Single Image Super Resolution Image Reconstruction Involving Examplar Based Inpainting

Although tremendous advancement happen in image processing domain, still “filling the missing areas” is area of concern in it. Though lot of progress has been made in the past years, still lot of work should be done. A novel algorithm is presented for examplar-based inpainting. In the proposed algorithm initially inpainting is applied on the coarse version of the input image, latter hierarchical based super resolution algorithm is used to find the information on the missing areas. The unique thing of the proposed method is easier to inpaint low resolution than its counter part. To make inpainting image less sensitive to the parameter, it has inpainted several times by different configurations. Results are combined using the loppy belief propagation and by using the super resolution details are recovered. The proposed algorithm results are compared with the different existing methods; results shown performance and efficiency are more accurate and reliable. Introduction In image processing “Filling the Missing Areas (holes)” is a problem in many image processing applications [1]. Although lot of research done still it’s an area of concern in many image processing applications. Image inpainting is the procedure of reconstructing lost or deteriorated parts of images. Existing methods are broadly classified into two sections a) Diffusion based approach b) Examplar based approach. These two existing methods are inspired from the texture synthesis techniques [2].Diffusion based approach generates the isophotes via diffusion based on variational structure or variational method [3], the main drawback of diffusion based approach is have a tendency to introduce some blur when the filling the missing area is very large. Latter method of approach is Examplar based approach which is quite simple and innovative, in this method copy the best sample from known image neighborhood. Initially examplar method approach is implemented on object removal as chronicled in [4], searching the alike patches is done by using the priori rough estimate method of the inpainted image values utilizing the multi-scale approach. The two varieties of methods (diffusion based approach and Examplar based approach) are then combined, for example by utilizing the structure tensor to calculate the priority of the patches to be filled in [5]. Latter the examplar approach is combined with the super resolution algorithm as shown in [6], it’s a two steps approach, firstly rough (coarse) version of the input image is inpainted then in second step originating the high clarity image from the inpainted image. Although lot of advancement done in the past decade on examplar based inpainting still lot problems to be addressed in all the main area of concern is patch size and filling the holes related to settings configuration. This problem is here addressed by several input inpainting versions to yield the final inpainting image after combining the all input inpainting versions. Note that Inpainting is applied on the rough (coarse) version of the input image when the filling area (hole) is very large which reduces the impact of computational complexity and robust behavior against noise entities. In this type of scenario final full resolution image is retrieved from the super resolution algorithm [6]. PROPOSED ALGORITHM The proposed inpainting algorithm presents the novel inpainting algorithm and also the process of combining the different inpainting images. NOVEL INPAINTING METHOD BASED ON EXAMPLAR APPROACH As described in the literature, filling the missing information or filling order computation and texture synthesis are the two classical steps .Based on these classical steps the INTERNATIONAL JOURNAL OF PROFESSIONAL ENGINEERING STUDIES Volume II/Issue 3/JUNE 2014 IJPRES proposed examplar based approach is presented. These two steps are discussed in latter section. A.PATCH PRIORTY: This section describes the first classical step i.e. filling order computation. The patch priority mainly focuses on two ideas; firstly differentiate the structures from the coarse version latter knowing the priority is salient step if the priority is high it indicates the presence of structure. By using the data term and the confidence term the priority of patch can be centered on Px. In order to know the data term in a detailed way tensor based [7] and Sparsity-based [8] have been used. THE PROPOSED ALGORITHM FRAMEWORK The priority term which is based on tensor approach is defined by a Di zenzo matrix or structure tensor is as follows = ∑ ‘J’ in above equation represents the sum of scalar structure tensors of the image Ii(R G B). The smoothing of the tensor is done without cancelling effects: Jσ=J*Gσ, where Gσ=1/2πσ exp (-(x+y/2σ)), with standard deviation σ. The main advantage of the structure tensor is that structure Eigen values is deduced from coherence indicator. Based on the accuracy that we are getting from the Eigen values we locate the anisotropic of the local region can be evaluated. The structure tensor Jσ is computed by using the structure tensor, here the Eigen vectors V1 V2 represent oriented orthogonal basis and Eigen values represent the structure variation. Then Eigen vector V1 represents the highest fluctuations and V2 is the local orientation. The data term D (p ) =∝ +(1−∝)exp (( ) ) Where is a positive value i.e. =8 ∝=0.8; lies between 0 and 1 The sparsity based priority is another method recently proposed by the professor Xu at el in [8].In this template matching is performed between the current patch and neighboring patch of the known pixel. By using the non local means of approach similarity weight is computed between the each pair of the patch as shown below D(p ) = × |N (p )| |N(p )| Where Ns and N stands for the number of valid patches and the is high then prediction of candidates is low, if is low then the predication of candidates is high. Training samples Singleimage SR n Inpainting