Research On Low-illumination Images Enhancement Algorithm Based On Streaming Model

With the increasing demand for computer vision technology,digital image processing has been widely applied in various fields,such as industrial production,medical imaging,video surveillance,intelligent transportation,etc.However,under low-light conditions,the color of the image is easily distorted and information is lost.For fields such as medical,intelligent transportation,and national defense,the loss of image information may cause serious consequences.Therefore,low-light image enhancement has become one of the essential technologies in these fields.Existing classical image enhancement techniques include methods based on grayscale value transformation,histogram equalization,and image fusion,but these methods often fail to accurately model the complex distribution of conditions from lowlight images to normally exposed images,resulting in excessive brightness,residual noise,and artifacts in the enhanced images.Although the deep neural network technology represented by the stream model has a better enhancement effect than the previous classical algorithm,the existing stream model network usually has two problems:(1)there is a certain uncertainty in the process of enhancing the lowillumination image to the normal exposure image,and the optimal image enhancement performance cannot be obtained;(2)The network stability of the flow model will decrease when there are many network layers;The above two problems greatly limit the application of stream model networks in real systems.Aiming at the above two problems,this paper first proposes a multi-channel parallel convolutional flow model to model the complex conditional distribution from the original low-illumination image to the enhanced normal exposure image,so as to improve the performance of the stream model for the enhancement of low-illumination images.Secondly,a reversible network structure of the flow model based on reversible residual connection is proposed to improve the stability and training efficiency of the flow model network in the case of a large number of network layers.The main research work and innovation points of this paper are as follows:(1)A low-light image enhancement algorithm(LRFlow)based on a multi-path parallel convolutional flow model is proposed.In this algorithm,the traditional flow model network’s conditional encoder is improved by using a multi-path parallel convolutional structure.A neural network structure with different-sized convolutional kernels is used to model the complex conditional distribution in low-light image enhancement,which better describes the one-to-many mapping relationship between low-light images and normal exposure images.Next,the conditional encoder with a multi-path parallel convolutional structure is used to generate corresponding features based on the input low-light condition information,and the illumination-invariant color map is extracted based on the Retinex theory,which serves as the prior information for low-light image enhancement.Then,a reversible network learns the conditional distribution of generating normal exposure images based on the features of low-light images,and finally generates the enhanced images through inverse mapping.The advantage of this model is that the inverse mapping direction in the training process is also constrained by the loss function of the flow manifold structure of normal exposure images,ensuring the accuracy of the inverse mapping results.Experimental results show that compared to other existing methods,this method achieves higher peak signal-to-noise ratio(PSNR),higher structural similarity(SSIM),and lower perceptual similarity(LPIPS)on both LOL dataset and VE-LOL dataset,successfully improving the performance of low-light image enhancement.(2)A low-light image enhancement algorithm(Res Flow)based on a reversible residual connection flow model is proposed.This algorithm is mainly designed to solve the stability problem of existing flow model networks and improve the training efficiency of flow model networks.In this algorithm,the reversible residual connection module is introduced into the reversible network part of the traditional flow model structure.As the reversible residual structure can better extract the illumination component details in low-light images,using this reversible residual connection in the reversible network of the traditional flow model can maintain the stability of the flow model network when the network layers are too many,reducing the possibility of gradient explosion as the network layers increase.In addition,this reversible residual connection can achieve better low-light image enhancement performance than the traditional flow model in cases of limited training data and further improve the training efficiency of network parameters.In the experiments of this paper,when the network layers are more,the Res Flow algorithm achieves better stability in low-light image enhancement and faster training speed than the traditional flow model network on the LOL dataset.In addition,in the case of small-scale training datasets,the algorithm also obtains less perceived similarity(LPIPS)and a more stable generation process than traditional stream model networks.