A Novel Approach to Low Light Object Detection Using Exclusively Dark Images

Abstract

The efficiency of our vision highly depends on the light’s intensity. In dark images, the intensity of light in our surroundings is generally lower, reducing the efficiency of vision and the capability to distinguish different objects. An analysis of lowlight images is possible with handcrafted and learned features. This process of object recognition also needs to take into consideration the intensity of light that is produced by a particular pixel varies depending on the color space used for a particular image since different colors produce different intensities of light. Therefore, the exclusively dark dataset has been used recently as a benchmark dataset for object recognition in the dark that contains 10 low light illumination types and 12 different categories of objects, and it has the potential to be used as the standard database for benchmarking research in the domain of low light. CSPNet is essential for the purpose of feature extraction. This reduces the computational load required by our model and also ensures that the accuracy does not significantly reduce. When it is coupled with the CNN, the results show potential for practical applications. The goal of this paper is to further improve the recognition rate of various objects in the dark.