Research On Multi-exposure HDR Imaging Algorithm Based On Convolutional Neural Network

In the natural world,the range of the radiation value of the scene is extremely wide.Due to the limitation of software and hardware,general imaging equipment can only display brightness changes of two orders of magnitude.Therefore,when the equipment performs non-linear compression on the radiation distribution of the original scene,the scene information beyond the display range is lost,and it is impossible to truly restore the scene seen by our eyes.High Dynamic Range Imaging(HDRI)is an image type that can more realistically represent the wide range of brightness changes.As people began to have more extreme pursuits and expectations for image quality,high dynamic range images gradually entered our vision.Multi-exposure synthesis method is to extract information fusion from multiple images with different exposure levels to get HDR image,which is currently the most extensive and effective method.However,traditional image processing methods rely too much on pixel alignment to avoid the problem of artifacts in dynamic scenes.Given that deep learning methods can learn the reconstruction of dynamic regions from a large number of samples and they have excellent performance in complex mapping tasks,this paper proposes a multi-exposure synthesis model based on convolutional neural networks.The research contents are as follows:(1)A three-layer cascade network is designed to generate HDR images.The low dynamic range images of different exposure levels are input to the network in sequence,and the reference frame image is preferentially processed through the first layer branch,which effectively avoids the mixing artifacts caused by the foreground displacement.The second branch and the third branch progressively extract the over-bright and over-dark information contained in the low-exposure image and the high-exposure image,which can steadily improve the quality of the synthesized HDR image.(2)Based on the cascade network,the use of long-short-term memory network and dilated convolution improves the model generation effect,and thus proposes the cyclic cascade network.The network strengthens the contrast of the differences between the multi-exposure images and helps to extract features at different levels.For the improved network,the single-frame data set is expanded to multi-frame samples for training,which overcomes the long-term problem of a single training set in the multi-frame HDR generation task,making the network model more robust.In the experimental part of this paper,a variety of HDR imaging methods are compared with our method,and the subjective image quality and objective image quality are evaluated respectively.The results show that the three-layer cascading network structure has good anti-artifact ability for HDR synthesis in dynamic scenes,and can reconstruct the texture details of the image and maintain good contrast.Through further data enhancement,the network can ensure the quality of HDR synthesis in scenes with different data distributions,and has further improved the accuracy of tones.