改进Retinex方法在低照度壁(岩)画中的图像增强

Low Illumination Mural(cliff)Painting Image Enhancement Through Improved Retinex Method

随着计算机算力的快速提升,低照度图像增强在文物数字化保护方面已得到广泛应用。Retinex是重要的低照度图像增强算法之一,其假设图像可以分解为反射图与光照图,然而大多数基于Retinex的方法在对图像进行分解时都会出现结果不确定的问题。针对此类问题,提出了一种基于Retinex算法的改进网络模型,该网络模型用较少的网络参数,并结合了注意力机制,相较于Retinex-Net,不仅处理速度提升大约1%,而且更为重要的是较大程度上改善了Retinex-Net的生成结果表现过于生硬的情况;同时将预训练模型发布在了https://github.com/neemperor/Retinex-Adv-model上。还构造了一个壁画图像数据集LLM,可用于壁画类图像增强任务网络模型的训练和测试,使输出的模型更适合完成低照度壁(岩)画图像增强任务。

With the rapid improvement of computer computing efficiency,low illumination image enhancement has been widely applied to the digital protection of cultural relics.Retinex is one of the important low illumination image enhancement algorithms.It assumes that the image can be decomposed into reflection image and illumination image.However,most Retinex-based methods have the problem of uncertain results in decomposing the image.Aiming at this problem,this paper proposed an improved network model based on Retinex algorithm.This network model uses less network parameters and combines with attention mechanism.Compared with Retinex-Net,it improves the processing speed by about 1%,and more importantly,it greatly improves the situation that the generated results of RetineX-Net are too rigid.The pre-training model was published in https://github.com/neemperor/Retinex-Adv-model.This paper also constructed a mural painting image data set LLM,which can be used to train and test the network model of mural image enhancement task,so that the output model is more suitable for low illumination mural(cliff)painting image enhancement task.