Multi-exposure image fusion using convolutional neural network

Purpose: The aim of the study is to obtain a new single high dynamic range image of a scene from the low dynamic range images of the same scene. Theory and Methods: In this study, a new MEF method using CNN is proposed. The proposed method is reasonably fast and effective. Since it is a costly task to create a new CNN model, a pre-trained CNN model was used. First, the fusion map employed for constructing the fused image was obtained from the source images using a pre-trained CNN model. The fusion map is then used to compute the fused image. In order to eliminate the saw-tooth effect in the fused images, weighting was performed on the fusion map. After that, a well-exposed fused image was constructed using the weighted fusion map. Results: The proposed method was applied to the MEF datasets, which are widely used in the literature, and the fused images obtained were evaluated using well-known quality metrics. The developed technique and other well-known MEF techniques are compared in terms of quantitative and visual evaluation. Conclusion: Convolution neural network, widely used and effective in multidimensional and nonlinear problem solutions, gives reasonably fast and effective results for multi-exposure image fusion. The results obtained show the feasibility of the proposed technique.