Radiometric Calibration of Smartphone Camera with Deep Learning Features

MasterTone and exposure control of photographs is one of the most important image enhancement techniques. Despite of its importance, adjusting tone and exposure of smartphone jpeg images is difficult due to lack of image irradiance data of the input. This thesis presents a novel radiometric calibration method which recovers irradiance data from smartphone jpeg images. We train a neural network architecture that maps smartphone jpeg pixel colors to their corresponding irradiance data. By considering not only image pixels but local characteristics of image patches, a framework of this thesis can estimate inverse camera response function from various input images accurately and robustly. Local characteristics are represented by features extracted by a convolutional neural network. With system of this study, tone and exposure of smartphone jpeg images can be adjusted accurately and effectively, as demonstrated in our results