Publication

Abstract : Scene depth and the corresponding ambient lights generally degrade low-light photos taken in a busy setting with uneven lighting.Vehicle object recognition is made more challenging by this degradation, which results in a significant loss of object details in the deteriorated picture format due to the poor contrast and the appearance of artificial light. On the other hand, current methods for identifying prominent objects rely on the unreasonable assumption that the photographs were captured in an adequately lit atmosphere. In this research, we refer to a method for improving images in dim light image detection of vehicle objects. The suggested approach immediately embeds the physical lighting model into the convolution neural network, where the traffic environment light is represented as a point-wise fluctuation that varies with local content, to explain the degradation of low light images. A Sensor Filter is also used to record the difference between an object’s local content and its local neighborhood hotspots. We generate a low light Images dataset with pixel-level human labeled ground-truth annotations for quantitative assessment, in addition to our benchmark dataset. We also show encouraging results on four other publically available datasets.