Video super resolution using non-linear regression and deep learning

Abstract

ABSTRACTIn this paper, a hybrid model is developed by integrating a non-linear regression model and optimization-driven deep learner for video super resolution. Initially, the low-resolution frames are subjected to framing, and each frame is provided to both Fractional-Group Search Optimizer-based Deep Belief Network (FrGSO-DBN) classifier and the nonlinear regression model. Then, the output frames of both methods are averaged by the proposed hybrid model and the enhanced video frame is generated. Here, the FrGSO is developed by modifying the update process of the GSO using the fractional theory, to train the DBN such that the weights in the DBN are selected optimally. Finally, the simulation results reveal that the proposed hybrid model achieves high values of 0.8697, 25.637 dB, 0.9171, and 49.821 dB, for the metrics Structural Similarity Index Measure, Peak Signal to Noise Ratio, Feature Similarity Index Measure, and Second Derivative like Measurement of Enhancement, respectively.