Research On Low-light Image Enhancement Algorithm Based On Retine

In recent years,due to the extensive application and development of digital image in people’s life,digital image has gradually become one of the important ways of information transmission in people’s life and work.However,due to the influence of lighting conditions and photography equipment,the degradation phenomenon of low brightness and loud noise often appears in the shooting image,which has a bad effect on the image quality.To a certain extent,the detailed texture of the image is blurred,resulting in poor visual effects for human eyes.Due to the adverse factors above,the image with low illumination degree was formed.Based on Retinex theory,two improved Retinex algorithm models were proposed respectively from the direction of classical algorithm and the direction of deep learning.The main research contents of this paper are as follows:(1)Aiming at the unbalanced enhancement effect of Retinex algorithm in bright and dark areas,an adaptive image enhancement algorithm based on Yiq-Retinex model was proposed.YIQ color model was introduced and the Y-component was enhanced by the improved Retinex algorithm in the wavelet domain.The nonlinear mapping function is used to project the I and Q components of the image to improve the color saturation.A new variation coefficient is used to classify the image,and two subsets with different structure information are obtained.In the subsets with complex changes and the subsets with uniform changes,different weight coefficients are used to fuse the corresponding branches of the Gaussian filter function.To solve the problem of image noise,an improved BM3 D algorithm is used in the image preprocessing part,which can adjust the size of reference block and the diffusion effect of generalized total variation according to the coefficient of variation of the image.The denoising effect is better than other traditional functions.Experiments show that this algorithm has better image processing effect than SSR,MSR and HE algorithms,and has certain advantages compared with subjective evaluation and objective indicators.(2)Aiming at the problem of uneven brightness of low illumination image,a Retinex-net network based on multi-scale feature extraction was constructed based on deep learning neural network model.The network model consisted of three modules: multi-scale feature extraction module,illumination component module and reflection component module.Firstly,Gaussian pyramid is used to extract multi-scale feature of the image,and illumination component module is used to estimate the illumination of each scale feature image.Among them,the illumination estimation model is based on the optimization model to solve the inherent under-exposure of low light images,and expand their optimization process to obtain a larger sensitivity field.Finally,the fusion network is used to fuse the feature images of various scales to obtain the final illumination estimation map.When the reflection component passes through the denoising network,the feature graphs of other scales are added into the denoising network to achieve detail compensation.Finally,the channel attention module is used to extract more effective feature mappings and get the reflection component.The results of ablation experiments prove the effectiveness of each module of the new algorithm.Compared with the current advanced image enhancement network algorithms,the new algorithm has a better effect in protecting image details and color saturation,and the objective index parameter of the new algorithm is the maximum.