一种特征自我保留的弱光图像增强方法

A Low-Light Image Enhancement Method with Feature Self-Preservation

现有的多数图像增强方法通常整体增强亮度通道,会导致过度增强、细节丢失及颜色失真等问题。为克服这些问题,提出一种基于生成式对抗网络(Generative Adversarial Networks,GAN)和特征自我保留的弱光图像增强方法SFPGAN。首先从颜色、亮度及纹理3个方向评判生成图像的真实性,其次引入特征自我保留损失以保留原始图像的特征,最后使用含有一定量正常亮度和过度曝光的图像训练模型使模型获得较强的鲁棒性。大量实验证明,提出的方法在视觉质量和客观指标上都优于其他方法,并且更适应真实的图像。

Most existing image enhancement methods usually enhance the brightness channel as a whole,which usually leads to problems such as excessive enhancement,loss of detail,and color distortion.To overcome these problems,a low-light image enhancement method SFPGAN based on Generative Adversarial Network(GAN)and feature self-preservation was proposed.Firstly,the authenticity of the generated image is evaluated from color,brightness and texture.Secondly,the loss of feature self preservation is introduced to retain the features of the original image.Finally,an image training model with a certain amount of normal brightness and overexposure is used to make the model more robust.A large number of experiments show that the proposed method is superior to other methods in visual quality and objective indicators,and more suitable for real images.