Review of image defect detection and inpainting techniques and scope for improvements.
The occurrence of defects in an image is a general phenomenon, but their automatic detection and restoration towards originality has been a latent research area. Image inpainting involves two steps i.e. identification of defected pixels followed by restoration towards their original value. The restoration should be accurate enough so that it is non-detectable by an observer who is not familiar with the original images. Basically the defects on images may occur due to three main reasons that are by addition of layer of defecting material, by removal of image material from their surface and printing or drawing defects. In all the cases, original image pixel values are changed. The images usually have typical damages from ink spray, paints, dust, scratch, scribbling etc. E. Ardizzone et. al.  compiled a report and provided a standard taxonomy for different type of possible defects. Lots of literature is available in this domain. This paper reviews the available techniques of automatic defect detection and image restoration, highlights the basic approach involved and identifies the gray areas to focus for future research.
This paper is organized in a way that after an introduction of the problem domain in section 1, section 2 gives a review and summary of automatic defect detection approaches, Section 3 gives the study of available restoration techniques with summary and finally we concluded in section 4 with future scope for research.
2 AUTOMATIC DEFECT DETECTION APPROACHES
Defect detection means locating the defected pixels accurately. Lot of work has been done for detection of different types of defects. The automatic defect detection algorithms available so far are defect specific and require human involvement for initial information about nature of defect. This makes them specifically a subjective approach. Presently available automatic defect detection schemes are locating the defected pixels basically by modeling the defect, exploring image background information, analyzing the spreading patterns of defect causing material, and analyzing the statistics of image. Based on these observations, following section classifies the available approaches into four categories, reviews them and further concludes with summarization.
2.1 Model based approaches
These approaches are based on developing a suitable model for a particular type of defect. This is achieved by studying the defect profile carefully. The prepared model of defect is then compared with various input image regions for identification of defects. These approaches are applicable where the shapes of defect objects are of some standard type e.g. scratches, lines, spots, blotches etc. The scratches or lines are assumed to be long-thin vertical sections of bright or dark intensity w. r. t. image background and have width of 3 to 10 pixels. From image processing point of view, a vertical scratch may be defined as a few columns of image where actual intensity information is entirely or partially lost. The first spatial model for scratch was proposed by Kokaram . In this model, the luminance cross-sections of scratches was explored which consists of vertical mean of each column of image intensity and eliminates the space varying local mean. The black scratches correspond to minima where as the white ones to maxima of means. As an example an image with scratches and corresponding intensity distribution is shown in Figure 1. The white scratch cross-sections profile may be modeled as in equation (1)
G (i, j) = I (i, j) + Ln (i, j) + e (i, j) (1)
where G(i, j) represents the degraded image intensity at position 'i, j', I(i, j) is the intensity of undamaged image, e(i, j) is an additive Gaussian noise, and Ln (i, j) is the term relative to scratch.
The Ln (i, j) may be modeled by a sinusoid represented by equation (2) below-
Ln = b * k [absolute value of (1 - n)] cos ([omega][absolute value of (1 - n)]t) (2)
Where Ln is the minimum brightness of particular scratch i.e. the peak of the sinusoid and b is slope and k is the decay coefficient depends on scratch width. 'n' represents the width of scratch e.g. for a scratch with width of 5 pixel, n ranges from 0 to 5. Before making comparisons with the finalized defect model for detection, the images are processed with Gaussian filter to smooth its texture and median filter to remove noise. After differencing both filter outputs, the difference image with Hough transform generates an input image for further defect detection processing. In this method, it is assumed that the lines and scratches are additive, spread over the entire vertical length of the image and also not curved. Later V. Bruni et al.  generalized Kokaram's approach by including curved scratches. L. Joyeux et al.  also presented their work for scratch detection and removal. Hongying Zhang et al , proposed algorithm for scratch detection. This technique is based on characteristics of uniform intensity and structural information of scratches. After applying certain transforms (top hat and low hat) followed by morphological processing, scratch regions were highlighted. B. Grosjean et al.  developed a spot defect detection model inspired from human vision system. They extended acontario framework for spot detection by developing a statistical model to predict the ability for detection of spot on a textured background. The proposed model is first validated theoretically with the available image texture and then generates a relationship between size of spot and its probable contrast w. r. t. background. Finally a threshold is estimated for spot detection. Y H Zhang et al.  introduced an intelligent color textured and classification model using genetic algorithms and the Elman neural network for fabric and stitching defect detection and classification.
X Si et al.  proposed a new regional growing pulse coupled neural networks (PCNN) model according to the features of fabric image and successfully applied it to fabric defect detection.
Xiaosong et al.  used Hidden Markov model (HMM) based approach, targeting dirt and blotches defects in digitized archive films. This approach is oriented towards defect detection by combining temporal and spatial information across a number of frames. A HMM is trained for usual observation sequences and then applied within a structure to detect defected pixels. Since the occurrence of defects is somewhat haphazard in nature and can be considered as a stochastic pixel change event, so rather than attempting to model the defects, they build a model of how distinctive the sequence of pixels is and any pixel sequence conflicting from this standard model can be marked as a possible defect. This approach results in large number of false detections because of errors in definition of defective pixels. But to differentiate defects with similar image objects and cases of thick scratches are the areas to further explore. These false detections results in loss of actual information during restoration.
2.2 Image textural/structural based approach
Any image is composed of structure image (edges and smooth areas) and texture image. Image texture is described by number and type of its primitives and their spatial association or arrangement. Texture is used as one of the most important characteristic for identifying defects or flaws. In fact, the task of detecting defects has been largely viewed as texture analysis problem. Because of high requirement of color texture analysis in application to visual inspection, these approaches are of significant importance. These approaches are applicable on variety of defected input images e.g. wood, steel, wafer, ceramics and even non flat objects such as fruits and aircraft surfaces. These are highly demanded by industry in order to replace the subjective and repetitive process of manual inspection. Visual inspection process often involves structural/color analysis and pattern classification for decision making. These approaches utilize the undamaged image structural contents to segregate defected portion. For better outcomes of these approaches, the images should have some predictable structures. One of the images having texture defect is shown in Figure 2. Initially, D. Vitulino et al.  and Z. Hongyinget et. al.  performed major contribution in structure based approaches for scratch detection. Prior to that J. Kittler et al.  proposed a defect detector, consisting of two stages i.e. training phase and defect detection phase.
In training phase, the detector is trained with number of correct images or image regions which are void of defects. In second stage i.e. defect detection phase, the input image is scanned and analyzed w. r. t. the information statistics generated in training phase to identify the location of defect. The procedure of training phase is illustrated below in Figure 3.
In second stage i.e. defect detection phase as shown in Figure 4, the input image after smoothening and noise removal processing, is scanned and analyzed w. r. t. the information statistics generated in training phase to identify the location of defect. Further a new color clustering scheme is explored to segregate the color image texture into various chromatic classes by using histogram based clustering and perceptual merging which is similar to human color perception. This scheme requires a prior knowledge about colors associated with the image texture. The resulting multiple images are then subjected to morphological processing and then represented by texture descriptors.
K.Y. Song et.al  talked about different approaches for generating the structural and texture information of actual image also utilizing its chromatic and luminance information. The structural features are extracted from various chromatic classes associated with the color image texture rather than using only RGB information. The structural texture information is extracted from various chromatic categories by measuring blobs structural statistics of similar colors and the small local chromatic variations are ignored. For each available chromatic category, all the pixels that can be confidently associated with other are defined by setting a single bit binary flag to unity. Thus for each chromatic class, a binary image is generated. An additional binary image is generated for the reject class which contains all the pixels that have not been accepted by any of the chromatic categories. The above process transforms the color macro texture image into a stack of binary blob images. As this process will invariably be noisy hence these images are subjected to morphological smoothing before any structural analysis about blob size, shape and distribution is carried out. Then color clustering and perceptual merging is done based on Euclidian distance criterion which makes it also close to human perception.
Later hybrid properties based algorithm was added to it to tackle the problem of image inspection on random macro color textures. The basic idea of this approach is to represent the random macro color texture by exploring the color texture features at macro level. But here the time consumption for detection is more. Z Wang  introduced hardware and software architecture of quartz wafer defect inspection system consisting of three modules, including image acquisition, image processing and defect detection. Here the image acquisition with dual light source at 45 degrees way of lighting for lighting the wafer was a new method. D. Weimer et al.  presented a machine vision system, which uses basic patch statistics from raw image data combined with a two layer neural network to detect surface defects on arbitrary textured. In this system the features are extracted first from data base and then a neural network model is generated to test the subject image.
These techniques are useful to find the printing defects or drawing defects and not applicable for defects because of external reasons. These approached finds difficulties to detect the defect if the texture is random in nature as an example is shown in Figure 5.
2.3 Properties of defect material based approaches
These approaches are applicable for the detection of defects caused by addition of material on image e.g. ink sprayed, dust, paints etc. In these techniques, the properties of defect creating material i.e. spreading pattern of material, transparency, density etc are explored to separate the defected portion from remaining part of image. An example image with ink sprayed defects is shown below in Figure 6.
For automatic detection of such defect, very little literature is available. RC Chang et al.  explored spreading patterns of sprayed defects in terms of their intensity distribution at their locations and found that these objects has uniform intensity distribution with low and smooth variation. They experimented with ink and wide scratch as defect objects and processed image in HSV space with filtering and iterations to find the smooth and low variance region in the histogram. Then morphological processing followed by image segmentation is done to locate probable defect objects based on size variation w. r. t. intensity variation. The objects with less size variation are declared as defect object. The complete procedure is shown below in Figure 7. Later, G. Rampoly et al.  presented a scheme for foxing detection and V. Bruni et al.  for water blotches using similar properties.
These techniques require number of iterations and give false detection of image objects with similar defect objects.
2.4 Image analysis based approaches
Here basically the statistics of subject images in terms of intensity distribution and contrast are analyzed in gray region. Otsu  provided a well known method for automatically selecting threshold value for defects. This method has limitations in case of images with close to unimodal or bimodal distributions in histograms. Later Hui Fuang et al.  modified Otsu method as Valley Emphasis method and provided good results even in case of unimodal histograms. In Valley emphasis method, the valleys i.e. lowest points of image histogram in gray region are taken as threshold values for image objects categorization. Jiu-Lun Fan et al.  presented revised valley-emphasis method by using the neighborhood information of the valley point. These methods are basically segmenting the image and require further information to categories defect objects. Lijuan Xu et al.  proposed a method to relate local nanostructure variability to nanostructure interactions under the framework of Gaussian Markov random field.
These methods require good contrast for better results. H Alimohamadi et al.  introduced new defect detection method in textile using morphological analysis and Gabor wavelet filters responses. This method consist of three main steps i.e. (i) optimal Gabor wavelet filter selection using image feature matrixes, (ii) morphological analysis of filter response and (iii) thresholding of defect. Basically these methods only divides the image into different objects based on gray level distribution and does not categorize the defect image objects in defect objects and image objects. For better results of these techniques, the image objects should also have good contrast with respect to background. Further to declare the defect objects as defect, these methods would require prior information about defect objects and image objects. It also gives false detection if image have objects with similarities to the defect objects.
All the above discussed defect detection techniques are summarized for their applicability and limitation in table -1.
3 IMAGE INPAINTING /RESTORATION TECHNIQUES
Restoration deals with methods used to recover the original information of degraded image portion. The goal of automatic inpainting techniques is to fill the missing or damaged regions in image utilizing spatial information of its neighboring region without human involvement. The concept of digital inpainting was first introduced by M. Bertalmio et al.  in their revolutionary paper "Image Inpainting" where the original idea of artistic inpainting was adapted by propagating the neighboring color and structure into the subject area. Since then, inpainting has become a popular topic in field of computer vision.
Till now different types of automated methods are available for filling of defected portions of image and all available methods are exploring the remaining undamaged image information, to estimate the possible filling contents for defected portion. In the following section, the available techniques are categorized into three classes according to their basic approach and reviewed for further area of improvements.
3.1 Statistical based methods
These methods are used for the texture synthesis. M. Bertalmio et al.  gave algorithm to synthesize arbitrary sized texture based on a source sample. The basic concept is shown in Figure 9.
The approaches available in this category are of two types i.e. pixels based and multi resolution based. Jing Xu et al.  described a pixel based method where it generates a new artificial texture by growing the borders of input texture pixel-by-pixel and then obtains the possible matches in available patterns. If match is found, the pixel at the center of the generated texture is copied at its destination. It is an iterative process and continues until all pixels have been filled in. This is an extremely flexible method, but its success depends on the size of the template window. A problem with using this approach is that it is very time consuming because the entire source sample area needs to be checked at each new pixel created.
Among texture synthesis methods, Heeger et al.  proposed an approach based on multi resolution sampling. This method creates multiple intermediate images called as image pyramid. These intermediate images represent the texture at different levels with the help of histogram. The histograms of pyramid images are then compared with original image histograms at each level. This results in a synthetic texture after a number of iterations. The process is straightforward and provides good results for stochastic textures, but the matching of histograms fails to capture more structured textures e.g. tiled floors. A similar texture synthesis algorithm was proposed by S. Bonet et al.  in their research related with multi resolution sampling procedure for analysis and synthesis of texture images which improves on Heeger and Bergen's limitations by preserving inter scale dependencies of image pyramids. This algorithm fails when the source texture is more complex e.g. rotationally symmetric textures. Igehy and Pereira  also modified the Heeger and Bergen's algorithm where they allowed smooth transition between the existing image and the synthesized texture. They used an additional composition step with mask of values in the range (0, 1) to have the ratio of the original image to the synthesized texture at each pixel. This is done to avoid any sharp transitions between the original image and the newly synthesized texture. This approach also has same shortcomings of time consumption when attempting to reproduce deterministic textures. Rane et al.  applied these approaches in compression application and wireless transmission of blocks. T.F. Chan et al  developed image statistical models in terms of joints, color patterns by extracting information from the available correct image statistics. Then to fill the defected portion of image, the process is started with input as defected image and keeps iterating to fill the defected portion until it matches the already developed models of the input image. Apart from still images, Nick C et al.  experimented with statistical models for the case of image sequences. The major drawback of these methods is their applicability is limited to highly stochastic textures caused by defects like smoke, water drops etc. These methods also fail in case image textures containing random structures.
3.2 PDE based methods
In these methods, the defected region of an image is filled by diffusion process i.e. smoothly propagating information from the boundary regions towards the interior of region to be filled. This process is simulated by solving nonlinear partial differential equation (PDE) of higher order. Lot of literature is available about these types of methods. M. Bertalmio et al.  proposed inpainting using the mechanism of PDEs and diffusion. The inpainting smoothly propagates the image information along the level lines directions (isophotes) from outside to inside the hole. The isophote direction, denoted by [nabla] [perpendicular to] u is normal to the gradient [nabla] u. It is the direction of least change in gray values and a relaxation on level lines continuity. In the case of image information propagation in defected image area, the smoothness measure is given by the Laplacian in equation (3) and the propagation of the change in image smoothness along the level lines is given by the PDE in equation (4).
L (u) = [u.sub.xx] + [u.sub.yy] (3)
[partial derivative]u/[partial derivative]t = [nabla] L (u) x [nabla] [perpendicular to] u (4)
This evolution only applies to u(x) when x [member of] [OMEGA], with boundary conditions given. In the steady state, when [partial derivative]u/[partial derivative]t = 0, the direction of the largest information change is perpendicular to the isophotes. Figure 10 shows an example of information propagation in an inpainting domain. This scheme based on two components of image i.e. structure and texture. The structure component is filled by using a PDE based method and second texture component is filled statistical based method. This method can inpaint preserving shapes, however it is possible to see the loss of sharpness in the inpainted regions.
Oliveira et al  worked on limitations of Bertalmio i.e. its effectiveness for holes only and design a fast and effective method solely based on diffusion model. They created diffusion barriers to halt diffusion process at pre-defined positions. Their method is quick, but limits itself to reconstruct larger areas and the areas containing texture. In addition, for deciding diffusion barriers, it requires user intervention. T.F. Chan  presented two different inpainting methods by using of second-order PDEs and third-order PDEs models. By using third-order PDE model, the algorithm allows for the curvature of isophote lines. This helps to capture the periodical curving of the isophote lines and also prevents the blurring of smooth areas. Its better results and additional ability to de-noise an image makes this algorithm versatile. The only significant problem with this method is that it is unable to reconstruct image texture, but this algorithm was a good starting point for approaches those can reconstruct both texture and structure. Chan et al.  have developed an elastic based variation. S D Rane et. al.  filled hole in an image by propagating image Laplacians in the isophote direction. A. Rare et al.  explored Edge-based image restoration. Another interesting technique in this field was presented by Tschumperl and Deriche . They presented a general vector-valued image regularization approach which can be applied to numerous computer vision applications e.g. de-noising and image magnifications. This method propagates the information from outside into area to be filled via a structure preserving diffusion method. The main problems with this method are the same as with many of the approaches presented so far i.e. the restored regions becomes blur and filling in of texture areas. P Chen et al.  have derived partial differential equations by formulating the image. W. Yao et al.  worked on PDE based texture synthesis. The main constraints of PDE based methods are applicability for images consists of thin elongated regions and assumption that the content of the defected region is smooth and non-textured.
3.3 Exemplar based methods
These methods have been the most successful up to now. Here the defected regions are filled by copying content from the actual undamaged portion of the image. This is generally achieved by pixel by pixel texture synthesis. Efros and Freeman  tried and overcome some of the shortcomings of pixel by pixel texture synthesis. These methods proposed a technique of stitching blocks of pixels together with one or more other blocks using image quilting process. Criminisi et al.  performed the propagation of textures using a block-based sampling process for removing large objects from digital images, pointing out that the order of the filling process is critical for achieving simultaneous recovery of image structure and texture. This technique fills in the target region with patches from the source region possessing a similar texture. The candidate patches are selected from the whole image, with special priority being given to those along the isophotes (lines of equal gray value) so as to preserve the linear structure during the filling in. Criminisi's method is into frame and assumes blotched areas to be known. So, if the missing regions cannot be found exactly from undamaged areas of the current frame, the restoration may fail. The major disadvantage of this method is the global searching, which not only leads to errors in the match but also greatly decreases system performance. It merely adopts a simple priority computing strategy without considering the cumulative matching error. A different block copying approach was proposed N. Komodakis et al.  where they emphasize the importance of fill ordering. They used patch priorities to decide the filling order based on a confidence term (which favors patches surrounded by more data) and a data term which tries to continue structure into the image. Komodakis et al  researched that the main characteristics of the previously mentioned texture synthesis approaches is that they work by synthesizing only a single pixel at a time. Although these methods work well, but the problem is that these are rather slow and even wasteful because the pixels value is usually predetermined based on their local neighborhood. This essentially allows the structure around the hole to be propagated as long as a certain amount of information around the edges is present.
These methods are relatively fast, accurately propagate and don't blur the image. In addition to this, these are less prone to create artifacts due to overshooting of edges.
The drawbacks of this algorithm is that it can only handle linear edges, bigger size of region to be filled and patch copying phase needs an accurate source for the patches.
At present, all the available restoration techniques are utilizing the remaining undamaged original image information to fill the defected areas. All are only estimating the probable contents for area to fill. But the probability to lose actual image objects which is lost or hidden because of defects is very high, if the objects have rare or nil occurrences in undamaged remaining part of image. This is demonstrated in Figure 12.
All the above discussed restoration techniques are summarized for their applicability and limitation in TABLE-2.
4 CONCLUSION AND FUTURE SCOPE
Automatic and accurate image defect detection and restoration is still a latent research area in the field of image in-painting. A lot of work has been done. The image restoration techniques are at good level. The available defect detection techniques are specific to a particular type of defect. All available techniques have some limitations. The future scope for work is identified as--
1. A generalized, efficient and universal technique for defect detection and restoration.
2. To differentiate the defect objects with similar image objects during defect detection.
3. To bring out the actual image objects which are lost or hidden because of defects.
4. To reduce the time consumption to make applicable for real time inspection so that corrective actions may be taken at same time.
The automatic in painting with above mentioned improvements may have several applications in civil, medical and in industrial field.
[1.] E. Ardizzone, H. Dindo and G. Mazzola. In Recent Advances in Signal Processing. Ch 9: Content-Based Image Retrieval as Validation for Defect Detection in Old Photos, Intech, 2009.
[2.] A. Kokaram. Detection and removal of line scratches in degraded motion picture sequences. Signal processing VIII, Trieste, 1996, 1: 5-8.
[3.] V. Brum and D. Vitulano. A generalized model for scratch detection. IEEE Transactions on Image Processing, 2004, 13(1): 44-50.
[4.] L. Joyeux, O. Buisson, B. Besserer, and S. Boukir. Detection and removal of line scratches in motion picture films. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 1999, 1: 2-7.
[5.] H. Zhang, B. Wu, L. Zhang, and Y. Wu. Uniform gray based arbitrary scratches detection for image inpainting. International Conference on Multimedia Technology (ICMT), 2010: 1- 4.
[6.] R B. Grosjean and Lionel Moisan. A contrario detectability of spots in textured back grounds. Journal of mathematical imaging and vision, 2009, 33 (3): 13-337.
[7.] Y. H. Zhang, CWM Yuen, WK Wong and Chi Wai. An intelligent model for detecting and classifying colored texture fabric defects using genetic algorithm and Elman neural network. Textile research journal, 2011, 81(17).
[8.] X. Si, H Zheng and X Hu. Fabric defect detection based on regional growing PCNN. Journal of Multimedia, 2012, 7 (3): 372-379.
[9.] Xiao Song Wang and Mazid Mir Mehdi. Archive film defect detection and removal: an automatic restoration framework. IEEE transactions on image processing, 2012, 21(8).
[10.] D. V. Bruni and P. Ferrara. Color scratches removal using human perception. Image Analysis and Recognition, lecture notes in computer science, Springer link, 2008, 51(12): 33-42.
[11.] Z. Hongying, W. Yadong, and K. Zhonglin. An efficient scratches detection and inpainting algorithm for restoration of old film. Proceedings of International Conference on Information Technology and Computer Science, 2009, 1: 75-78.
[12.] J. Kittler, R. Marik, M. Mirmehdi, M. Petrou and J. Song. Detection of defects in color texture surfaces. IAPR Workshop on Machine Vision Applications, 1994,13-15.
[13.] K Song J. Kittler and M. Petrou. Defect detection in random color textures. Image and Vision Computing, 1996, 14(9): 667-683.
[14.] Z. Wang. Design of detection system of quartz wafer's defect. International Workshop on Information and Electronics Engineering (IWIEE), Procedia Engineering, Science Verse Direct, Elseviewer, 2012, 29: 3708-3712.
[15.] D. Weimer, H. Thamer and B Scholz Rieter. Learning defect classifiers for textured surfaces using neural networks and statistical feature representations. Forty Sixth CIRP Conferences on Manufacturing Systems, Procedia CIRP, 2013, 7: 347-352.
[16.] R. C. Chang, Y. L. Sie, S. M. Chou and T. Shih. Photo defect detection for image inpainting. Seventh IEEE International Symposium on multimedia, 2005.
[17.] G. Rampoly, F. Stanco and L. Tenze. Virtual restoration of vintage photo-graphic prints affected by foxing and water blotches. GEI journal of Electronic Imaging, 2005, 14(4).
[18.] V. Brum, A. Crawford, D. Vitulano, and F. Stanco. Visibility based detection and removal of semi-transparent blotches on archived documents. VISAPP (1)'06, 2006, 64-71.
[19.] N. Otsu. A threshold selection method from gray level histograms. IEEE Transactions on Systems Man Cybernet, 1979, 9(1): 62-66.
[20.] H. F. Ng. Automatic thresholding for defect detection. Pattern Recognition letters, Science direct, 2006, 27(14): 1644-1649.
[21.] Jiu Lun Fan and Bo Leia. A modified valley-emphasis method for automatic thresholding. Pattern Recognition Letters, 2012, 33(6): 703-708.
[22.] Lijuan Xu and Q Huang. Modeling the Interactions among Neighboring Nanostructures for Local Feature Characterization and Defects Detection. IEEE Transactions on Automation Science and Engineering, 2012.
[23.] H. Alimohamadi, A Ahmadyfard and E Shojaee. Defect Detection in Textiles Using Morphological Analysis of Optimal Gabor Wavelet Filter Response. International Conference on Computer and Automation Engineering (ICCAE '09), 2009, 26-30.
[24.] M. Bertalmio, L. Vese, G. Sapiro, and S. Osher. Simultaneous structure and texture image inpainting. IEEE Transactions on Image processing, 2003, 12(8): 882-889.
[25.] Jing Xu, Daming Feng, Jian Wu and Zhiming CuiAn. Image inpainting technique based on 8-neighbourhood fast sweeping method. International Conference on Communications and Mobile Computing, 2009, 626-630.
[26.] D. Heeger and J. Bergen. Pyramid based texture analysis/synthesis. SIGGRAPH, 1995, 229-238.
[27.] S. Bonet. Multiresolution sampling procedure for analysis and synthesis of texture images. Annual conference series on Computer Graphics, 1997, 31: 361-368.
[28.] H. Igehy and L. Pereira. Image replacement through texture synthesis. International Conference on Image Processing, ICIP, 1997, 3:186-189.
[29.] S. D. Rane, G. Sapiro, and M. Bertalmio. Structure and texture filling-in of missing image blocks in wireless transmission and compression applications. IEEE Transactions on Image Processing, 2003, 12: 296-303.
[30.] T. F. Chan and J. Shen. Variational image inpainting, in Communications on Pure and Applied Mathematics. 2005, 8(4): 579-619.
[31.] Nick. C. and T. K. Shih. Digital inpainting and video falsifying. IEEE International Symposium on Information Technology (ITSim 2008), 2008, 1: 1-8.
[32.] M. Bertalmio, G. Sapiro and C. Ballester. Image inpainting. Siggraph 2000, Computer Graphics Proceedings, Addison Wesley, 2000, 417-424.
[33.] M. M. Oliveira, B. Bowen, R. McKenna, and Y.-S. Chang. Fast digital image inpainting. VIIP, 2001, 261-266.
[34.] T. F. Chan. Nontexture inpainting by curvature-driven diffusions. Journal of Visual Communication and Image Representation, 2001, 12: 436-449.
[35.] T. F. Chan, S. H. Kang, and J. Shen. Euler's elastica and curvature driven diffusions. SIAMJ, 2002, 63(2): 564-592.
[36.] S. D. Rane, G. Sapiro, and M. Bertalmio. Structure and texture filling-in of missing image blocks in wireless transmission and compression applications. IEEE Transactions on Image Processing, 2003, 12: 296-303.
[37.] A. Rare, M. J. T. Reinders and J. Biemond. Edge-based image restoration. IEEE Transactions on Image Processing, 2005, 14: 1454-1468.
[38.] D. Tschumperle and R. Deriche. Vector-valued image regularization with PDEs: A common framework for different applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005, 27: 506-517.
[39.] P. Chen and Y. Wang. A New Fourth-order Equation Model for Image Inpainting. Sixth International Conference on Fuzzy Systems and Knowledge Discovery, 2009.
[40.] W. Yao, J. xiang Sun, G. Zou, S. Teng, and G. Jian Wen. PDE image inpainting with texture synthesis based on damaged region classification. 2nd International Conference on Advanced Computer Control (ICACC), 2010, 5: 447-450.
[41.] A. A. Efros and W. T. Freeman. Image quilting for texture synthesis and transfer. SIGGRAPH, Computer Graphics Proceedings, ACM, 2001,341-346.
[42.] A. Criminisi, P. Perez, and K. Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing, 2004, 13(9):1200-1212.
[43.] N. Komodakis and G. Tziritas. Image completion using efficient belief propagation via priority scheduling and dynamic pruning. IEEE Transactions on Image Processing, 2007, 16(11):2649-2661.
Neelam Bhardwaj, Suneeta Agarwal
Asst. Prof., Computer Science Department, Hindustan Institute of Technology & Management, Agra, UP, India.
HOD, Computer Science Department, Motilal Nehru National Institute of Technology (MNNIT), Allahabad, UP, India.
Table 1. Summary of defect detection techniques Approaches Model Based Image Structural based Applicability Fixed shapes Texture defects, defect e.g. printing defects, Scratch ,lines, drawing defects spots etc Limitations a. False detection in case similarity between image objects and defect objects b. All available schemes are type of defect specific c. Prior database about object images is required d. More time consuming Approaches Defect Image Analysis properties based Based Applicability Defects by Defect objects addition of with good materials e.g. Ink contrast sprayed Table 2. Summary of Inpainting techniques Methods Statistical PDE based Exemplar- based methods Methods Based Methods Images Texture images With Small Images with properties hole type Structural regions uniformity Limitations a. Probability to lose actual image objects, lost or hidden beneath defects, is very high, if the objects have rare or nil occurrences in undamaged remaining part of image. b. No scheme is explored toward bringing out the traces of lost or hidden because of defects. c. No universal scheme, all available schemes are type of defect specific. d. Difficult to fill big size areas accurately. e. Max PSNR achieved is 41
|Printer friendly Cite/link Email Feedback|
|Author:||Bhardwaj, Neelam; Agarwal, Suneeta|
|Publication:||International Journal of Emerging Sciences|
|Date:||Mar 1, 2015|
|Previous Article:||Robust method for codebook generation using embedding algorithm for large image collections.|