Color Image Demosaicking using Deep CNN-based Self Ensemble Approach with Guided Image Filtering

Abstract

This article focuses on development of advanced demosaicking using deep convolutional neural network (D-CNN) model with self-ensemble method to reduce the computational complexities. The proposed D-CNN model consisting of densely connected residual blocks with the densely connected residual network (DRDN) for the training of various mosaic patterns and CFAs. Thus, this architecture reduces the vanishing-gradient problems generated during the training process with the utilization of efficient sub-pixel convolutional neural network (ESPCN) layer. The test images are applied to the D-CNN+DRDN architecture and performs the initial demosaicking operation using the local features of block-wise convolutional layers. Finally, improved guided image filter (IGIF) method is employed to preserve the edge intensity values in output demosaicked image. Extensive simulation results shows that proposed color image demosaicking model gives the enhancive subjective and objective performance with least mosaic pattern effects and reduced color errors. Performance evaluation compared to the demosaicking approaches from the literature like DDEMO, DRDN, and DRDN+ in terms color peak signal-to-noise ratio (CPSNR), and structural similarity (SSIM) index.