Single-stage Face Detection under Extremely Low-light Conditions
Jun Yu, Xinlong Hao, Peng He
Abstract
Face detection has been well studied for many years. One remaining challenge is to detect faces from low-light images. The brightness of the image captured under extremely low-light conditions could be very low and the contrast will be severely reduced. It is easy to cause confusion during feature extraction and affects the performance of face detection. In this paper, we propose a single-stage low-light face detection method. First, we design an improved MSRCR method to increase the image quality under the condition of ensuring that the colors of the image are not distorted. It shows better enhancement effect than other methods in the DARK FACE dataset, especially the low-resolution face details are well preserved. There are a number of small, blurred and partially occluded faces. To address this, the Pyramidbox algorithm is a very effective face detection algorithm. Moreover, we conduct multi-scale tests to further develop the performance of the model and integrated the results through Soft-NMS method to obtain final results. Integrating these techniques, this paper has achieved high accuracy and obtained excellent results in the face detection task of the DARK FACE dataset.