Research On Deep Learning Based High Dynamic Range Reconstruction Algorithm

With the continuous development of multimedia information technology and digital image technology,high dynamic range images are widely used in video,games,medical detection and other industries,which have great commercial and scientific value.However,due to the limitation of the dynamic range of the imaging sensor,a high-quality high-dynamic-range image cannot be obtained.It is generally more expensive to capture high dynamic range images by improving hardware performance.Therefore,how to make up for the shortcomings of the camera's dynamic range through algorithms to produce high dynamic range images is an urgent problem.Based on the above background,the work done in this thesis mainly includes the following two points:Aiming at the problems of loss of high dynamic range image details and insufficient brightness constraints caused by sensor dynamic range overflow,an algorithm for high dynamic range image reconstruction of weak motion scenes based on deep learning is proposed.The algorithm consists of a mask network and a backbone network.The mask network first extracts the reflection map and illumination map of the multi-frame low dynamic range image through a convolution separation module designed based on the theory of human eye retina,and then the attention guidance module based on the residual unit design is used to further extract high-level information in the illumination map and the reflection map to provide a spatial monitoring mechanism for the backbone network to better restore the details and brightness of the image.The backbone network first down-samples the image through the encoder to reduce noise and deformation while obtaining low-frequency features of the image,then uses the fusion block design based on the residual block to fully interact with the features,and finally uses the decoder sampling reconstructs the target image.In addition,in order to reduce the high-frequency information lost during the encoding process,a skip connection is added to the backbone network part so that the high-frequency information can be effectively multiplexed during the decoding process.Aiming at the problem of large moving targets in the scene,which causes artifacts and color patches in the multi-exposure fusion result,this paper proposes a high dynamic range image reconstruction algorithm based on deep learning for target significant motion scenes.The entire network consists of an alignment part and a fusion part to form an end-to-end model.In alignment network,firstly,the optical flow between the two low dynamic range images with different exposures and the reference image is estimated through a convolutional neural network.Secondly,based on the idea of the spatial converter,the estimated optical flow information is used to register it with the reference image to reduce the artifacts caused by the pixel offset on the fusion result,while maintaining the rotation invariance of the network dynamically.In the fusion network part,the network framework proposed in Chapter 3 is used,and the registered image is spliced with the corresponding original image channel as input,which reduces the pixel offset and ensures the expansion of the dynamic range of the reconstructed image.Meanwhile,in order to optimize the network performance,a loss function is designed to accelerate network convergence and achieve better reconstruction results.The experimental comparison results show that compared with the existing high dynamic range image reconstruction algorithms,the proposed algorithm has significantly improved the objective indicators PSNR,SSIM,and HDR-VDP-2.From a subjective perspective,the restoration of details in highlights and shadows is better,the phenomenon of color patches and afterimages is greatly improved,and the dynamic range of the reconstructed image is greatly expanded.The color gamut space is also richer.