SUPER PIXEL BASED VIRTUAL TEXTURE MAPPING OF IMAGE SYNTHESIS

: Image synthesis is the process of generating images of a 2D model or 3D model by means of computer programs. This involves several different subjects like global illumination, local illumination, rendering and visual perception. We concentrate on the subject of rendering which can be projection of environment into an image. Sometimes, it may be different to that what we really want to capture, i.e., rendering artifacts may occur. In existing method, object removal of an image takes place by the exemplar-based inpainting. While in proposed method, occlusion culling and shadow removal of an image to reduce the rendering artifacts. But, the regions are filled with a texture according to those neighbourhood pixels for the consideration of super pixel. Finally, these methods generate more natural composite images especially for a proposed method the image of selected region affecting the texture, the illumination and the colour of objects lying in the region.


I. INTRODUCTION
In computer vision [1], super pixels also known as sets of pixels which can be of partitioning a digital image into multiple segments. To rearrange or progress the representational for a image into something that is additional serious and also simpler to examine. Image segmentation will be transformed from claiming appointing a name to each pixel clinched alongside a image such-such pixels with the same name offer specific qualities. The result is a set of segments that collectively cover the entire image, or a set of contours extracted from the image. Each of the pixels in a region is similar with respect to some characteristics or computed property, such as colour, intensity, or texture. Contiguous regions are fundamentally separate with admiration to the same qualities.
Texture is a vague word and in the ambience of texture synthesis [2] may have:  In common speech, the word 'texture' is used as a synonym for 'surface structure' which has different properties as coarseness, contrast, directionality, linelikeness and roughness.  In 3D computer graphics [3], a texture is a digital image applied to the surface by texture mapping to give the model a more realistic appearance.
 In image processing, every digital image composed of repeated elements is called a 'texture'. Texture mapping [4] is a method for defining high incidence specify, surface texture, or colour in sequence on a computer-generated graphic or 3D model. It is a method that plainly wrapped and mapped pixels from a texture to a 3D surface.
Image synthesis is the evolution of creating images from scenes is called rendering. This processing prepares the atmosphere for rendering [5] through such possessions as the propagation of energy throughout the environment or optimizations to demonstrate. An overview of general process as follows:  The object model is determined starting with those displaying procedure that we analyzed.  The Illumination model describes how light (and other energy components) is defined and managed.  Texture refers to the group of techniques (texture mapping, bump mapping, reflection mapping etc.) used to add effects to the surface of objects.  Shading refers to variations in plane lighting that result from the application of an illumination model.  Image formation is the process of bringing together the other components to form a synthetic image, usually called rendering or image synthesis. The applications of image synthesis are Computer -Aided Design (CAD), Entertainment, Simulation, Computer Art, Augmented Reality (Image analysis + Image Synthesis).
The DIBR (Depth Image Based rendering) method has limitations due to inherent artifacts in the warped image [5]. As a result, perception of the depth and the desired experience are affected due to the reduced visual quality of virtual view. Rendering artifacts are severe in the extrapolated view, since there is no additional information to handle, whereas in the intermediate view the artifacts are partially filled with the information from the other warped view. Possible artifacts in the warped images are classified as shown in Fig 2. The artifacts in the warped image are ghosting and holes (i.e. uncovered areas).

Fig 2: DIBR artifacts classification
Ghosting artifacts are mixtures of colors at the edges in the original image which are projected into the neighbouring objects in the warped image. The cause of the ghosting artifact is the depth and texture misalignment and mixing of neighbouring colours at the depth discontinuities.
Holes are undefined pixels in the rendered images. They appear due to uncovered (not captured) regions because they were occluded by foreground in the original view. They are classified into three types, namely cracks, disocclusions and out-of field areas.
Cracks are one to two pixel-wide empty regions in the warped image. Cracks appear in the virtual view due to assigning warped pixel coordinate positions to the nearest integer coordinates.
Translucent cracks appear for the same reason as the cracks, but these areas possess background information as consequence of occlusions in virtual view.
Disocclusions [6] occur at depth discontinuities near the object borders. These are the result of regions being revealed in the virtual view as they have been occluded by the foreground in the original view. Therefore occlusions in the original view translate into disocclusions and appear as holes in the virtual view as the result of the warping process. When there are more than two depth layers, we define foreground and background as relative terms. Occlusion areas between a relative foreground-background pair are referred to as overlaid occlusions. It is worth noting that the size of the disocclusions depends on the size of the baseline and the scene depth.
Translucent-disocclusions differ from common disocclusions by exposing texture information that is present behind the relative background. Translucent disocclusions occur only when the depth has three or more layers, and where there are occlusions between layers. The magnitude of the disturbance created by translucent disocclusion artifacts depends on the placement of occlusions in the scene and the occlusion area. However, translucent disocclusion artifacts need to be addressed since they can deteriorate the perceived visual quality.
Out-of-field areas are also holes, which occur at image borders. They occur due to the limited field of view in the original view.
Unnatural contours are the pixilation of "new" edges between background and foreground in the warped image due to the colors of the pixels at the edge not being blended.
To reduce the DIBR artifacts a number of state-of-the-art solutions have been introduced. However, disocclusion handling still remains a challenging problem.

II. RELATED WORK
Image -based modelling [7] and rendering techniques have recently arriving with much attention as powerful substitute to usual geometry-based techniques for image synthesis. Instead of geometric primitives, a collection of sample images are worn to render novel views. Examples comprise the well known panoramas, light fields, and variants, concentric mosaics, etc. The reconstruction model (i.e., rendering) is treated as a multidimensional sampling problem, where new views are generated from compactly sampled images and depth maps instead of building precise 3D models of the scenes.
For instructive purposes, imagebased representations can be classified according to the geometry information used into 3 main categories: 1) Representations with no geometry 2) Representations with implicit geometry 3) Representations with explicit geometry 2-D panoramas, 3-D concentric mosaics, 5-D McMillan and Bishop's plenoptic modelling, 4-D ray-space representation, and light fields/ lumigraph belong to the first category, while layered-based or object-based representations using depth maps fall into the third. Conventional 3-D computer graphics models and other more sophisticated representations belong to the last category. At the other intense, light field or lumigraph rendering relies on opaque sampling (by capturing more image/videos) with no or very petite geometry information for rendering without recuperating the exact 3-D models. An important lead of the latter is its superior image quality, compared with 3-D model building for complex real world scenes and the other is that it requires noticeably less computational resources for rendering not considering of the scene complexity, because most of the quantities involved are precomputed or recorded. This has fascinated to considerable attention in the computer graphics community, recently in developing quick and efficient rendering algorithms for real-time relighting and soft shadow production.
Since capturing a 3-D model in real time is still a very difficult problem, light-field or lumigraph based dynamic IBR representations with slight amount of geometry information have received extensive attention in immersive TV (also called 3-D or multiview TV) applications. In particular, tremendous rendering quality has been demonstrated using the pop-up light field, the object-based approach and the layered based rendering approach. Later sections will be faithful to these representations, which use fairly accurate/ incomplete geometry in the appearance of depth maps. Segmentation, mapping, and depth estimation techniques are essential to these approaches.
On the other dispense, since 3-D models of the objects and scenes are not available, user communication is limited to the change of viewpoints and sometimes restricted amount of relighting. In contrast, more user interaction such as real-time relighting and soft-shadow computation has been found to be practicable using IBR concepts [8] and the associated 3-D models using precomputed shadow fields. This opens up a new chance for very fast interactive visualization /graphic systems with low complexity. If approximate geometry of objects in a scene can be recovered, then interactive editing and relighting of real scenes are in principle practicable. This has important applications in computer games, scientific visualization, and relighting of IBR objects in future generations of IBR systems.

III. OVERVIEW
Texture is an important visual system and it is used to distinguish one object from other objects clearly. Especially in color images which provide extra information that can be used on its own or combined with texture information to segment images. Texture indexing is an act of classifying and in order to make items easy to retrieve.
On exposed area of an image do not have corresponding information about the reference view, their texture attributes are uncertain. Those uncertain values can be filled based on the neighbour pixels by linear interpolation [9], extrapolation [10] or inpainting [11].
Inpainting is one of the arts for restoring the lost or damaged parts of an image and also get rid of the unnecessary objects depending on the background of that image. This should be done in an untraceable way. Previously, in past days, the artists used to paint by their own manually of the damaged paints. As the technology is recuperating day to day, this image inpainting processes the inpainting automatically.
The subject of rendering which can be projection of environment into an image involves different subjects as we discussed here are object removal, occlusion culling and shadow removal.

A. Object Removal
The idea behind this part is to fill in larger areas of an image by repeating the textures that surrounds the target region with some level of stochasticity. This is done in an onion-peel like approach starting from the outer layers and moving inwards towards the centre of the target region.
Exemplar-based technique which it takes a few samples and considers the high-quality matching texture patches from texture synthesis. It gives higher results for putting off of objects from the picture. The desired residences as of filling the location which has (i) correct propagation of linear systems, (ii) robustness to modifications in form of the goal region, (iii) balanced simultaneous structure and texture propagation all in single efficient set of rules as criminisi.
First, given an input image, the user selects a target region, Ω, to be removed and filled. The source region, Φ, may be defined as the entire image minus the target region (Φ = I − Ω), as a dilated band around the target region, or it may be manually specified by the user. Size of the template window Ψ must be specified as of 9×9 pixels, but in practice require the user to set it to be slightly larger than the largest distinguishable texture element, or "Texel", in the source region.
In our criminisi algorithm [12], each pixel maintains a colour value (or "empty", if the pixel is unfilled) and a confidence value, which reflects our confidence inside the pixel price, and which is frozen once a pixel has been filled. Throughout the course of the algorithm, patches alongside the fill the front are also given a transient precedence cost, which determines the order in which they may be filled. Then, our set of rules iterates the subsequent 3 steps till all pixels have been filled:

Computing patch priorities:
The synthesis task through a best-first filling strategy that depends entirely on the priority values that are assigned to each patch on the fill front. The priority computation is based toward those patches which: (i) are on the continuation of strong edges and (ii) are surrounded by high-confidence pixels.
Given a patch Ψ p centred at the point p for some p∈δΩ, we define its priority P(p) as the product of two terms: P(p) = C(p)D(p) (1) We call C(p) the confidence term and D(p) the data term, and they are defined as follows: Where |Ψ p | is the area of Ψ p , α is a normalization factor (e.g., α = 255 for a typical grey-level image), n p is a unit vector orthogonal to the front δΩ in the point p and ⊥ denotes the orthogonal operator. The priority P(p) is computed for every border patch, with distinct patches for each pixel on the boundary of the target region.
During initialization, the function C(p) is set to C(p) = 0 ∀p∈Ω, and C(p) = 1 ∀ p∈I−Ω. The confidence term C(p) may be thought of as a measure of the amount of reliable information surrounding the pixel p. The intention is to fill first those patches which have more of their pixels already filled, with additional preference given to pixels that were filled early on (or that were never part of the target region).
At a coarse level, the term C(p) of (1) approximately enforces the desirable concentric fill order. As filling proceeds, pixels in the outer layers of the target region will tend to be characterized by greater confidence values, and therefore be filled earlier; pixels in the centre of the target region will have lesser confidence values.
The data term D(p) is a function of the strength of isophotes hitting the front δΩ at each iteration. This term boosts the priority of a patch that an isophote "flows" into. This factor is of fundamental importance in our algorithm because it encourages linear structures to be synthesized first, and, therefore propagated securely into the target region 2. Propagating texture and structure information: Once all priorities on the fill front have been computed, the patch Ψˆp with highest priority is found. We then fill it with data extracted from the source region Φ.
On the contrary, we propagate image texture by direct sampling of the source region. Similar to, we search in the source region for that patch which is most similar to Ψˆp. formally, Ψˆq = arg min Ψq∈Φ d(Ψˆp,Ψ q ) (4) Where the distance d(Ψ a ,Ψ b ) between two generic patches Ψ a and Ψ b is simply defined as the sum of squared differences (SSD) of the already filled pixels in the two patches. Pixel colors are represented in the CIE Lab colour space because of its property of perceptual uniformity. Having found the source exemplar Ψˆq , the value of each pixel-to-be-filled, p|p ∈ Ψˆp∩Ω, is copied from its corresponding position inside Ψˆq.

Updating confidence values:
After the patch Ψˆp has been filled with new pixel values, the confidence C(p) is updated in the area delimited by Ψˆp as follows: This simple update rule allows us to measure the relative confidence of patches on the fill front, without imagespecific parameters. As filling proceeds, confidence values decay, indicating that we are less sure of the color values of pixels near the centre of the target region.
Finally, pixels are classified as belonging to the target region Ω, the source region Φ or the remainder of the image by assigning different values to their alpha component. The image alpha channel is, therefore, updated (locally) at each iteration of the filling algorithm.

B. Occlusion culling
Virtual reality is a new and rapidly expanding area seeking to immerse the user in completely new surroundings. Total immersion is currently hindered in prerecorded environments by the user's inability to move inside the scene. Additional views can be synthesized which causes disocclusion holes -visual artifacts due to spatial regions that were not visible in the reference view. These holes can be filled with traditional inpainting (using known portions of the image to fill the unknown portions), but the additional depth data can be leveraged to produce higher quality, more accurate results.
The criminisi method and the proposed modified daribo method which simply added another term to the two main equations as: ( ) = ( ) ( ) ( ). (6) L(p) is the level regularity term, or the inverse variance of the depth patch , described by: Where | | is the area of and is the mean of . The level regularity term gives higher priority to patch overlaying at the same depth level, which naturally favours background pixels in the case of holes caused by disocclusions.
The best exemplar calculation went from minimizing the SSD of the known pixels of the target patch and patches of the known portion of the image to Their algorithm fills the depth map prior to inpainting the image. This is done for two reasons: first, depth maps are usually constant or slowly changing, making them easier to fill, and second, the greater quantity of depth data assists in the image inpainting process. Ružić et al. [13] found that filling in entire rows of the depth map holes reasonably interpolates the depth data. Their algorithm fills the rows by extracting a suitable value from a local window located on the background side of the disocclusion hole.
During inpainting, all depth pixels are considered, and the top L best exemplars are averaged when filling in the target patch. Averaging the top exemplars adds a slight blur to the details added to the image, but it also helps prevent the algorithm from creating and propagating false structures in the image. The final change made is to limit the search area while the algorithm is checking for the best exemplar.
The method outperforms the depth-less Criminisi algorithm, where a premade mask was applied as the target area. Premade mask is part of the foreground, the background is automatically targeted first for inpainting, resulting in perfect reconstruction of original image.

C. Shadow removal
Shadows are ubiquitous in image and video data, and their removal is of concentration in both computer vision and graphics. Although shadows can be useful cues, e.g. shape from shading, they can also involve the performance of algorithms (e.g. in segmentation and tracking). Their removal and editing is also often the pain-staking task of graphical artists. An unbeaten shadow removal method should seamlessly relight the shadow area while keeping the lit area unchanged. The umbra is the darkest division of the shadow whilst the penumbra is the transitional shadow boundary with a non-linear intensity change between the umbra and lit area. The textures in shadowed surface generally become weaker that contrast artifacts can appear in shadow areas due to image post-processing.
The important contribution of this work is the first proven and multi-scene class ground truth for shadow removal algorithms. This data set containing 200 images eliminates in consistencies between shadow and shadowfree images and provides a range of various shadow types such as soft, textured, colour and broken shadow. Using these facts, the most thorough comparison of state-of-the-art in user-aided shadow removal technique is performed.
A shadow image I c may be taken into consideration as a Hadamard product [14] of a shadow scale layer S c and a shadow-free image I c as shown in Eq (9) where c is a RGB channel. The scales of the lit area are 1 and other areas scales are between 0 and 1. I c = I c * S c (9) For interactive shadow removal especially consists of 4 steps as Pre-processing, Penumbra Unwrapping, Relighting and Color Correction. Each step may be explained briefly with technical information for each of its components. In this we denote 6 undetermined parameters as h1,h2,...,h6.

1) Pre-processing
An initial shadow mask (Fig. 5(b)) is detected using a KNN classifier trained from data from two rough user inputs (e.g. Fig.5 (a)) are sample lit and shadow pixels. A fusion image, which magnifies illumination discontinuities around shadow boundaries, is generated by fusing channels of YCbCr color space and suppressing texture (Fig. 5(c)). Highlighted pixels RGB intensities in the Log domain are supplied as training features and used to construct a KNN classifier (K = 3). Euclidean distance is used as the distance measure and the majority rule with nearest point tie-break as the classification measure. Spatial filtering with a Gaussian kernel (size = h1, ) is applied to the obtained image of posterior probability and binarize the filtered image using a threshold of 0.5 (e.g. Fig. 5(b)).
To assist unwrapping of the penumbra, an image is derived that magnifies illumination discontinuities around the shadow boundaryalso assisting penumbra locationwhich is called the fusion image (e.g. Fig. 5(c)). There are 2 steps in this process:

i) Magnification of illumination discontinuity
We derive an initial fusion image F that maximises the contrast between shadow and lit areas by linearly fusing the three channels (C l ) of YCbCr space as eq.(10) as follows: Where a l is the fusing factor of C l (positive). The best fusing factors are derived by minimising the following objective function E b : where a is the vector of fusing factors and F S and F L are the two sets of shadow and lit pixels marked by user scribbles. In this paper, σ and µ are defined as functions that respectively compute the standard derivation and mean of a set of values. The first term ensures larger distinction between pixels of lit and shadow regions and the second term ensures smaller variation for pixels of the same lit or shadow regions.
ii) Suppression of Texture The noise due to image texture is reduced by applying a median filter with an h2-by-h2 neighbourhood to F. YCbCr color space offers perceptually meaningful information. Empirically, illumination information appears dominantly in one of its channels. The illumination information in RGB channels are usually affected by texture noise. 2) Penumbra Unwrapping Based on the detection shadow mask and fusion image, pixel intensities of sampling lines are sampled perpendicular to the shadow boundary (Fig. 5(d)). Noisy samples are removed and remaining columns stored as the initial penumbra strip (Fig. 5(e)). The initial columns illumination changes are also aligned (Fig. 5(f)) by a finescale alignment.

1) Pre-processing 2) Penumbra Unwrapping 3) Relighting 4) Color Correction
The start and end points are initially set as the boundary point (x b ,y b ) and the direction vector ∆v as the normalized gradient vector of (x b ,y b ). Instead of processing unaligned and unselected samples individually, we transform these samples into unified columns of the initial penumbra strip to enable fast batch processing. And, only the good samples are kept for shadow scale estimation.
To avoid outliers, e.g. sampling lines at occlusion boundaries, invalid samples are filtered based on an assumption of similar shadow scales. A scale vector Y c = T l −T s is first computed where T l and T s are the average Log-domain RGB intensities of the lit and shadow halves of a sampling line. Y c is then converted to spherical coordinates and considered as feature vector Y s . DBSCAN clustering [15] (radius: h3) is applied to Y s for all samples, and samples that belong to the largest cluster are stored as valid ones with valid illumination. For finer scale estimation, valid clusters are further divided into a few sub-groups using mean-shift [16] (band width: h4) and the samples of invalid subgroups, whose total members are less than 10% of the largest sub-group's, are discarded. Fig. 5(d) shows an example of the above outlier detection.
The parameters of this fine-scale alignment for each column are estimated by minimizing the following energy function Ea: E a = MSE(L n −L a ) (12) where L n = Γ(A s ,A k ,L o ), whose A s and A k are the stretching shift and the centre shift in the fine-scale alignment respectively, L o is the scales of original column, L a is the reference of alignment which is the average scale values of all valid columns (i.e. column-wise mean of the penumbra unwrap), L n is the aligned unwrap, Γ is a function that aligns L o according to the estimated alignment parameters, MSE is a function that computes mean squared error.

3) Relighting
From the penumbra strip, multi-scale shadow scale estimation is applied to quickly and robustly estimate the illumination change along sampling lines and derive the sparse scales for all sampled sites (Fig. 5(g)) which are propagated to form a dense scale field (Fig. 5(h)). Shadows are removed by inverse scaling using this nonuniform field (Fig. 5(i)).
A pyramid (e.g. Fig. 5(b)) of horizontally filtered penumbra unwraps using 5 averaging kernels in different sizes are computed so that texture noise can be cancelled. The sizes of averaging kernels are specified as 1-by-2˜n where ˜n∈{2,3,4,5,6}. The filtered intensities of the pyramid are then converted to shadow scales.
The optimum shadow scales for each column are selected from different layers of the pyramid. Column intensity with higher localness (i.e. filtered by a smaller kernel) and lower roughness are preferred. However, higher localness leads to higher roughness, so an optimum solution should balance these two properties. The roughness of intensity change E s (˜c, ˜n) is measured as follows: E s (˜c, ˜n) =∫(∂ 2 u(˜r,˜c,˜n)/∂ ˜r 2 ) 2 d˜r (13) Where u is the penumbra unwrap, ˜ n is the layer index of pyramid, ˜ c and ˜ r are the column and row coordinates of the penumbra unwrap respectively. The optimum scales for each column are selected using a threshold of roughness T s which is computed as the mean of all values in E s.

4) Color Correction
Post-processing effects may cause inconsistent tone and contrast in shadow removed areas compared with the lit areas. Without introducing additional artifacts, a multiscale color correction is proposed to remove these inconsistencies (Fig. 5(j)). The multi-scale color correction aligns the color variance at different scales from coarse to fine. On each single scale, the initial input image (c1) exhibits inconsistency of local variance between lit and shadow areas. The higher-frequency variation (c3) of shadow and lit areas are aligned in (c4). The corrected output (c5) can be obtained by adding (c4) to (c2).
Statistics are collected from the lit side pixels P l and the umbra side pixels P u both near the penumbra as the reference and source of color correction respectively.
The algorithm for alignment is described in Algorithm2. Where s is a scale, β is the maximum image dimensional of I ra , bfilter is an operation that bilaterally filters [17] the input image (first parameter) using a standard deviation of the space (second parameter) and a range Gaussian (third parameter), I h is an image of intensity variation, where c is the channel index, m is a function which computes the median absolute deviation.
Finally, to smooth the color correction result, alpha blending is applied in RGB color space according to the shadow scale as shown in Eq. (14).

5) Error Measurement
Our objective function E p minimizes the sum of all error measurements as the follows: Where e k is the k th error measurement and w k is the weight for e k . We assume that the weights for all error measurements are the same (i.e. equally important), e.g., w k = 1.

D. Evaluation
In this section, we first describe experiments that highlight our algorithms behaviour given variable user inputs. The quality of our new ground truth versus existing state-of-the-art ground truth is then quantitatively evaluated. Finally, our algorithm is evaluated versus other state-of-theart shadow removal methods based on our new dataset.
Given user-supplied single pairs of strokes of lit and shadow pixel samples, our shadow detection generates stable results in different conditions (e.g. Fig. 7(a)). In some cases, e.g. where the surface color is very shadow-like, the detection results can be improved by supplying more than one pair of strokes (e.g. Fig. 7(b)).
Fewer gray pixels indicate higher stability, i.e. the image should only show black (0% probability) and white (100% probability) pixels when it is absolutely stable. The bottom row shows examples highlighting how additional strokes can improve the detection result (binary mask).

1) Ground Truth Quality
Ideal pairs of ground truth images should have a minimum intensity difference in the common lit areawhich will also indicate whether registration is poor (due to camera shake or scene movementwhich should be rejected). This is utilized to assess the quality of ground truth candidates. The error image ∆I = I s −I g and the ratio image I r = Φ(I s ) ø Φ(I g ) are first computed, where I s and I g are the original shadow image and its shadow-free ground truth image (which differs from the processed shadow-free outputs I f in Eq. 14) respectively, ø is element-wise division and Φ is a function that converts RGB image to gray-scale image. The set of pixels P r of I r that satisfies I r (P r ) ≥1 are regarded as lit pixels. Due to some unavoidable minor global illumination changes and the inaccuracy in camera exposure control, the lit intensities in the shadow image can be higher than those in the shadow-free ground truth image. I r can therefore be greater than 1. The ground truth error Q d is computed as follows: Q d = µ(|∆I(P r )|) + σ(∆I(P r )) (16) Ground truth pairs in our data set with Q d > 0.05 are removed. To test robustness, the standard deviation for each measurement is also computed. Unlike previous uncategorized test, our removal test is based on our data set of 214 cases, which contains challenging soft, broken and color shadows and shadows cast on strong textured surfaces. Each case is rated according to 4 attributes, which are texture, brokenness, colourfulness and softness, in 3 perceptual degrees from weak to strong which were aggregated by 5 users.

2) Quantitative evaluation of shadow removal
The quality of shadow removal is measured by directly using the per-pixel error between the shadow removal result and shadow-free ground truth. However, a shadow in a smaller size or a lighter shadow can result in a smaller initial error between the original shadow image and its shadowfree ground truth. It is thus unfair to judge that a method is better only because the error between the shadow removed image and its shadow-free ground truth is smaller. In our work, we cancel out the affects of the size and darkness of the shadow. We therefore compute the error ratio E r as our quality measurement: Where E n is the error between the ground truth (no shadow) and shadow removal result, and E o is the error between the ground truth (no shadow) and the original shadow image. This normalized measure better reflects removal improvements towards the ground truth independent of original shadow intensity and size. We assess E n and E o using Root-Mean-Square-Error (RMSE) of RGB intensity. According to Eq. 17, E o for all pixels is lower than E o for the shadow pixels only because the intensity errors of the lit pixels are close to 0 and E o measures the RMSE. In addition, the new RMSE E n for both all pixels and shadow pixels only are very close after shadow removal, the error ratio E r for the shadow pixels only is therefore generally lower.
Our method requires users to supply reasonable inputs. We have not considered its tolerance for very careless user inputs (e.g. mistakenly marking many shadow pixels as lit samples). Besides, insufficient user inputs may result in poor shadow detection. Since a sufficiently trained shadow classifier may be robust to this issue, another future work could be improving shadow detection by combining user inputs with the shadow masks generated from an automatic shadow classifier.

IV. CONCLUSION
We have presented a texture mapping based on uncertain values of neighbour pixels by inpainting concentrated on rendering artifacts of occlusions and a number of state-ofthe-art solution in case of user-aided shadow removal method is performed using matlab. Occlusion culling method utilizes aspects of several disocclusion inpainting methods to achieve desirable and visually believable results. It consistently outperforms a similar function which does not consider depth information. Shadow removal of our technique balances the complexity of user input with sturdy shadow removal performance. Our quantitatively-verified ground truth data set overcomes issues of mismatched illumination and registration in present data sets. Except the opportunities for improving shadow removal quality for the categorized shadows in our dataset, the detection and elimination for fantastically-complex shadows, such as overlapping shadows precipitated multiple mild resources with one of kind mild colors, and shadows as a result of transparent gadgets with complex internal shape and color, continues to be an open trouble for the community.