Research On High Dynamic Range Image Synthesis Algorithm Based On Dynamic Scene

In the real world,the dynamic range of the scene can reach dozens of magnitude.However,due to the constraints of hardware conditions,the existing image acquisition and display devices can only display a limited dynamic range,which leads to the obtained image could not fully represent all the information of the scene.In order to record the shooting scene more realistically and completely,the existing algorithms often use a method of combining multiple different exposure images into one high dynamic range image to improve the dynamic range of the image.When there is a moving object in the shooting scene,directly merging images of different exposures will result in ghosting in the generated high dynamic range image,which seriously affects the quality of the image.Aiming at the possible interference object intrusion and main object motion in the scene,the corresponding high dynamic range image synthesis algorithm is designed to solve the ghost problem that may occur.Aiming at the interfering objects that may occur during the shooting process,this thesis proposes a multi-exposure image synthesis algorithm based on background modeling.In order to completely remove these interferences,this thesis first uses the median filtering in the time domain to perform background modeling and uses the obtained background model as a reference image.The local background difference method and the median threshold bitmap detection method are used to perform double detection of ghost on each input image.The double detection can effectively compensate for the deficiency of single detection and improve the detection accuracy.Then,the zero-mean normalized cross correlation coefficient is used to guide the detection of the motion region and remove the ghost region of the misjudgment,the information entropy factor is introduced to correct the weight map,thereby further improving the dynamic range of the resulting image.Finally,the multi-scale pyramid fusion algorithm is used to fuse the input multi-exposure image with the weight map.Aiming at the situation that the moving object is the main object during the shooting process,this thesis proposes and builds an end-to-end convolutional neural network based on the VGGNet and the concept of automatic encoder.In this thesis,the input image is simply registered by optical flow method to generate a set of structurally similar image copies,and use this as the input of the network.In order to ensure the integrity of the high-frequency information and low-frequency information of the synthesized result image,a skip connection layer is added at the end of each set of convolutional layers,so that the high-frequency information lost during encoding can participate in the reconstruction process of the high dynamic range image.By training the network with the data set,the network can directly generate a high dynamic range result image without ghost,and finally map it to the display device through the tone mapping algorithm.The thesis evaluates the algorithm results from both subjective and objective aspects.The experimental results show that the two synthetic algorithms proposed can effectively remove ghosts generated by moving objects,and can generate high-quality images with better visual effects while expanding the dynamic range of images.