A Research Of Image Fusion Algorithms Based On Deep Learning

Image fusion is the process of merging two or more images obtained by different sensors into one image after precision matching in the same scene.The fused images not only preserve all the important information of the source image,but also are clearer than the image presented by a single sensor.The information is richer and easier to be understood by humans or machines.Because of these advantages,image fusion has been widely used in many fields,such as digital images,medical images,remote sensing images,and the machine vision.Deep learning model is built on probability statistics and applied mathematics.Thanks to more powerful computers,larger data sets and the ability to train deeper network technologies,the popularity and practicality of deep learning has grown tremendously,with many important issues in computer vision,speech recognition,image processing and natural language processing.There are outstanding performances on the top.In this paper,deep learning is applied to image fusion work,and three image fusion algorithms are proposed.Firstly,the multi-focus image fusion task is regarded as the classification problem of the focused image block and the defocused image block.Based on the original Alex Net network model,the new network model is used to block and classify the multi-focus images.Secondly,the multi-spectral image(MS)is spatially enhanced by the super-resolution image reconstruction network.The image components of the corresponding channel are subjected to non-downsampling shearlet(NSST)transform to obtain low frequency sub-bands and several high frequency direction sub-bands,respectively.And the inverse NSST transform recons-tructs the obtained fusion images.Finally,an visible infrared image fusion algorithm based on convolutional neural network is presented.Firstly,the scaled-aware edge protection filter composed of the guided filter and the Gaussian filter is used to perform multi-scale decomposition on the input source image,and the base layer is fused by the weighted average fusion rule of the pixel intensity distribution,and the detail layer uses the convolutional neural network to space Details are extracted and blended.The experimental results show that compared with the traditional image fusion algorithm,the fusion images obtained by using the three image fusion algorithms mentioned in this paper have better performance in both subjective visual and objective evaluation indicators,indicating that the deep learning method is in the image.The effectiveness of processing.