Improved content aware scene retargeting for retinitis pigmentosa patients
© Al-Atabany et al; licensee BioMed Central Ltd. 2010
Received: 16 April 2010
Accepted: 16 September 2010
Published: 16 September 2010
In this paper we present a novel scene retargeting technique to reduce the visual scene while maintaining the size of the key features. The algorithm is scalable to implementation onto portable devices, and thus, has potential for augmented reality systems to provide visual support for those with tunnel vision. We therefore test the efficacy of our algorithm on shrinking the visual scene into the remaining field of view for those patients.
Simple spatial compression of visual scenes makes objects appear further away. We have therefore developed an algorithm which removes low importance information, maintaining the size of the significant features. Previous approaches in this field have included seam carving, which removes low importance seams from the scene, and shrinkability which dynamically shrinks the scene according to a generated importance map. The former method causes significant artifacts and the latter is inefficient. In this work we have developed a new algorithm, combining the best aspects of both these two previous methods. In particular, our approach is to generate a shrinkability importance map using as seam based approach. We then use it to dynamically shrink the scene in similar fashion to the shrinkability method. Importantly, we have implemented it so that it can be used in real time without prior knowledge of future frames.
We have evaluated and compared our algorithm to the seam carving and image shrinkability approaches from a content preservation perspective and a compression quality perspective. Also our technique has been evaluated and tested on a trial included 20 participants with simulated tunnel vision. Results show the robustness of our method at reducing scenes up to 50% with minimal distortion. We also demonstrate efficacy in its use for those with simulated tunnel vision of 22 degrees of field of view or less.
Our approach allows us to perform content aware video resizing in real time using only information from previous frames to avoid jitter. Also our method has a great benefit over the ordinary resizing method and even over other image retargeting methods. We show that the benefit derived from this algorithm is significant to patients with fields of view 20° or less.
There are thought to be 38 million people suffering from blindness worldwide, and this number is expected to double over the next 25 years . Additionally, there are more than 124 million people who have severely impaired vision. The low vision pathologies of this latter group can be divided mainly into two categories; those that predominantly suffer from a loss of visual acuity such as Macular Degeneration (MD), and those that predominantly suffer from a reduction in the overall visual field, such as Retinitis Pigmentosa (RP). RP in particular (population prevalence ~1:4000 ) causes a tunnel vision with decreasing peripheral fields as the condition progresses.
For those with central visual impairment, conventional low vision aids (LVAs) can provide magnification in order to compensate for reduced visual acuity. Also, electronically enhanced visual aids have been proposed which offer a number of distinct advantages over conventional LVAs by enhancing the contrast without the need of image magnification [3–6].
Severe visual field (VF) impairment (those with a 20° in remaining tunnel or worse) can greatly affect a patient's mobility and navigation. Despite ongoing research into genetic and pharmacological therapies , there is currently no effective treatment for RP patients which can significantly slow or arrest the disease. Traditional low-vision aids for these patients have included de-magnifying optics to expand the remaining visual field of those patients. However, such demagnification comes at the cost of optical (fish-eye) distortion and a loss of resolution (i.e. the objects seem more distant).
Recently, Peli et. al. developed an augmented vision system  which multiplexing minified edges over the original scene on a see-through display. However, there is the potential for inattentional blindness, which is the inability of observers to maintain awareness of events in more than one of two superimposed scenes .
This paper introduces a new method for image retargeting for those with peripheral vision impairment without degrading the resolution or adding more complexities to the visual scene.
Image resizing is an interesting topic in the image processing field, due to the increasing demand for displaying images and videos on a variety of display devices of different resolutions or aspect ratios. Standard image resizing techniques, such as scaling and cropping, are not efficient. Scaling is applied uniformly by reducing the sampling over the whole image. As with its optical (demagnifying) counterpart, it results in the key features becoming smaller and appearing further away. An alternative approach is to use scene cropping, as performed by Suh et al.  and Chen et. al. , which involves finding the best rectangular sub-window in the image to be cropped. This is useful only if there is a single important feature in the image then the image can be cropped and scaled to fit. However, Images with multiple important features present a more challenging case for image cropping.
Recently, important progress has been achieved in the development of content-aware image and video resizing techniques. Liu and Gleicher  proposed a different image retargeting algorithm, which determines a region of interest (ROI) and then applies a novel fisheye-view warping that applies a piecewise linear scaling function in each dimension to the image to achieve a target image size. Their algorithm is simple, but the warping may cause distortions that look unnatural.
Setlur et al. [13, 14] proposed an alternative approach for retargeting large images to small size displays by segmenting the proposed image into a background layer and different ROIs objects, then cuts the ROIs from the image, and fills the holes with an inpainting scheme. Then it rescales the background image and finally pastes the ROIs back to the image. Despite the quality of the resized image, this approach relies strongly on the quality of segmentation which can be difficult, requiring complex feature recognition to be able to be performed properly.
Avidan and Shamir  recently provided a new algorithm called seam carving. This algorithm alters the dimensions of an image by removing a connected path of pixels, called a seam, from an input image repeatedly to achieve a target size. Although the technique shows good results for static images, it has limitations. When an image is overly compressed, the algorithm starts to carve out important objects, yielding unnatural artefacts.
Extending seam carving to video, by treating each frame as an image and resizing it independently, creates jittery artefacts due to the lack of temporal coherency. Rubinstein et al.  improved the algorithm by treating video as a 3D cube and extending seam carving from 1-D paths to 2-D manifolds in a 3-D volume. There are two limitations; Firstly, if there are moving objects which travel from top to down and from left to right, then it will be so difficult to find a 2-D plane that avoid crossing these objects causing unnatural artefacts. Secondly, this algorithm cannot process real time video, and is therefore not useful for those with restricted field of view.
Wolf et al.  proposed a system to retarget video by using non-uniform global warping. Given an input image, their algorithm first computes the importance of each pixel, based on spatial edges, face detection and motion detection. Then, based on the importance map, it forms a system of linear equations solved by a least squares manner.
A similar system was later proposed by Zhang et al. , where they used the same importance map of the Wolf method, but they then used it to calculate a shrinkability matrix. Each pixel of the image is compressed according to it is shrinkage value. This algorithm shows fewer artefacts than the seam carving method because it shrinks the image rather than removing pixels. However, the compression efficiency of this method in keeping the original size of the important objects while minimizing less important features is low.
Kim et al.  recently provided a strip based image and video retargeting technique. Their approach divides the image into strips and compresses each strip individually based on the gradient complexity within each strip. The advantage with this approach is that it can more uniformly distribute the compression across the image. However, this technique can distort the objects in the image. This is because part of one object may be scaled with a scaling factor different to the other parts when an object resides between two strips. Furthermore, this algorithm is not efficient in video as moving objects can cross different stripe boundaries causing significant artefacts in the resultant target video.
From a comprehensive review of the literature it is clear that there are some trade-offs in these prior methods. In this paper we combine the advantages of the seam carving and the pixel shrinkability method, by designing a novel importance map based on the seam carving which we use to shrink the pixels. Also, we show how our method can be scalable to perform content aware video resizing in real time. This is possible as our method only needs to consider previous rather than future frames.
A per-pixel importance matrix is computed indicating the significance of each pixel. This is a combination of two measures: a local saliency gradient map, and a block difference motion detector map.
A modified importance map is computed based on the seam pixel locations.
A shrinkage map is computed from the modified importance map.
This shrinking map is scaled to resize the whole image to the desired k columns.
Finally, a remapping algorithm is applied to re-map each pixel in the original image into its new location in the retargeted image. Thus, a pixel with a low shrinking value is mapped to a distance of approximately one from its left neighbor, while a pixel of high shrinking value is mapped closer to its neighbor.
These five steps are repeated for the vertical resizing.
A) Generating the importance matrix
Our method of generating an importance matrix has similarities to that of Wolf . However, to calculate the saliency map, we use a pyramidal edge detection method to improve computational efficiency. Each pixel in the importance matrix is a combination between the saliency of the current pixel in the source image and the dynamic information of this pixel compared to its location in the previous frame. Values range between 0 and 1, where 0 refers to non- important pixels and 1 refers to high importance pixels.
1) Spatial saliency map
As simple high frequency (small kernel) derivatives of this form can be lossy in their boundary detection, we therefore use a multi-scale pyramidal approach with three kernel sizes to obtain lower frequency (large kernel) spatial derivatives .
2) Temporal saliency map
A Motion saliency map is a map used to identify moving objects. Because the human eye is very sensitive to motion, retargeting dynamic scenes while preserving the temporal context is very important. Motion is detected based on a method proposed by Liu et al. , which is relatively easy to implement and requires low computational power. The image is divided into NxN blocks and motion in each block is calculated by taking the weighted average of intensity difference of each pixel, so that the motion map W T (x, y) is set to one if the block containing the pixel (x, y) has motion, and zero otherwise. We uses N = 4 in all the processed images and videos in this paper.
B) Modifying the importance matrix
In the seam carving method, a cumulative energy map is generated based on the spatial importance map similar to that described in equation 3. We modify this to include temporal saliency as given in equation 5. We then search for seams with lowest energy and then give them very low importance values.
Where x(i) represents the column number for a given row, as the seam path should has only one pixel in each row of the image. This condition |x(i) − x(i − 1)| ≤ 1 is to make sure that the pixels along the seam path are connected.
Then starting from the last row in the M matrix, we search for the minimal cumulative pixel. After that, we work backwards from this pixel to obtain an optimal vertical seam by finding the minimum of the three neighboring pixels of this pixel in the previous row and then save this pixel to the seam path. At this point, Avidan and Shamir  would adjust the image width by removing this optimal vertical seam. These steps are repeated until the desired size is achieved.
C) The shrinkability matrix
The summation of S(j) over j columns equals 1 if K is 1. For higher values of K, i.e.number or columns to be removed/shrunk, the summation of S(j) is equal to K. This equation is repeated for the rows.
D) Scaling the Shrinkability matrix
We repeat equations (16) and (17) until the summation of S'(J) equals to K. To avoid the total removal of any pixels, the threshold value is set to 0.9 and not 1. This method is much faster and more accurate than the one described by Zhang et al. , which applied a binary search method to find the best Ko value that is close to the required K.
E) Remapping algorithm
The scaled cumulative shrinkability is used to resample the source image to the retarget image by using an algorithm suggested by Karl M. Fant . The algorithm is a 1 D method used in separable transformations defined in terms of forward mapping functions. It maps a limited line of discrete input pixel intensity values into a limited line of discrete output pixel intensity value. A full description can be found in Fant's paper, but we will summarise it briefly here.
Where S is a variable scaling factor, a size factor of the output data in relation to the input data which changes from pixel to pixel in the same line and 1/S is the inverse of the size factor which indicates how much of an input pixel contributes to each successive output pixel.
F) Dynamic scene retargeting
To extend our algorithm into dynamic scene retargeting, we take into consideration some constraints from the previous frame. As previously mentioned, calculating the importance map for each frame individually generates jittering artefacts in the retargeted video sequence. We therefore start by considering spatio-temporal importance maps rather than considering the spatial importance alone (equation 4). If we relied purely on the spatio-temporal importance map, seams could cross to the other side of a significant feature, causing significant jitter. To counteract this, movement of seams from one scene to the next need to be constrained.
We therefore calculate the seams of the first frame or first couple of frames, and then the actual locations of these seems are stored into an arbitrary matrix T to be used in calculating the seams for the next frame. These seams' locations are adapted, for the forthcoming frames, if there are dynamics in the scene. If the objects in the frame are static then the locations of the seams will be the same, but if these objects move then the seam locations within the same areas through which the objects move around will also move to avoid crossing the moving objects.
By iteration, we have found that the best threshold value Z is 20% greater than the energy sum of the previous seam. Lower values of Z result in more frequent re-evaluations of the seams. This is more sensitive but results in jitter effects. Higher values of Z, increases stability with respect to jitter, but could result in distortion of the geometry of significant features.
G) Downsampling to improve computational efficiency
There is one additional advantage of our algorithm over the seam carving method. In the seam carving approach, the process of generating the energy map and searching for the seams should be applied on the full size of the image to remove the seams repetitively. However, in our approach we generate seams to create an importance map rather than removing them from the scene. Hence, we can resize the source image into smaller one and generate the importance matrix from this downsampled image. After that, we resize the importance matrix into the original size of the source image. For example, if we down sized the source image by half, the overall processing time will be reduced by more than 60%, as most of the processing time is spent in generating and updating the importance matrix. This approach works well for high contrast scenes, though less well for more complex lower contrast scenes.
Evaluation of Performance
The evaluation process is divided into two sections; synthetic evaluation based on testing the performance of our algorithm compared to other retargeting methods, and evaluations based on testing our algorithm on real subjects.
If not done so already with automatic gain control, the video's frames luminance intensities are normalized between 0 and 255. We then applied our video retargeting algorithm for different retargeting sizes on this video file, from 20% to 62% of the original size, and did the same for the seam carving and shrinkability methods. The efficiency of the three algorithms was compared with respect to the ordinary video resizing by interpolation. For comparison, and to avoid subjective human interpretation, we measured three objective parameters:
1) Recognisability with respect to compression ratio
To measure this parameter we used optical character recognition (OCR) to count the number of correctly recognized words along the whole video frames. We developed our OCR interpreter based on the character template matching approach .
We measured the ratio between the average sizes of the three text boxes; when using the three algorithms, and when using the interpolation method.
In the original video file we inserted four small blue circles (top, bottom, left and right). We measured the average vertical and horizontal displacement of these circles along the whole video frames to determine alignment changes after processing.
In addition, our algorithm was objectively and subjectively evaluated on real subjects. In the objective evaluation, our algorithm was tested on 20 volunteers using a simulated tunnel vision. Two pair of goggles, covering the whole eye area, was artificially painted, apart from an aperture, to simulate 10° and 20° tunnel vision effects. Objective testing was divided into 4 sections.
The first section was to assess the affect of our algorithm on speed of recognition. 26 images were displayed to the participants (13 with original size and 13 retargeted to 53% using our method). Images were projected to the participants using the NEC LT280 projector with resolution of 1024 × 768 and maximum projection brightness (At a distance of 2 m from the projection wall) of 2500 Lumens, in a darkened room. The projected area for the original image was (225 × 180 cm) which equivalent to 70° field of view. The 13 images included 3 synthetic images with colored shapes of squares, triangles and circles. The remaining 10 real-world images included football players with colored shirts on a grass background. Participants were asked to count the number of certain colored shapes (synthetic images) or shirts (real-world images) and allowed to move their head freely. The time taken to count these objects was measured for both of the original and retargeted images. This experiment was done for two simulated tunnel vision degrees; 11° and 22° field of view (FOV), respectively. Of the 20 participants, 6 were used as control subjects to do the same task without wearing the goggles. Those 6 subjects did both the whole test and the control part. Samples of the projected images are presented in the Appendix.
The second section of the test measured the efficiency of our method compared to the ordinary (linear) rescaling method. Three different images were compressed to 10%, 20% 40% and 50% of the original size using standard rescaling and our seam assisted shrinkability algorithm. The images were projected to the participants starting from the smallest size (10%) to the largest, and the ordinary rescaled images were displayed before our retargeted images. Participants were able to move their head freely and asked to count the number of colored shapes and the number of people in the images. The percentages of recognized objects were measured with respect to the total number of objects (total of 21 objects in the 3 images) in each scale for the ordinary resized images and our retargeted images.
The third section determined to what extent the compression algorithm could increase the effective field of view. The participants were asked to watch 3 video files (1 synthetic and 2 real-world) which included specific actions and objects. The participants were asked to count the number of times they perceived certain events in their peripheral field. The test was repeated for 40% retargeted versions using our algorithm. The percentage of detected objects was measured with respect to the total number of objects and actions (a total of 17 objects and actions occurred in the 3 video files) for both, the original and retargeted video. In two of the videos, the participants were asked to follow the moving object and in the third one they were free to move their heads. A Full description for the three video files used in this section is explained in the Appendix.
The last section measured the efficiency of our method over the ordinary rescaling method in recognizing certain actions in video. We compressed video files to 25%, 35% and 45%, respectively, of the original size using the ordinary rescaling and our method. The video was projected to the participants starting from the smallest size (30%) to the larger one, and the ordinary rescaled one is displayed before our retargeted one. The subjects were allowed to move heads freely. The percentages of recognized actions were measured with respect to the total number of actions (total of 4 actions) in each scale for the ordinary resized and our retargeted version. The description of the 4 actions is presented in more detail in the Appendix.
The subjective test of this part is divided into two sections. Firstly, the performance of our seam assisted shrinkability method was subjectively compared to the seam carving and shrinkability only methods on healthy subjects. We then separately performed a comparison of our seam assisted shrinkability algorithm with ordinary (linear) resizing. This second study is to explore the tradeoffs' between improved object sizes with inevitable levels of jitter. 6 video files were used in this evaluation, which were each retargeted with the three algorithms plus retargeting them with the ordinary resizing technique. The first video sequence consisted of four subjects passing a tennis ball to each other while they are standing and not moving. In the second video sequence, 5 subjects pass a basket ball while they are moving in a circular form. The third video consists of a subject moving (in the middle of the scene) towards the screen from a distance while three other subjects enter and exit from the frame. In the fourth video, a subject is playing with golf balls, so the motion in this video was considerably slow. The fifth video consists of a subject skating on the water. The final video consisted of eight subjects dancing in a small room. This was the busiest of the 6 videos.
Sizes of the original and retargeted videos used in the preference experiment
240 × 320
240 × 180
240 × 320
240 × 200
240 × 320
240 × 200
400 × 700
400 × 300
210 × 320
210 × 160
200 × 420
200 × 200
We present the results in two sections; the synthetic performance of our algorithm compared existing methods, and results obtained from human testing with simulated tunnel vision.
A) Results from synthetic testing
The middle row of Figure 8 shows the effect of increasing the noise level on the compressibility performance. When a small compression rate is required and the noise level is low (which means the background gradient is low), our seam assisted shrinkability method efficiency is high. This is because the seams will be accumulated in the low gradient areas, causing these areas to be shrunk more. Increasing the noise level (gradient of background becomes high) drops the compressibility performance in all algorithms. Our algorithm performs better than shrinkability only, but starts to approach this latter algorithm at high noise levels during lower compression.
Also from the last row of Figure 6 and Figure 8, we see that the number of unaligned pixels tends to decrease with increasing the noise level and when high compression rates are required. This is because increasing the noise level gives the pixels increasingly equal energy, resulting in some of the seams crossing into the text box. Hence, the seams will tend to go straight from down to top causing less misalignment effect.
B) Results from human trials
The synthetic results indicate that our method is superior to both the standard resizing method and previous seam carving and shrinkibility retargeting methods. Simulated patient trials indicate that greatest benefit is achieved in those of remaining field of view less than 20°. The algorithm would therefore be also useful in retinal prosthetic devices  which will in initial implementations only return a limited visual field. The recent optogenetic approaches  look particularly promising and could potentially return higher resolutions. However, assuming the technology can approach a level of development of present head mounted displays for virtual reality, the fields of few will still be limited to typically 40°. Thus some form of compression would greatly assist such users. As the visual acuity may still be low, scene simplification and enhancement operations may additionally be needed .
Ultimately, this work has shown efficacy in scenes whereby the camera is static and objects in the scene move. In this situation creating an energy map for the motion energy is straight forward and obtained from the movement of objects across the frames. However, in the situation where the camera moves relative to the environment, the movements of all features would need to be subtracted from the background movement. This background movement would then have to be calculated from optic flow analysis on the images or from accelerometer motion sensors.
A further point to be considered is that this presented algorithm has been designed for monocular vision. In binocular cases, the situation is more complex as slight differences in left and right scenes may vary the compression such that objects to not appear to overlap in 3D space. We will thus look to further develop this technique in future for binocular use.
Downsampling effect on the processing time
Processing time (sec)
Shrinkability & Fant
We can see that the processing time of the importance map drops approximately by 4 when downsampling the input frames by half of the original size. However, the time taken in generating the shrinkability map and remapping by Fant's algorithm is fixed because these processes applied on the whole size of the input frame.
Currently, all of the algorithm's parts are running serially on the computer CPU. However, some parts of the algorithm can be optimized to run in parallel on a graphic processing unit (GPU) or even on a portable field programmable gated array (FPGA) device . For example, most of the importance map parts are based on the convolution kernels which can be implemented in parallel. Not only can the importance map be parallelized, but also the process of generating and scaling the shrinkability map. Such parallel processing can be performed in portable GPU or FPGA devices. Because the Fant's algorithm remaps each row individually on the retarget frame, it is also possible to implement this in parallel architecture. In future work we hope to implement this algorithm on a portable parallel processing platform so as to perform real time RP patient trials. As is generally accepted in the graphics processing community, parallel processing using GPU architectures can speed up the processing time in the range of 10-100× depending on the level of parallelism of the program. Although 25frames per second is sufficient for video rate, 50 frames per second is widely accepted as the minimum to reduce strain through motion blur. We believe this target can be achieved in portable GPU architectures.
In this article we have described a novel content aware scene retargeting technique developed for patients suffering from retinitis pigmentosa. Our approach tackles the shortcomings of the two current techniques for image retargeting (seam carving and image shrinkability). In particular we have improved upon issues such as discontinuity artefacts and jitter in real time video sequences.
We combined the advantages of both techniques, by designing a novel importance map which gives the pixels along the seams paths lower importance values. Then we used the modified importance map to build a shrinkability matrix to shrink the pixels according to their shrinkage values.
Results show the robustness of our approach compared to the seam carving and image shrinkability techniques in preserving the scene contents intact and in the compressibility performance.
Written consent was obtained from the participants for publication of this paper and accompanying images. A copy of the written consent is available for review by the Editor-in-Chief of this journal.
Appendix: Description of the trials
Figure 17(b), shows sample from the second part of the test. The lowest size of the image is displayed first to the participant to count the number of objects in the image (objects here is the total number of people), as shown in the right pair of images. Then the larger size is displayed to count the number of objects, as shown in the left pair. The image to the left in each pair is the one resized linearly and the one to the right is retargeted using our method, respectively.
The last section in the test includes only one video which compressed into 25%, 35, and 45% of its original size using linear rescaling and our retargeting method. The movie includes 4 persons moving around while doing specific actions (reading, drinking, eating a banana, and holding a child). The participants are asked to recognize the actions of these persons in each scale. Figure 18(c, d), shows a snapshot from the 25% version of the video, when linearly resized and when using our retargeting method, respectively.
We would like to acknowledge and thank The University of London Central Research Fund (AR/CRF/B), the Royal Society Research fund, the National Institute for Health research (BRC) fund and the EPSRC (F029241) for supporting this research. Also Mr. Walid Al-Atabany would like to thank the Egyptian government, who are sponsoring him for his PhD. We would also like to thank Dr. Muhammad Ali Memon and Dr. Susan Downes for our collaborations on past, present and future patient trials in this augmented vision area.
- Foster A, Resnikoff S: The impact of Vision 2020 on global blindness. Eye 2005, 19: 1133–1135. 10.1038/sj.eye.6701973View ArticleGoogle Scholar
- Hamel C: Retinitis pigmentosa. Orphanet Journal of Rare Diseases 2006, 1: 40. 10.1186/1750-1172-1-40View ArticleGoogle Scholar
- Vargas-Martín F, Peláez-Coca MD, Ros E, Diaz J, Mota S: A generic real-time video processing unit for low vision. International Congress Series 2005, 1282: 1075–1079. 10.1016/j.ics.2005.05.107View ArticleGoogle Scholar
- Fullerton M, Peli E: Post Transmission Digital Video Enhancement for People with Visual Impairments. J Soc Inf Disp 2006, 14: 15–24. 10.1889/1.2166829View ArticleGoogle Scholar
- Wolffsohn JS, Mukhopadhyay D, Rubinstein M: Image enhancement of real-time television to benefit the visually impaired. Am J Ophthalmol 2007, 144: 436–440. 10.1016/j.ajo.2007.05.031View ArticleGoogle Scholar
- Atabany W, Degenaar P: A Robust Edge Enhancement Approach for Low Vision Patients Using Scene Simplification. Cairo International Biomedical Engineering Conference CIBEC 2008, 1–4. full_textGoogle Scholar
- Bennett J: Gene therapy for retinitis pigmentosa. Current Opinion in Molecular Therapeutics 2000, 2: 420–425.Google Scholar
- Peli E, Luo G, Bowers A, Rensing N: 22.4: Invited Paper: Augmented Vision Head-Mounted Systems for Vision Impairments. SID Symposium Digest of Technical Papers 2007, 38: 1074–1077. 10.1889/1.2785492View ArticleGoogle Scholar
- Simons DJ: Attentional capture and inattentional blindness. Trends in Cognitive Sciences 2000, 4: 147–155. 10.1016/S1364-6613(00)01455-8View ArticleGoogle Scholar
- Suh B, Ling H, Bederson BB, Jacobs DW: Automatic thumbnail cropping and its effectiveness. In Proceedings of the 16th annual ACM symposium on User interface software and technology. Vancouver, Canada: ACM; 2003:95–104. full_textView ArticleGoogle Scholar
- Chen LQ, Xie X, Fan X, Ma WY, Zhang HJ, Zhou HQ: A visual attention model for adapting images on small displays. Multimedia Systems 2003, 9: 353–364. 10.1007/s00530-003-0105-4View ArticleGoogle Scholar
- Liu F, Gleicher M: Automatic image retargeting with fisheye-view warping. In Proceedings of the 18th annual ACM symposium on User interface software and technology. Seattle, WA, USA: ACM; 2005:153–162. full_textView ArticleGoogle Scholar
- Setlur V, Takagi S, Raskar R, Gleicher M, Gooch B: Automatic image retargeting. In Proceedings of the 4th international conference on Mobile and ubiquitous multimedia. Volume 154. Christchurch, New Zealand: ACM; 2005:59–68. full_textView ArticleGoogle Scholar
- Setlur V, Lechner T, Nienhaus M, Gooch B: Retargeting images and video for preserving information saliency. IEEE Computer Graphics and Applications 2007, 27: 80–88. 10.1109/MCG.2007.133View ArticleGoogle Scholar
- Avidan S, Shamir A: Seam carving for content-aware image resizing. ACM Transactions on Graphics 2007, 26: 3. 10.1145/1276377.1276390View ArticleGoogle Scholar
- Rubinstein M, Shamir A, Avidan S: Improved seam carving for video retargeting. ACM Transactions on Graphics 2008, 27(3):1–9. 10.1145/1360612.1360615View ArticleGoogle Scholar
- Wolf L, Guttmann M, Cohen-Or D: Non-homogeneous content-driven video-retargeting. 2007 IEEE 11th International Conference on Computer Vision 2007, 1–6: 1418–1423.Google Scholar
- Zhang YF, Hu SM, Martin RR: Shrinkability Maps for Content-Aware Video Resizing. Computer Graphics Forum 2008, 27: 1797–1804. 10.1111/j.1467-8659.2008.01325.xView ArticleGoogle Scholar
- Kim J-S, Kim J-H, Kim C-S: Adaptive image and video retargeting technique based on Fourier analysis. Proc IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops CVPR Workshops 2009, 1730–1737.Google Scholar
- Fleck MM: Some defects in finite-difference edge finders. #IEEE_J_PAMI# 1992, 14: 337–345.Google Scholar
- Canny J: A Computational Approach to Edge Detection. #IEEE_J_PAMI# 1986, PAMI-8: 679–698.Google Scholar
- Jobson DJ, Rahman Z, Woodell GA: A multiscale retinex for bridging the gap between color images and the human observation of scenes. #IEEE_J_IP# 1997, 6: 965–976.Google Scholar
- Liu SC, Fu CW, Chang SY: Statistical change detection with moments under time-varying illumination. IEEE Transactions on Image Processing 1998, 7: 1258–1268. 10.1109/83.709658View ArticleGoogle Scholar
- Bovik AC: Handbook of Image and Video Processing. Orlando, FL, USA: Academic Press, Inc; 2005.Google Scholar
- Fant KM: A Nonaliasing, Real-Time Spatial Transform Technique. IEEE Computer Graphics and Applications 1986, 6: 71–80. 10.1109/MCG.1986.276613View ArticleGoogle Scholar
- Spitz AL: Shape-based word recognition. International Journal on Document Analysis and Recognition 1999, 1: 178–190. 10.1007/s100320050017View ArticleGoogle Scholar
- Degenaar P, Grossman N, Memon MA, Burrone J, Dawson M, Drakakis E, Neil M, Nikolic K: Optobionic vision-a new genetically enhanced light on retinal prosthesis. Journal of Neural Engineering 2009, 6: 035007. 10.1088/1741-2560/6/3/035007View ArticleGoogle Scholar
- Al-Atabany W, Memon M, Downes S, Degenaar P: Designing and testing scene enhancement algorithms for patients with retina degenerative disorders. BioMedical Engineering OnLine 2010, 9: 27. 10.1186/1475-925X-9-27View ArticleGoogle Scholar
- Atabany W, Degenaar P: Parallelism to reduce power consumption on FPGA spatiotemporal image processing. Proc IEEE International Symposium on Circuits and Systems ISCAS 2008, 1476–1479.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.