Time Based Image Synthesis: Light and Color Conversion Using Machine Learning Techniques

The progress made in camera technology and computational imaging has been astounding when it comes to photography. A specialized neural image signal processor is currently used by many top smartphone cameras and DSLR to convert chaotic raw sensor photos into the final, polished output. A sizable dataset of image pairs-one with a noisy, short-exposure raw picture and another with a clean, long-exposure ground-truth raw image-is needed to train these neural ISPs for low-light, processing It takes a lot of effort and time to capture these aligned image pairs, and motion blur in the long-exposure ground truth frequently makes the process difficult. Artificial nighttime datasets can be used to train various image processing algorithms and neural ISPs that are more resilient and effective. In order to overcome this difficulty, we proposed a model that uses to convert image colour and lighting of a raw picture to artificially create realistic image of different time zone raw photos, brighter photos are typically easier to take, have less noise, and are less likely to suffer from motion blur. The objective of the suggested processing framework is to transform raw photographs taken during the different time into simulated images of different timing with different levels of noise and other low-light features. More ease and flexibility in training neural ISPs for time-based rendering is made possible by this artificial time- based picture dataset. Additionally, in our proposed method the performance of the trained neural ISPs can resemble that of models trained only on real different timings photos if they mix their synthetic data with a tiny percentage of real data. The potential to produce accurate raw photos of midnight and different night time in a synthetic manner holds significant implications for computational photography and image processing applications and even object recognition under difficult nocturnal conditions.