Panchromatic Color Filter Arrays with white signal were introduced a while ago, e. g. RGBW Color Filter Array (CFA), assuming to have better resolution in lowlight due to panchromatic signal. However, there is no successful RGBW image sensor in the industry targeting mobile cameras until now. In this work, we introduce a novel Samsung RGBW image sensor and we study its performance in a popular remosaic scenario. We propose a DePhaseNet -a deep fully convolutional network to solve RGBW remosaicing or demosaicing problem. We propose to have 3 layers of phase differentiated inputs and custom frequency-based loss function for each layer. Proposed method successfully suppresses False Colors inherent to RGBW sensor due to heavily undersampled colors. By using this method, we were able not only to increase details preservation, but also increased color reproduction. We found that RGBW sensor is beneficial not only in low light scenarios, but also in widely spread remosaic scenarios. Experiments show improvement in image quality, yielding CPSNR of 42dB for Kodak dataset, reaching the bar of Bayer CFA demosaicing result. Proposed method advances state-of-the-art in RGBW demosaic by 6dB in CPSNR.

Recently, commercial vision sensors hit the mobile market. To achieve that, computer vision networks had to be quantized. However, this topic was not studied well for Image Signal processor (ISP) challenging image restoration tasks, being crucially important for hardware implementation, as well as for deployment on hardware accelerators, e.g. Neural Processors Units (NPU).In this paper, we studied the effect of the quantization of deep learning network on image quality. We tried various quantization on raw RGBW image demosaicing. Experimental results show that 10 bit weight quantization can sustain image quality at the similar level with floating-point network. 8 bit quantized network shows slight degradation in objective image quality and mild visual artifacts.Although network weight's bit-depth can be significantly reduced for computer vision tasks, our finding shows that it is not true for raw image restoration tasks: at least 10 bit weights are required to provide sufficient image quality. However, we can save some memory on feature maps bit-depth. We can conclude that network bit depth is critical for raw image restoration.