Single Image Super Resolution using a Hybrid Feature Dictionary for Remotely Sensed Images

In this paper, a sparse representation based single image super-resolution reconstruction (SRR) method using a self-learned hybrid feature dictionary is proposed. Efficient feature extraction plays a very important role in image SRR and most of the models use gradient-based feature extraction. These models, being an intensity-based, can be easily influenced by the intensity gradient to ignore essential edge profile and also fails to address the edge profiles like delta, roof and ramp edge which are very crucial for efficient reconstruction. Inspired by this, a hybrid feature based over-complete dictionary is formed by extracting various features of the LR input image. Features are extracted using the Fast Fourier Transform (FFT) procedure along with the first-and-second-order gradients. This dictionary is jointly trained with the HR image patch features using Orthogonal Matching Pursuit (OMP) and K-singular value decomposition (K-SVD) algorithm. By considering the similarity between the LR and HR image patch pairs, the output HR image is reconstructed using the sparse recovery model. Experimental analysis and results for remotely sensed data demonstrate that the proposed method reduces the computational complexity as well as outperforms other state-of-the-art methods in terms of qualitative and quantitative parameters. The average signal to noise ratio (PSNR) has improved significantly by +0.6 dB along with a notable boost in computational time and perceptual quality.