Research On Low-light Image Enhancement Based On Variational Autoencoder

The low-illuminance environment causes the image quality to be degraded,with more image noise and lower contrast.The effect is not ideal when the image used for image classification,target recognition,image understanding and analysis,super-resolution reconstruction and other image processing.Therefore,it is necessary to enhance the illumination of this type of image,that is,increase the overall and partial contrast of the image,denoise,and appropriately adjust the image background and edges.This thesis uses variational autoencoder as a key technology to study low-illuminance image enhancement,aiming to analyze the characteristics of low-illuminance images from different angles,combine various technologies and methods to improve both the image structure,the contras and the details of low-illuminance images,and finally obtain high-quality images with complete structure,natural colors and clear details.The following three image enhancement methods are mainly proposed.Aiming at the multiple distortion characteristics of low-illuminance images(low brightness,multiple noise and blurring,etc.),a Multiple Reconstruction-Variational Auto Encoder(MR-VAE)is proposed based on the variational autoencoder,which is gradually generate high-quality low-illuminance enhanced images from coarse to fine.The MR-VAE is composed of three modules: Feature Probability Distribution Capture(FPDC),Global Reconstruction(GR),and Detail Reconstruction(DR).The core idea is to reconstruct global features and local features in stages,and gradually solve multiple distortion problems.The GR is used to builds image global features and improves the global brightness to obtain a rougher image.The DR weighs denoising and deblurring to generate images with more realistic details,less noise and more suitable local brightness.In addition,a multiple loss function,which is defined to replace the VAE loss,guides the network to generate highquality images.The experimental results show that the design of multiple reconstructions and multiple loss functions improve the performance of the network to generate complex images and handle multiple distortion low-illuminance images,and improve the quality of the generated images,signal-to-noise ratio and visual characteristics.The task of low-illuminance image enhancement requires operations such as color restoration,denoising,and illumination enhancement at the same time,but most of the current work only considers illumination enhancement,so it is difficult to generate highquality enhanced images.In order to solve this problem,a novel deep probabilistic framework that integrates both attention mechanism and context encoding into a unique variational autoencoder(ACE-VAE)is proposed based on the variational autoencoder.The key modules of ACE-VAE include a skip refinement modulue(SRM)and a context encoding module(CEM).The SRM uses the attention mechanism to weight skip information to filter noise and extract important features.The CEM consists of an attention feature pyramid(AFP)and a global context extractor(GCE).Among them,AFP is used to encode effective context information and adaptively assign different weights to local areas;GCE is used to capture global context information.In addition,a polynomial loss function including content generation,denoising and basic variational autoencoder loss is proposed,and the loss function is used to guide the network to achieve a better balance between enhancing illumination and maintaining proper saturation.Experiments show that ACE-VAE can obtain better performance.For the adaptive enhancement problem,on the one hand,learning the long-distance dependency between pixels is beneficial for the network to judge the illuminance of the surrounding area of a pixel because the low-illumination region is scattered;on the other hand,the shape and area of the low-illumination region are diverse,so learning the multiscale representation ability is beneficial for the network to capture such features.Therefore,by combining VAE with Transformer and introducing multi-scale feature extraction and selfcalibrated convolution,trans-VAE is proposed to be used for low illumination image enhancement.Different from MR VAE and ACE-VAE,trans-VAE has more efficient performance of low illumination enhancement.The strategy of "one step in one place" is adopted to enhance the low illumination image,which reduces the algorithm complexity and improves the enhancement efficiency.Trans-Vae is mainly composed of BSA(Bottleblock of Scale Aggregation)and BSAT(Bottleblock of Scale Aggregation with Transformer).BSA is used to capture multi-scale features.BSAT is obtained by replacing the ordinary convolution layer of BSA with Transformer,which is different from standard Transformer,which obtains Query,Value and Key through simple linear transformation.In this paper,global average pooling and adaptive pooling are used to compute Query and Value,so as to obtain global and local context representation,and promote the network to enhance local region adaptively according to global information.Experiments have shown that trans-Vae can effectively enhance ultra low light images due to classic and latest low light image enhancement algorithms.