关键词: 人脸检测, 图像增强, 注意力机制, 损失函数, 低照度环境

Abstract: Abstract: Images captured in low-light environments suffer from low brightness, low contrast, and a lot of noise. At present, face detection technology has become mature, but the accuracy of face detection in extreme conditions such as low illumination is usually very low. Aiming at the problem that the detection performance of the face detection model is significantly reduced in low-light environment, a face detection method based on image enhancement is proposed. First, image enhancement methods are used to preprocess low-illumination images to enhance the effective feature information of human faces. Secondly, the attention mechanism is introduced after the model backbone network to enhance the network's attention to the face area, and at the same time reduce the negative impact of non-uniform illumination and noise. In addition, an attentional bounding box loss function (Wise-IOU) is introduced to improve the detection accuracy of the network for low-quality faces. Finally, a more effective feature fusion module is used to replace the original structure of the model. Compared with the original algorithm, the detection accuracy AP@0.5 of the improved model on the low-light face dataset DARK FACE is improved by 2.4 percentage points. Experimental results show that the proposed improved method can improve the face detection ability of the face detection model in low-light environments without introducing additional parameters and calculations. At the same time, the results on the other two low-light face datasets show the effectiveness and robustness of the method proposed in this paper, which is suitable for low-light face detection in different scenarios.

Key words: Keywords: face detection, image enhancement, attention mechanism, loss function, low illumination environment