ExDark-DETR: Enhancing Object Detection in Low-Light Images using a Transformer-based Framework

This paper presents the idea of enhancing object recognition in Exdark dataset in the absence of sufficient illumination by training the identification Transformer (DETR) model. The ExDark dataset, comprising a vast majority of low illumination images, poses significant challenges to conventional object detection algorithms. The DETR model has been fine-tuned and parameters have been adjusted as per requirements of Exdark dataset. Also sophisticated preprocessing method such as CLAHE is applied along with data augmentation techniques during the course of training to remedy problems such as noise and poor contrast. Benchmarking experimental results against existing baselines reveals that the new DETR model compares favorably against standard detection performance metrics, notably mean average precision.