New approaches for multi-focus image fusion

Multi-focus image fusion is the process of combining two or more images of the same scene with different focus points to create a all-focus image. In this thesis, two unique approaches are proposed to obtain near-perfect all-focus images. In the first approach, the detail information of the source images is increased by bicubic-based super-resolution method and separated into subbbands with the help of Stationary Wavelet Transform. As a result, images are combined with the gradient-based fusion rule. The second proposed approach is based on learning that includes innovations at every step. For this approach, a rich dataset consisting of two or more color source images is created. Also, a new CNN model is designed to classify focused and unfocused parts. The source images are first divided into 8x8 overlapping blocks, which is the input layer size of this proposed architecture, and given to the proposed network, and initial decision maps are created for each source image. Then, with the help of the improvement mechanism based on focus metrics, which has not been used before in the literature, the cases where the proposed network is unstable are minimized. After this step, morphological operations are applied to the initial decision maps and final decision maps are created. Finally, the fused images are created with the help of the fusion rule with dynamic decision mechanism. It is seen that the approaches are quite successful in multi-focus image fusion.