Dark light enhancement for dark scene urban object recognition

Abstract

This paper presents a low‐light image enhancement method to improve the performance of autonomous piloting tasks based on deep learning methods. In the low light environment, camera sensors cannot capture enough effective photon signals causing poor performance in vision‐based autonomous piloting. Moreover, the lack of training data makes single‐frame low‐light enhancement and denoising algorithms hard to generalize in real‐world scenarios. By analyzing the noise patterns of real‐world cameras under low‐light environments, a noise generation method is proposed to mimic real‐world dark‐light noise and generate noisy‐clean data pairs for training. A method that can calibrate the camera shot noise parameters is then designed to learn low‐light enhancement from the synthesized noisy data pairs. The neural network trained on a synthesized dataset can effectively enhance the dark light image quality. Consequently, the network improves the downstream auto‐piloting applications such as object detection and semantic segmentation. Qualitative and quantitative experiments have been conducted to demonstrate that the method outperforms previous methods in low‐light enhancement.

Object recognition based on shape interest points descriptor

“…Introduction: Object recognition [1,2] is widely used in medical imaging, video surveillance, virtual reality, robotics, remote sensing, visual navigation and digital libraries. It includes mainly object feature extraction, representation and classification [3][4][5].…”