Selective Detail Enhanced Fusion with Photocropping

Transcription

1 IJIRST International Journal for Innovative Research in Science & Technology Volume 1 Issue 11 April 2015 ISSN (online): Selective Detail Enhanced Fusion with Photocropping Roopa Teena Johnson PG Student Jawaharlal College of Engineering and Technology, Palakkad, Kerala Manu G Thomas Assistant Professor KJSCE, Vidyavihar(E), Mumbai, India Palakkad, Kerala AmbikadeviAmma T Professor Jawaharlal College of Engineering and Technology, Palakkad, Kerala Abstract The display of a natural scene which exhibits high dynamic range (HDR) on a conventional low dynamic range (LDR) display is normally a challenging task. To solve this problem multiple differently exposed images are captured and fused together into a detailed image. In this paper, a ghost removal algorithm is performed to convert non-consistent pixels into consistent pixels and the corrected image is fused using a selectively detail enhanced exposure fusion algorithm. Thus a detail enhanced images is produced as a result. To this detail enhanced image a photocropping technique uses an unsupervised fuzzy clustering algorithm which converts the image into atomic regions. A manifold embedding algorithm is used for image-level semantics and image global configurations with graphlets or a small-sized connected subgraph. Bayesian network (BN) makes the photo into the framework derived from the multi-channel post-embedding graphlets of the image data. The cropping parameters are calculated by Gibbs sampling method and finally, the enhanced cropped image will be obtained. Keywords: Cropping, Bayesian network, Exposure Fusion, Graphlets, Natural scene I. INTRODUCTION The display of a natural scene which exhibits high dynamic range (HDR) on a conventional low dynamic range (LDR) display is a challenging task. To avoid this challenge the multiple differently exposed images is used to produce a more detailed image. Exposure fusion technique is used to merges all input images information together. The input images may have intensity gap, thus the exposure fusion algorithm is used to make the objects intensity changes. Ghost removal algorithm is used to correct the non-consistent pixels in the input image. The ghosting artifacts are removed to produce a corrected image. The bilateral filter is mainly used to decompose each input image into base layer and its details. The exposure of each image is different and it extracts the fine details based on a bilateral filter used. Then the fusion of the base layer is to generate mask in which the exposure, contrast, and saturation are measured that guides the fusion process. Finally, the fused base layer and the detail layers is combined together for all input images by using selectively detail enhanced exposure fusion algorithm. Then the enhanced image is cropped to each object. Photocropping is a challenging problem mainly due to three reasons: 1) A cropping system only employs a small number of manually defined semantics based on a specific data set. 2) Global spatial configurations are not well preserved in existing cropping models. 3) Existing cropping methods cannot automatically adjust the importance of multi-channel visual features from An image region. This photocropping is a new approach that uses image-level semantics it captures the local composition of each photo by using unsupervised fuzzy clustering algorithm to model the atomic regions and its spatial arrangements. A manifold embedding algorithm is derived to preserve the image-level semantics of each photo. The sampling of a number of candidate cropped photos will produce a multi-channel post embedding graphlets from each candidate cropped photo. A Bayesian network (BN) will measure the quality of cropped photos and finally the Gibbs sampling for parameter inference is being applied. These photocropping steps will help out to crop image through its edges. The manifold graphlet embedding is a new algorithm that will encode image-level semantics and also the photo spatial configurations into graphlets and the Bayesian Network (BN) which will weights the multi-channel visual cues automatically in the post embedding graphlets during the transferring process. The exposure fusion algorithms in earlier times focus on the static scenes. Selectively detail enhanced exposure fusion algorithms and the photocropping process focus on dynamic objects as a new system to produce the enhanced cropped image as a result with better efficiency. All rights reserved by 401

2 The remainder of this paper is organized as follows. Section II reviews the literature survey. In Section III, the methodology of Selective detail enhanced fusion with photocropping, followed by result and the conclusion of the system. II. LITERATURE SURVEY High Dynamic Range (HDR) images in [8] can be generated by taking multiple exposures of the same natural scene. When fusing information from different natural images, least changes in these scenes can generate artifacts. This a technique capable of dealing with a large sum of movement in available exposures, patches which are consistent to a reference image previously selected from the stack. The HDR image has averaging the radiance estimates of all such regions and compensates for camera calibration errors by removing probable seam. The method works even in cases when many moving objects cover large regions of the scene A new image based technique in [9] is used for enhancing the shape and feature of an object. The input is a small set of photographs taken from a fixed perspective, but under varying lighting conditions. For each image, compute a multiscale decomposition based on the bilateral filter and then reconstruct an enhanced image that combines detail information. The bilateral filter is a good choice for using multiscale algorithm because it avoids the halo artifacts commonly associated with the traditional Laplacian image pyramid. Thus develop a new scheme for computing a multiscale bilateral decomposition The paper proposes [7] is a novel method for automatically cropping a photo using a quality classifier that assesses whether the cropped region is agreeable to users. This quality classifier uses large photo collections available on websites where people manually insert quality scores to photos. First trim the original image and then decide on the candidates for cropping. The cropped region with the highest quality score by applying the quality classifier to the candidates. They are not always pleasant to users because they do not take into account the quality of the cropped region. The quality classifier outperforms a method that takes into consideration only the user s attention for automatic photo cropping This paper [6] introduces a graphlets that is a small connected sub graphs to represent a photo's aesthetic features, and a probabilistic model to transfer aesthetic features from the training photo onto the cropped photo is proposed here. In particular, by segmenting each photo into a set of regions, a region adjacency graphs represents the global aesthetic features and graphlets are then extracted from the region adjacency graph. The graphlets capture the local aesthetic features of the photos. Finally, the photo cropping is a process based on a probabilistic model, and it infers the parameters of the cropped photos using Gibbs sampling and the method is fully automatic. III. METHODOLOGY The selective detail enhanced fusion with photocropping will enhance the details of the images at different exposure. The algorithm used to enhance the static and dynamic scene is: Ghost Removal Algorithm. Selectively Detail Enhanced Exposure Fusion Algorithm. A. Ghost Removal Algorithm: The ghost removal algorithm is composed of mainly two modules: Detection Module. Correction Module. In which the detection module detects the non-consistent pixels and. the correction module will correct the non-consistent pixels. Thus, all pixels in the corrected images will become consistent after performing this ghost removal algorithm. The detection module detects the non-consistent pixels by placing the images at different level of pyramid and the normalization of pixels by bidirectional method: In where are the normalized images that is detected and corrected in (1) where the value of η is empirically chosen as 216 for 8-bit images.the ( ) ( ) are images at different pyramid levels. All images are detected and corrected in an order of ( 1 ),...,1,( + 1 ),...,(N 1 ) and N. All images are concerned in the correction of non-consistent regions. As a result, the quality of moving objects and its details will be usually high in the final image. The corrected image is taken as the intermediate image for further process. All rights reserved by 402

3 B. Selectively Detail Enhanced Exposure Fusion Algorithm; The corrected images (1 k N) is combined to form a differently exposed image. They are fused by using an exposure fusion algorithm which includes a unique detail extraction module. The fusion process includes a gradient domain bilateral filter in which the edges of the image are preserved well and the intermediate image is combined with the extracted features. The first step of this exposure fusion algorithm is to build up the vector field which includes fine details of all the corrected images. The corrected image is represented by the variations of the luminance component in log domain. Normally, the gradient of a pixel with the largest absolute value along different exposures are taken. However, the maximum gradient values will be noisy, especially in dark regions of an HDR scene. The vector field is formed using the exposedness level of gradients overall exposures. Thus the weighting factor of a well exposed pixel is larger than that of an under/over-exposed pixel. Weighting factors of a gradient vector (P) also the be the right and bottom pixels of the pixel p and computed in (2). Where the weighting function γ (z ) is defined as: By calculating the gradient parameter and weighting factors the selective detail enhanced fusion is performed on the corrected images. The selective detail enhanced fusion with photocropping is the proposing system. The photocropping technique needs an edge preserving nature so that the images can be cropped well through the edges. So this selective detail enhanced exposure fusion algorithm makes an input image possible for better photocropping. A crop module is added to the proposing system. To the enhanced photo, cropping is done effectively since the noise is less and the image can be cropped out through the fine edges. Fig. 1: Block diagram of Selectively Detail Enhanced Fusion with Photocropping Selective Detail Enhanced Fusion with Photocropping is defined by a block diagram and is shown in Fig. 1.The first block represents the input images and the detection module will detect the non-consistent pixels in the image and the correction pixel are corrected by correction module. The fusion module fuses the input images to detailed enhanced images and on this testing photo the cropping module is performed to crop out fine edges of each object in the image. The algorithms used in photocropping part: Unsupervised Fuzzy Clustering Algorithm Manifold Graphlet Embedding Gibbs Sampling method C. Unsupervised Fuzzy Clustering Algorithm: The Fuzzy clustering algorithm is an iterative clustering method which produce a best possible partition by minimizing the weighted sum of squared error objective function within groups. Apply this algorithm, segmentation to decompose each photo into atomic regions; extract {1,, T} - sized multi-channel graphlets from training photos based on random walking. Thus, construct a three-level spatial pyramid for each atomic region to be labeled. Atomic region s has corresponding cell is divided into a coarse-to-fine manner. This algorithm helps to divide the image into atomic section. Graphlet method can be performed on these atomic regions. All rights reserved by 403

4 D. Manifold Graphlet Embedding: The manifold embedding is used to transform color and texture channel graphlet into dimensional feature vectors. The graphlets from different atomic region are transferred to concatenate appearance feature vectors which contribute to photo aesthetics. First, define two matrices headed for symbolizing the atomic regions and structure. Given a t-sized graphlet in color channel that characterizes all its atomic regions by a matrix, each row of which denotes a 9-dimensional feature vector signifying the color moment of an atomic region. The manifold graphlet embedding is done by: Where Golub-Werman distance and M is the matrix values. Where Y= [,, ], in which and are column vectors standing for the d-dimensional representations of the i-th and the j-th graphlets form the h-th photo and is a function measuring the semantical difference between graphlets.(4) represents two parts. First one is to preserve graphlets Golub- Werman distance and the second is representing image-level semantics. E. Gibbs Sampling: Gibbs sampling is a Markov chain Monte Carlo (MCMC) algorithm used for optimal cropping parameter selection. The colour channel and a texture channel have a probability distribution to calculate the approximate marginal distribution of one of the variables or the subset of the variables or to compute an integral. The Gibbs sampling is done by: Where in (5) p(i(η)) is the probability of a photo I cropped using parameter η. Thus the plotted edges are cropped from the enhanced image. IV. RESULT The selective detail enhanced fusion with photocropping is an improved technique which enhances the image overall details and the enhancement is efficient compared to the early fusion process. The selective detail enhanced fusion will have detection module will detects the pixel and later which will convert the non-consistent pixels into consistent pixels by a correction module. The fusion module will fuse the images together to an enhanced fused image. The cropping applied on an enhanced image will crop out the edge of an object by use of a bilateral filter. The cropping technique is done by graphlets. Thus this is a new system which includes cropping with fusion. Thus it gives better efficiency and the enhanced cropped image is obtained. V. CONCLUSION The HDR images are natural scenes with high range of light variation. The quality of the images may change since it taken at different exposure rates. Selectively exposure fusion algorithm is performed on both static and dynamic scenes. The images of dynamic scenes may have ghosting artifacts and these ghosting artifacts are removed by the ghost removal algorithm. Then fusion process is done to combine the images at multiple exposures. The selectively detail enhanced exposure fusion algorithm is more efficient than the earlier used simple exposure fusion algorithm. Thus detail enhancement is finer and the edge preserving nature of final fused image is due to the gradient domain bilateral filters used. The cropping technique has many drawbacks earlier; due to this the cropping is not properly done. To solve this problem a detail enhanced image is given as an input. The edge preserving nature of the input image will help to plot the fine edges in the image and photocropping will be perfect. The photocropping includes an unsupervised fuzzy clustering algorithm which converts the image into atomic region. Graphlets technique is used to define the edges of the object and the manifold embedding algorithm will combine all the graphlets. The final step is to find the optimal cropping parameters by the Gibbs sampling method. Thus photocropping will be better and efficiently cropped due to this proposing system. Drawbacks of the earlier system are improved by this selective detail enhanced fusion with photocropping method. By using this system will produce a better and efficient enhanced cropped image as the output. REFERENCES [1] E. A. Khan, A. O. Akyuz, and E. Reinhard, Ghost removal in high dynamic range images, in Proc. IEEE Int. Conf. Image Process., pp , Oct [2] F. Pece and J. Kautz, Bitmap movement detection: HDR for dynamic scenes, in Proc. Conf. Vis. Media Prod., London, U.K, pp. 1 8, Nov [3] J. Harel, C. Koch, and P. Perona, Graph-based visual saliency, in Proc. NIPS, pp , All rights reserved by 404

5 [4] J. Hu, O. Gallo, K. Pulli, and X. Sun, HDR deghosting: How to deal with saturation? inproc. IEEE Conf. Comput. Vis. Pattern Recognit., pp , Jun [5] L.Zhang, M.Song, Y.Yang, Q.Zhao, Q.Zhao and N.Sebe, Weakly Supervised Photo Cropping, in Proc. IEEE, pp , Jan [6] L. Zhang, M. Song, Q. Zhao, X. Liu, J. Bu, and C. Chen, Probabilistic graphlet transfer for photo cropping, IEEE Trans. Image Process., vol.22, no. 2, pp , Feb [7] M. Nishiyama, T. Okabe, Y. Sato, and I. Sato, Sensation-based photo cropping, in Proc. ACM Multimedia, pp , [8] O. Gallo, N. Gelfand, W.-C. Chen, M. Tico, and K. Pulli, Artifact free high dynamic range imaging, in Proc. IEEE Int. Conf. Comput. Photogr. (ICCP), pp. 1 7, Apr [9] R. Fattal, M. Agrawala, and S. Rusinkiewicz, Multiscale shape and detail enhancement from multi-light image collections, ACM Trans. Graph., vol. 26, no. 3, pp. 51:1 51:10, Aug [10] S. Raman and S. Chaudhuri, Bilateral filter based compositing for variable exposure photography, in Proc. Eurograph., Munich, Germany, pp. 1 4, Apr [11] Y. Luo and X. Tang, Photo and video quality evaluation: Focusing on the subject, in Proc. ECCV, 2008, pp [12] Z. Farbman, R. Fattal, D. Lischiski, and R. Szeliski, Edge-preserving decompositions for multi-scale tone and detail manipulation, ACM Trans. Graph., vol. 27, no. 3, p. 67, Aug [13] Z.Li, J. Zheng, S.Wu and Z. Zhu, Selectively Detail-Enhanced Fusion of Differently Exposed Images with Moving Objects, in Proc. IEEE, pp , Oct All rights reserved by 405