Blind Separation For Permutation Image Based On Deep Learning Models

Blind source separation,which is also known as blind separation,can separate the input source signals from the observed mixed output signals without too much prior information of source signal and channel.It is also an advanced technology which widely used in the processing of images,voice signals,medical signals and so on.Blind separation of permuted alias images is a special case of blind source separation,alias mode of which permuting alias mode rather than traditional superposition alias mode.Therefore,the blind separation algorithms for the images aliased in traditional overlaying way are no longer suitable for the images in permuting way.So,new method and theory are necessary to study blind separation of this permuted image.As the prior information is all unknown,such as location,size,number and type of the permuted region,it is of great challenge to separate the image in the permuted region based simply on the information of the image itself.However,even the prior knowledge are unknown,the unsupervised neural network can automatic discover the internal feature of the data as well as the difference between them.Hence,in this paper,unsupervised Sparse Auto-Encoders is used to separate permuted alias image with noise and Restricted Boltzmann Machine to separate with blur.The main contents of this paper are summarized as follows:The noises are different between the permuting and permuted regions and have no sparse representation.We choose Sparse Auto-Encoders to separate permuted image with noise.First,the image is divided into small pieces.We turn the pieces into operable dataset.Then,we construct the sparse automatic encoder network structure by forward propagation,and use the input dataset to train the network to obtain the decoded image.Finally,according to the difference image which is obtained by subtracting the decoded and input image,we use maximum interclass variance method to find appropriate threshold and obtain the permuted image with noise through threshold operation.The experimental results show that the proposed algorithm can effectively separate the permuted images with different positions,sizes and numbers when the noise type and variance size are unknown.For the permuted alias blur image,blind separation for permuted alias image is proposed based on Restricted Boltzmann Machine.First,we convert this kind image into operable dataset.Then,we construct and train the limited Boltzmann model to get the probability matrix.Finally,the input data set is reconstructed with the trained network,and the blur mixed aliasing image is separated according to the characteristic difference between the reconstructed data set and the input data set.The experimental results show that the proposed algorithm can effectively separate the permuted image even though the location,size,blur type and ambiguity are unknown.