Survey on Image Fusion: Hand Designed to Deep Learning Algorithms

Abstract—Image fusion is the process ofcombining/integrating multiple images to generate the singleimage having all meaningful information. The input/sourceimages are captured through various sensing devices underdifferent parameter setting. It is impossible to focus on allinformation or all small objects in single image. Hence, Imagefusion methods provide the composite image known as fusedimage with complementary information. Fused image should bemore suitable for human as well as machine perception.Therefore, several methods have been developing to improve thequality of images. Traditional methods include spatial andtransform domain based image fusion in which spatial domainincludes the fusion methods with pixel, blocks or segmentationbased processing. Whereas, transform domain utilizes thetransform theories to transform the image in another domainrather than same domain of the input and performs fusion ruleon transformed image. Spatial domain methods produce spatialand spectral distortion in the fused image whereas transformmethods often perform inadequately when images obtained fromdifferent sensor modalities. Recently the concept of deep learninghas enlarged to enhance the development in image processingand computer vision problems such as segmentation,classification, super-resolution, etc. Deep learning algorithmssuch as convolutional neural network (CNN), deep autoencoder(DAE), and deep belief networks (DBN) with different categoryof images such as multi-modal, multi-resolution, multi-temporaland multi-focus have been proposed for image fusion. Itsapplications include disease analysis, disaster assessment,providing a complete information for diagnosis, detection ofchange, etc