Zero-DCE网络的自适应损失函数改进

Improvement on the adaptive loss function for the Zero-DCE network

针对轻量化微光增强网络Zero-DCE在处理亮度变化范围较大的微光图像时,存在不同区域亮度增强不一致导致的图像不清晰问题,本文提出了一种基于伽马变换的自适应损失函数,在原损失函数的基础上降低了网络对图像曝光差异的敏感性,明显改善了微光增强效果。该方法通过在卷积神经网络(convolutional neural network, CNN)中添加CBAM模块提高网络对微光图像特征的表达能力,使用网络增强图像灰度平均值与增强特征图均值的对数距离作为伽马变换自适应系数,最后计算网络增强图像和伽马变换后的图像之间的灰度参数距离。实验表明,与原网络相比,改进后的方法处理效果提升显著,其中在图像评价指标方面,均方误差提升9.7%,峰值信噪比提升13.8%,结构相似性提升6.7%。

For light-weight low level light intensifying network,blurred image issue caused by inconsistent light intensifying degree in different area can occur when Zero-DCE handles the low level light image with a bigger brightness variation range.This paper introduces a self-adaptive loss function based onγtransform,on the basis of the original loss function,decreases the sensitivity of the network on image exposure difference and dramatically improves the low level light intensifying effect.In this method,CBAM module is added into the convolutional neural network(CNN)to increase the expression ability of the network to low level light image feature,in addition,the logarithm distance between the average value of gray level of the network intensifying image and the average value of intensifying feature image is selected asγtransformed self-adaptive factor,and finally,the gray level parameter distance between network intensifying image andγtransformed image is calculated.The experiment shows that the performance of this method is dramatically improved comparing to the original network,in which in aspect of image evaluation index,the error mean square is increased by 9.7%,the peak signal to noise ratio is increased by 13.8%,and the structure similarity is increased by 6.7%.