Research On High Dynamic Range Imaging Algorithm Based On Frequency Domain Learning

With the widespread application of deep learning in recent years,high dynamic range imaging technology has made a breakthrough,but there are still many shortcomings in practical applications,such as problems such as inability to retain complete details when fusing multi-exposure images in dynamic scenes,and incomplete elimination of artifacts caused by foreground motion.This thesis takes high dynamic range imaging as the research background,and uses a combination of attention mechanism,multi-scale feature fusion and frequency domain learning to conduct an indepth study,the main research content and innovation points are as follows:Firstly,to address the problem of not being able to retain complete details when fusing multi-exposure images in dynamic scenes,this thesis proposes an algorithm based on attention mechanism and multi-scale feature fusion to extract multi-scale feature maps from low dynamic range image inputs and generate corresponding spatial attention maps to guide the reference image to identify unaligned parts,thereby alleviating ghosting in the fusion stage.With the help of the spatial attention map,the fusion module merges the mapping of the feature LDR images and ultimately reconstructs the high dynamic range output at different scales.Secondly,to address the problem of incomplete artefact elimination due to foreground motion,this thesis proposes a transform domain artefact elimination algorithm.The algorithm first transforms the feature map into the frequency domain and combines a learnable bandpass filter to learn a subtle frequency domain prior to separate out the artifact components and obtain a clean artifact-free HDR image.This thesis also uses a joint function combining Sobel loss and L1 error to optimise the algorithm.Experimental analysis demonstrates that the algorithm provides significant image quality improvement over the current optimal algorithm in the multi-exposure HDR image synthesis task,effectively eliminating artefacts and recovering high quality HDR images.In order to verify the contribution of each module,this thesis conducts a complete ablation experiment,which demonstrates the effectiveness.