Investigation Of High Dynamic Imaging Technique Based On Convolutional Neural Network

Limited by the response characteristics of camera imaging devices,the dynamic range of digital images acquired by most common imaging systems is small,which is not conducive to the analysis of scene information.High dynamic imaging technology can generate high dynamic images closer to the real natural scene by using signal processing technology to reconstruct the details of over-exposed and under-exposed regions in the low dynamic range images without changing the physical structure of the imaging system.At present,high dynamic imaging technology has been widely used in military,medical,photography and gaming fields.The existing high dynamic images are mostly obtained by multi-exposure fusion,but the calculation time is long to meet the real-time requirements,and artifacts are generated in dynamic scene,affecting the image effect.Therefore,it is of great research significance to use a single frame of low dynamic range image for high dynamic imaging.Traditional single-frame high dynamic imaging methods are usually model-driven.Whether it is based on the camera imaging principle or inverse tone mapping,it contains a large number of parameters.Each scene requires precise tuning,which is difficult to deal with frequent scene switch.In view of this,neural network methods with excellent feature extraction capabilities and end-to-end parameter free operation have become current research hotspots.Existing neural network methods are mostly based on the U-Net structure.The downsampling operation will lose detailed information,and the upsampling operation will cause checkerboard artifacts.The imaging effect is difficult to meet the application requirements.In order to solve the above problems,this paper proposes a single-frame high dynamic imaging method based on a multiscale convolutional neural network model.The model uses a multi-scale feature extraction unit to extract low-frequency,intermediate-frequency and high-frequency features of the image respectively.And the network can selectively suppress or enhance overexposure and underexposure regions by comparing local information with global information.In addition,a feature refining unit based on the attention mechanism is specially designed to select multi-scale features from channel domain and spatial domain,which can remove noise and overexposure features while enhancing the effective features.Experimental results show that this method is superior to existing methods in both subjective and objective evaluation.In addition,this paper also incorporates the potential prior knowledge of images into the network,and proposes a high dynamic imaging method based on traditional prior and deep learning.This method utilizes undecimated wavelet transform to obtain multi-directional high-frequency detail features of the luminance component,and then fuses the transform domain and spatial domain features to generate a more accurate luminance gain surface.Finally,the luminance gain surface is utilized to perform an equal map on the three RGB channels to prevent color distortion.Considering that the bright and dark channels contain rich details and structural information in the underexposed and overexposed regions,this method introduces the bright channel prior and dark channel prior to reconstruct the details of the poorly exposed regions respectively.In addition,a new loss function based on structural similarity and cosine similarity is designed to optimize the network to better preserve the image structure and color information when expanding the image dynamic range.The experimental results show that the high dynamic image generated by the method can more accurately reflect the actual situation of natural scenes with richer details and clearer hierarchy.And both subjective and objective evaluation are superior to existing methods.