Low-light True Color Image Enhancement Algorithm Based on Adaptive Truncation Simulation Exposure and Unsupervised Fusion
Han Yongcheng,Zhang Wenwen,He Weiji,Chen Qian
DOI: https://doi.org/10.3788/gzxb20235209.0910002
2023-01-01
Abstract:Low-light true color image enhancement is an important branch in image processing. The images obtained in low-illuminance environments often have low brightness, low contrast, noise, and color distortion. Due to the complexity and diversity of target scenes and imaging equipment, it is difficult to directly obtain satisfactory high-quality images in low-illumination environments. There are many problems in the information content of these low-light true color images, and the image is not good for viewing and is not conducive to advanced image tasks in the later stage. Aim at solving the problems of detail loss in the dark area and excessive enhancement in the bright area, a low-light true color image enhancement algorithm based on adaptive truncation simulation exposure and deep fusion is proposed. The algorithm has the function of brightness suppression. We design an adaptive truncation simulation exposure method and use an unsupervised network model to fuse the exposure sequence to achieve a flexible and efficient fusion of multiple exposure images of fixed size. First, an exposure sequence about the original low-light true color image is generated by simulation, and then the convolutional network is used to learn the weight map corresponding to the exposure sequence. We can obtain the final enhanced results by weighted fusion within the network. Most classic simulated exposure algorithms either map the image linearly or use existing enhancement algorithms such as histogram equalization. The number of simulated exposures is often determined artificially in pursuit of as many exposure sequences as possible covering different brightness levels, which results in many redundant images in the simulated exposure sequence. To address these problems, we propose an adaptive gamma correction method which can effectively avoid this redundancy. The light and dark regions of the image are segmented first, then truncated adaptive gamma correction is carried out. Finally, the appropriate exposure sequence is obtained by guided filter denoising. After obtaining the multi-exposure sequence, an efficient fusion method is necessary. At present, the fusion method for single low light image enhancement is mainly weighted hierarchical fusion, which has large computation and low robustness. It is easy to produce halo and seam phenomenon. To address these problems, we propose an unsupervised network model, including a context aggregation network based on dilated convolution which can achieve low resolution weight map efficiently, and a deep guided filter that can strike a balance between image quality and efficiency. And the final enhanced result is obtained by the weighted average of the simulated exposure sequence and the weight map. To verify the superiority of the algorithm, we collect vast low-light true color image datasets and compare the enhanced results with many state-of-the-art low-light image enhancement algorithms from subjective and objective perspectives. And a laboratory environment dataset is collected using a low-light night vision camera and a three-channel true color camera. Experimental results show that the algorithm we proposed has the best NIQE scores of the public datasets, and the best PSNR and SSIM scores of the laboratory environment dataset, among which NIQE is reduced by 4.49, PSNR is increased by 4.28 and SSIM is increased by 1.94%. In addition, the color reproduction effect of the algorithm is very good, and the color difference of the proposed algorithm is the smallest. At 8. 71x10(-2) lx illuminance, the color difference of the proposed algorithm is reduced by 14.83% than the suboptimal algorithm, and at 1.02x10(-2) lx, it is reduced by 3.05%. The algorithm proposed can significantly improve the brightness and contrast of the image, has good robustness, and will not produce excessive enhancement. It can effectively restore the image details while taking into account the color information and enhance the results to be true and natural.