基于知识蒸馏的夜间低照度图像增强及目标检测

Nighttime low-light image enhancement and object detection based on knowledge distillation

为了实现夜间低照度图像的增强,提高目标检测模型在夜间低照度条件下的检测精度并减小模型的计算成本,提出了一种基于知识蒸馏和数据增强的夜间低照度图像增强以及目标检测多任务模型,基于高质量图像模型进行知识蒸馏,利用高质量图像的特征信息指导模型训练,从而使模型在夜间低照度图像中提取到与高质量图像类似的特征信息。利用这些特征信息可以实现图像的对比度增强、去噪以及目标检测。实验结果表明,提出的蒸馏方法可以提升16.58%的夜间低照度目标检测精度,且用该方法增强的图像可以达到主流的基于深度学习的图像增强的效果。

In order to enhance the quality of nighttime low-light image,improve the accuracy of the object detection model under the nighttime low-light condition and reduce the calculation cost of the model,a multitask model for nighttime low-light image enhancement and object detection based on knowledge distillation and data enhancement was proposed.Knowledge distillation was performed based on the high-quality image model,and the feature information of the high-quality image was used to guide the model training,so that the model could extract the feature information similar to that of the high-quality image in the nighttime low-light images.These feature information could be used to achieve enhancement of image contrast,denoising and objects detection.The experimental results show that the proposed distillation method can improve the object detection accuracy of nighttime low-light by 16.58%,and the image enhanced by this method can achieve the effect of mainstream image enhancement based on deep learning.