Akademik Veri Yönetim Sistemi
31st IEEE Conference on Signal Processing and Communications Applications (SIU), İstanbul, Türkiye, 5 - 08 Temmuz 2023
Today, images presented to the users by digital cameras are created by sequentially processing the raw data received from the sensor by different signal processing modules. In this traditional approach, in which the signal processing modules are designed independently of each other in accordance with the purpose, the final image quality might be adversely affected due to distortion effects occurred at modules and carried throughout the camera pipeline. In order to eliminate these effects and to ensure the optimization of the camera bus from beginning to end, it is seen that researches on modeling the camera pipeline with deep network architectures have increased in recent years. Within the scope of this paper, the images in the Zurich dataset created for this purpose were edited using Topaz Gigapixel AI and Photomatix Pro 5.0 software, and the final model performance was examined by retraining the CNN (Convolutional Neural Network) based PYNET[1] network architecture, which is one of the pioneering studies in this field. Considering the PIQUE (Perception based Image Quality Evaluator) metric, which does not need a reference image, the performance of the created model is increased; at the same time, it has been seen that HDR tonemapped image can be constructed by revealing the details in the low and high exposure regions.