Max Olli Blom and Thomas Johansen, MPCAS

In recent years, convolutional neural networks have outperformed previous state-of the-art methods on several computer vision tasks. However, prior to the network training, there are in general several non-trainable image processing steps that map the raw sensor data to the input image, and even though each of these sequential steps can be individually hand-tuned to maximize the perceptual image quality, since they are non-trainable, they are not optimized for computer vision nor the end task of the network. Our work investigates the possibility of removing and/or replacing traditional image signal processing operations using a deep learning approach with the goal of increasing performance and reducing the runtime of an object detection system. We show that the use of raw camera sensor data together with end-to-end trainable image processing yields an increase in performance, even with very few operations. Our results show the potential of this approach, which easily generalizes to other areas of computer vision. ​