Deep Learning and Entropy Regularization for Efficient Autoregressive Image Demosaicing
C. Anitha Mary, A. Boyed Wesley · 2025
The pixels of a digital camera with a single sensor are captured by a Color Filter Array (CFA), which records only one color element per pixel. As a result, the resolution of the color image is reduced due to noise and artifacts introduced during the reconstruction of the color image. In this paper, a novel image demosaicing method is proposed that combines Convolutional Neural Networks (CNN) with an Autoregressive approach to enhance the image demosaicing process. The goal is to reconstruct a full-resolution color image from the sampled data provided by digital cameras using CFA. The input image is first processed using a Local Polynomial Approximation (LPA) filter. The final demosaic image is generated by integrating the CNN and LPA filter with an autoregressive model. The effectiveness of the proposed method is evaluated using performance metrics such as Peak Signal-to-Noise Ratio (PSNR) and Second Derivative Measurement (SDME). The output of the proposed method is simulated using MATLAB software.