Three Kinds Of Image Fusion Algorithms Based On Convolutional Neural Network

As one of the most important branches of machine learning,deep learning has developed rapidly in recent years based on the mass data of social digitization.Deep learning is widely used in image processing.Deep learning theory can be used to process at different levels of image fusion,which provides a new tool for image fusion processing.Due to restricted by technical factors,from the part of the sensor image information is not complete,different sensors will receive different image information,the image fusion is the process of the information collection to a single image,combined with the deep learning theory can better to tackle the problem of image fusion using convolution neural network training,integration can generate a single image collection more advantages of two or more images,It expresses more complete information than the traditional fusion algorithm,and presents more natural and clear information to people in machine vision.In this paper,image fusion problem is researched by means of deep learning theory,discrete wavelet decomposition and operator theory.Three image fusion algorithms are proposed.Firstly,the image fusion algorithm based on Crop-VGG convolutional neural network is proposed to deal with the problem for multi-focus image fusion.Through the operation of the input data preprocessing,image input modules with four-characteristic value are designed based on wavelet transform and Sobel operator.Thus,the precision of the whole training network is improved.By adjusting the details of the network,the convergence speed of the Crop-VGG model is accelerated.The clear modules and fuzzy modules of the left-focused images and right-focused image are blurred respectively in the high frequency detail parts.Then the weighted summation fusion strategy is applied.The clear-everywhere focused images are obtained.Secondly,according to the different characteristics of multispectral images and panchromatic images,discrete wavelet decomposition is impremented on them,respectively.This paper a combination of wavelet decomposition image super-resoluti--on reconstruction network VDSR detail component operating conditions is proposed,and adjust the network parameters,through the experiment training finally through IHS low frequency weighted processing,with high frequency part of discrete wavelet inverse transformation,get the fusion image.Finally,an infrared image fusion algorithm combining discrete wavelet and GAN network is proposed,which decomposes the high and low frequency images of the original infrared light and visible light.Through the adaptive weighting coefficient fusion on the base layer and GAN network is used to generate images with high-frequency texture information and location information on the detail layer.The base layer and the detail layer are reconstructed to get a complete fusion image.Experiments show that compared with the traditional image fusion algorithm,the three image fusion algorithms proposed in this paper based on deep learning theory and wavelet method have better effect in both subjective and objective evaluation.