety of measures such as PSNR, SSIM, RMSE, Run time analysis, and FSIM. Popular datasets IAPR TC-12, DIV 2K, and CVDS are used for validating the proposed model. This model outperforms the state-of-art interpolation techniques in its performance, by yielding an increase of 8% in PSNR, 6% in SSIM, 1.2% in FSIM, and a decrease of 38.79% in RMSE, 5.875 times in run time analysis.