ECASR: Efficient Channel Attention Based Super-Resolution

Despite recent advancements in single image super-resolution (SISR) methodologies, reconstruction of photo-realistic high resolution (HR) image from its single low resolution (LR) counterpart remains a challenging task in the fraternity of computer vision. In this work, we approach the problem of SR using a modified GAN with specialized Efficient Channel Attention (ECA) mechanism. CA mechanism prioritizes convolution channels according to there importance. The ECA mechanism, an extension of CA, improves model performance and decreases the complexity of learning. To capture the image texture accurately low-level features are used for reconstruction along with high-level features. A dual discriminator is used with GAN to achieve high perceptual quality. The experimental result shows that the proposed method produces better results for most of the dataset, in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and mean-opinion-score (MOS) over the state-of-the-art methods on benchmark data-sets when trained with same parameters.