A Deep Unsupervised Residual - Convolutional Auto-Encoder: Multifocus Image Fusion based on Local Binary Pattern

Abstract

Image fusion has become a frequently used approach for improving the visual awareness for various applications such as medical diagnosis, surveillance, military , computer vision, robotics, enhanced vision systems, biometrics, and remote sensing and many. From literature, most of the Image fusion technique requires sophisticated pre-processing and post-processing steps to create image patches to train the network, generate the focus map, consistency verification but the fused image results in blocking artifacts and smoothening effects. This paper focuses on simple novel Deep Unsupervised Residual-Convolution Auto-Encoder Network (R-CAEN) to address these limitations. R-CAEN is trained with proposed Euclidean Loss Function in keras-tensorflow platform. The encoder-decoder model learns to extracts the deep features from the source images in a residual approach. A normalized Local Binary Pattern (LBP) codes for the corresponding encoder features are calculated for fusion. The decoder decodes the fused features and the fused image is evaluated with the objective measures: Correlation Coefficient, Peak Signal to Noise Ratio, Measure of Structural Similarity Index, Entropy and Visual Information Fidelity. The subjective and objective analysis of R-CAEN is compared with six existing methods and the results shows that the performance of R-CAEN outperforms the existing methods.