Research On Image Enhancement Algorithms Under Non-uniform Illumination Conditions Based On Retinex Theory

Due to the limitations of lighting conditions and the dynamic range of digital devices,the captured photos usually suffer from degradations such as low visibility,low contrast and noise amplification.Such low-quality images are not only visually poor,but also difficult to support advanced computer vision tasks in the later stage.In view of the above problems,in the traditional direction,based on the limitations of the Retinex principle,this thesis proposes a Retinex-based nonuniform illumination image enhancement algorithm,which is structure and texture aware;in the direction of deep learning,based on the shortcomings of Retinex-Net network model,such as low color fidelity and amplified noise,a non-uniform illumination image enhancement network is proposed to improve Retinex-Net.The main work and innovations of this thesis are as follows:(1)The traditional Retinex algorithm is difficult to accurately estimate the illumination map and reflection map,and at the same time,the enhanced image may have low naturalness and low contrast.Aiming at the above problems,this thesis proposes a texture and structure aware image enhancement algorithm based on Retinex for non-uniform illumination.The algorithm mainly includes three parts,namely Retinex decomposition based on texture and structure perception,contrast enhancement based on contrast limited adaptive histogram equalization and histogram correction,and illumination adjustment based on look up table.After the observed image passes through the exponential local average variation filter,the weighted method is used to realize the Retinex decomposition of texture and structure aware,and the more accurate illumination components and reflection components are flexibly estimated at the same time.Then,the content-based adaptive contrast enhancement is performed on the sub-blocks of the reflection map by means of histogram correction.Finally,the brightness is adjusted by learning look-up table.The experimental results show that the proposed algorithm has stable performance and good applicability.While maintaining the naturalness of the image,there is no degradation phenomenon such as halo artifacts,excessive enhancement and color distortion.(2)Inspired by the Retinex-Net network model,this thesis proposes a non-uniform illumination image enhancement network that improves Retinex-Net for its problems such as color distortion and amplification noise.The network consists of three modules,namely the decomposition module and the reflectance denoising module and illumination adjustment module.For the decomposition module,dual branches and residual connections are used to maintain more image details,and the illumination consistency loss function is added to further suppress the weak edge information of the image and retain more strong edge structure information;for the reflectance denoising module,Double residual structure is used to retain more details and textures,and channel feature attention module is introduced to learn the key content of the image and reduce the response to irrelevant features such as noise;for the illumination adjustment module,in order to alleviate the problem that the local area of the illumination component is too dark or too bright,upsampling is used to reconstruct the illumination intensity,and at the same time,in view of the fact that the color distortion of the enhanced image exists in the Retinex-Net network,an attention module and a color loss function are added to the illumination adjustment modul.The experimental results reflect that the algorithm in this thesis gets better results in improving the image contrast and the details of the saturated area,and there is no obvious color distortion and large-area artifacts,so the naturalness of the image is higher,and compared with other contrast methods,the PSNR and SSIM index values of the algorithm in this paper are the largest.