A framework for measuring sharpness in natural images captured by digital cameras based on reference image and local areas
© Nuutinen et al; licensee Springer. 2012
Received: 26 October 2010
Accepted: 10 May 2012
Published: 10 May 2012
Image quality is a vital criterion that guides the technical development of digital cameras. Traditionally, the image quality of digital cameras has been measured using test-targets and/or subjective tests. Subjective tests should be performed using natural images. It is difficult to establish the relationship between the results of artificial test targets and subjective data, however, because of the different test image types. We propose a framework for objective image quality metrics applied to natural images captured by digital cameras. The framework uses reference images captured by a high-quality reference camera to find image areas with appropriate structural energy for the quality attribute. In this study, the framework was set to measure sharpness. Based on the results, the mean performance for predicting subjective sharpness was clearly higher than that of the state-of-the-art algorithm and test-target sharpness metrics.
Keywordsimage quality reduced-reference metric no-reference metric test-target metric camera phone benchmarking subjective sharpness
Image quality can be measured using objective or subjective methods. Different imaging applications require different methods and metrics. Objective methods for characterizing the performance of digital cameras can be based on test-targets or algorithms. For a test-target metric, a target with known physical properties is captured, and the reproduction is measured. Test-target measurements are tedious and require a controlled laboratory environment. Compared to a real scene, test-target views are easy to interpret and process as desired. Subjective tests, however, cannot utilize captured test-target images. Subjective tests for consumer cameras should be performed using natural-scene views captured under typical photographic conditions. Only natural scenes can show the naturalness and usefulness of images captured by a camera.
Algorithmic methods can facilitate natural-scene picture assessment, and the same image files can be used for both subjective and objective measurements. The algorithmic methods can be classified into no-reference (NR), reduced-reference (RR), and full-reference (FR) methods. This classification is related to reference images' availability and use. An NR metric does not need a reference image, an RR metric needs some information about a reference, and an FR metric needs a pixel-wise reference image. "Pixel-wise" means that the corresponding pixels in two images are found at the same pixel coordinate locations. This is not the case when characterizing cameras, e.g., for benchmarking purposes.
Objective image-quality research aims to develop methods that predict the subjective quality experience. The FR methods are fairly close to achieving this goal . There are different approaches to the FR methods. One is based on modeling the human visual system , and another is based on the structural similarity (SSIM) between images . In addition, natural image statistic (NSS) metrics are promising . The SSIM metric is simple and has many variations [5–8]. Recently, eye-tracking and salience algorithms have been integrated into FR methods [9–12]. These algorithms weight the attractive areas of an image when the spatial values of the quality metrics are pooled into a single quality number. Learning-based models are another new direction. Examples include Moorthy and Bovik's  learned-support vector machine, Eerola's  learned-Bayesian network, and Cui and Allen's  learned neural networks for estimating image quality.
An FR metric is often general. Its output is a single number (e.g., a mean of a spatial distortion map) that provides an overall quality estimation. If the image-quality space is multi-dimensional, as with digital cameras, a single number or distortion map does not explain the quality . Image-quality evaluation can be seen as a hierarchical model. The model includes higher- and lower-level attributes. The higher-level attributes are more subjective. Personal preferences affect their values more than those of the lower-level attributes. For example, naturalness and clarity are higher-level attributes for a consumer digital camera. Graininess, brightness, sharpness, and contrast are lower-level attributes. The lower-level attributes connect to the higher-level attributes. Leisti et al.  claimed that brightness, sharpness, and higher contrast make an image seem clear, and graininess and color brightness affect naturalness. By contrast, faded colors are associated with a lack of clarity.
Higher-level attributes could be predicted using models composed of lower-level attributes. Before such models can be composed and tested, we need robust metrics for the lower-level attributes. There are two reasons why robust metrics are unavailable for digital cameras. The first is that the quality space of digital cameras is multi-dimensional and complex. Digital camera pictures have many different and interacting distortion sources. Signal sharpening, noise removal, and color correction operations also affect the perceived image quality. The second reason is that there is no pixel-wise reference image, due to geometrical differences between images captured by different cameras; therefore, mature FR methods cannot be applied to digital cameras. NR methods do not need reference images, but they are content-dependent, developed for specific distortion types and interact with other distortions.
The method proposed in this study is a compromise between the FR and NR methods. It can be classified into the group of RR methods. The method uses a reference camera, and its application area is camera benchmarking. Camera benchmarking aims to rate the quality of consumer camera systems and determine the reasons behind the differences. The RR method proposed is partly analogous to the test-target and NR methods. The analogy to the test-target methods relates to known patches. Variation in a patch describes the test-target, and variation in a particular area describes the attribute value for the proposed metric. The analogy to the NR methods relates to local area searching. An NR noise metric tries to find smooth areas, and an NR sharpness metric tries to find edges from distorted images. The proposed method is more precise than the NR methods because the areas are located in a high-quality reference image. Compared to the test-target methods, the proposed method reduces the work load in camera benchmarking study because a controlled environment is unnecessary. The potential for finding connections between subjective and objective data is better when the same image can be used for both measurements.
The novelty of the proposed method arises from using a reference image and local areas to compute the image-quality attributes of the natural pictures captured by the cameras to be benchmarked. The contribution of the method lies in utilizing scene features to measure digital camera quality attributes. This is accomplished by identifying local areas from a given scene digitized by capturing with reference camera. A high-quality reference camera plays a key role. The corresponding areas for the images captured by the cameras to be benchmarked are located using area descriptors. The difference from the earlier RR metrics [18, 19] is using local regions for the attribute measurements.
The rest of the article is organized as follows. Section 1 introduces the study. Section 2 reviews the test-target, NR, and RR methods. Section 3 defines the proposed method in detail. Section 4 describes the test setup, and Section 5 shows and analyzes the results. Section 6 concludes the study.
2. Earlier studies
2.1. Test-target metrics
Digital camera quality attributes have widely been measured using test-targets [20–23]. The ISO 12233 standard  describes the method for sharpness and resolution measurements. This method is based on the frequency response of a slanted edge. MTF50 is the spatial frequency at which MTF = 50% (i.e., at which contrast has fallen to half its value at low spatial frequencies). Koren  has argued that the MTF50 value of the frequency response correlates well with the perceived sharpness. In Section 5, we use the MTF50 value as a reference value to assess the performance of the proposed metric.
The test-target methods are good tools for camera characterization. Signal sharpening strength can be estimated by the peak of the spatial frequency response [25, 26]. Recent research has focused on estimating the influence of the combination of signal sharpening and noise removal. The difficulty is that noise removal can filter the image structure while signal sharpening sharpens the edges. Based on the kurtosis  or dead leaf metrics , the sharpening and noise removal combination can be estimated for test patches. However, a test-target metric does not describe the reproduction of specific real scenes. System performance under the conditions in which a typical consumer takes photographs is more interesting for camera benchmarking purposes. Illuminance, color temperature, and scene complexity differ between laboratory and real-scene environments, and cameras use different signal processing parameters for different lighting conditions. Compared to a real-scene, the patterns or colors of test-targets are easy to interpret and process in the camera pipeline as desired.
2.2. NR metrics
The NR methods are applicable for digital cameras. They can be divided into local and global metrics. Local metrics select specific areas from an image, and global metrics use all the image's pixel values in the calculations. Furthermore, the methods can be based on gradients, kurtosis, or singular values, wavelet-decomposition or edge-widths. An NR metric often combines these metrics and transformations.
Edge-width metrics are local. They expect that natural images include sharp edges. In Section 5, we use the Marziliano et al. , Ferzli and Karam , and Narvekar and Karam  NR metrics as benchmarks for the proposed metric. The metrics are based on edge-width analyses. Marziliano et al.  calculated the sharpness value using the edge-intensity profiles after Sobel filtering. Ferzli and Karam  described the just-noticeable blur (JNB) concept. Their sharpness metric compares the edge width and contrast-dependent JNB values. If the edge width is higher than the JNB value, the probability that the image is not sharp increases. Narvekar and Karam  utilize JNB with a cumulative probability of blur detection (CPBD). The CPBD algorithm calculates the percentage of edges at which blur cannot be detected. Liang et al.  have also proposed NR metrics based on edge widths. Liang et al. computed the histogram of vertical and horizontal gradient profiles. The shape of the histogram described the sharpness of the image. Caviedes and Gurbuz  calculated sharpness values using the kurtosis of DCT values from an edge neighborhood.
Global metrics are calculated from given statistical properties of an image. A global metric often uses image gradient values. Singular values and wavelet decomposition have also been used. Zhu and Milanfar  measured sharpness using the singular values of a gradient image. Chen and Bovik  calculated sharpness using the distributions of the gradient and wavelet-decomposition values. Wee and Paramesram  estimated sharpness using the highest eigenvalues of a normalized image. They expected that the dominating eigenvalues would relate to sharpness and the less dominate eigenvalues to noise. Sheikh et al.  described how the NSS model provides reference for the metric. They expected that distortions in the nonlinear dependences of the NSS would be due to image distortions.
2.3. RR metrics
The RR methods suggested so far cannot be directly applied to digital camera characterizations. An RR metric needs information about a reference image. The reference-image framework proposed here makes it possible to use RR methods for digital cameras. We used Wang and Simoncelli's RR metric  as a benchmark for the proposed metric in Section 5.
where k is the number of sub-bands, p k and q k are the probability functions of the k th sub-band in the reference and distorted images, respectively, is the estimate of the KLD between p k and q k and D0 is a constant used to control the scale of the metric.
Other RR metrics also operate in the wavelet domain. Li and Wang  computed RR features using a divisive normalization method for the local normalization of wavelet coefficients. Zhang et al.  calculated a difference vector between the reference and distorted images using the singular values of wavelet decomposition. Xue and Mou  used the Weibull distribution of the wavelet decomposition, whereas Cheng and Cheng  used the Laplace distribution of the gradient image. A literature review of RR metrics can be found in .
3. Reference-image framework
3.1. Framework components
The inputs to the framework are the reference image Ir and the images captured with different cameras, called test images It. Before the analysis, the reference and test images are scaled to the same resolution. The "Scene analysis" component characterizes the scene using the reference image and selects candidate blocks for measurement. The output of the "Scene analysis" component is the vector of candidate points, CP, which includes the pixel coordinates of the so-called candidate blocks. The "Correspondence areas" component locates the blocks in the test images that correspond to the candidate blocks. This location is based on the correspondence features Fr and Ft between the reference and test images. The correspondence features are searched using the well-known SIFT algorithm (scale invariant feature transform). The SIFT algorithm is implemented in the "Correspondence features" component. The correspondence blocks (B i , i = 1,2,...,m) are cropped from the test images and fed to the "Quality metric" component. Finally, the "Quality metric" component applies the attribute metric to the correspondence blocks.
The objective function, f(), is a feedback control of the "Quality metric" component. The objective function is used both for candidate block searching and the quality metric. In this study, the objective function was the standard deviation of the wavelet coefficients. The dual task of the objective function is to locate the high-energy areas from the reference image and measure the corresponding areas' energy from the test images. The proposed reference-image framework is modular, and other quality attributes can be calculated by changing the objective function. The next subsections describe the components' functions in more detail.
3.2. Scene analysis
The "Scene analysis" component computes local values for the candidate blocks from the reference image using the function f(). It aims to find image areas with appropriate structural energy for the quality attribute. The structural energy of the selected areas should change if a change in the quality attribute is perceivable. Smooth regions cannot be used for sharpness measurements because their energy levels can remain unaltered after some low-pass operations.
CP y and CP x are the center coordinates of the block, and M is the block size. Equation (3) maximizes the standard deviation of the first-scale diagonal wavelet coefficients, x, within the block when d(IP, CP) < T. The function d() computes a distance between the IP and CP. The IP are the center coordinates for the predetermined blocks. The CP are the center coordinates for the candidate blocks.
Figure 2b shows how the candidate block locations depend on the scene structure. Because Equation (3) maximizes the first-scale wavelet energy, many candidate blocks include high-frequency structural objects, such as the trees in the image center. In this scene, the candidate blocks do not sample the strong edges between the roof and sky, as the higher-energy regions (trees) can be found in the neighborhood. The most important parameter of the method is the tolerance, T, which limits the size of the search window. Without tolerance, all measuring blocks would migrate to the highest-energy area of the scene. With tolerance, the sampling is more extensive, and the metric acquires more sampling points.
3.3. Correspondence features
The proposed framework utilizes the SIFT method  for correspondence-block searching. The "Correspondence features" component computes feature vectors for the test and reference images (Fr, Ft). The "Correspondence areas" component locates correspondence blocks with the aid of the correspondence features. The SIFT algorithm was selected for the framework because SIFT-based methods are invariant to scaling, translation, and rotation, and they are partially invariant to brightness changes and perspective . In addition, the features are fairly robust to noise.
The SIFT algorithm used  makes a scale-space transformation and calculates the difference of Gaussian (DoG) for the images at different scales. Local extremes are calculated by comparing the DoG sampling points to eight neighboring points in the same scale and nine neighboring points in higher and lower scales. The DoG point is a key point if its value is the highest or lowest in the neighborhood. Other points are excluded. Next, key points with low contrast or along an edge are excluded. Orientation histograms with 36 bins are calculated for the remaining key points. The orientation histogram is based on gradient orientations calculated from the surroundings of a key point after Gaussian filtering. The highest peak of the orientation histogram describes the key point orientation. If the histogram includes peaks with heights that are at least 80% of the highest peak, they are marked as the new key points for that direction. Finally, the key-point descriptors are defined. A key-point descriptor includes the orientation histograms from the neighborhood of the key point. In our implementation, the descriptors have four orientation histograms with eight bins that describe the cell values for the feature vectors.
3.4. Correspondence areas
The SIFT algorithm was applied, and the 20 nearest correspondence features of the candidate block center in the reference image were used for the correspondence-block searching in the test images. The block centers in the test images were located by calculating the angle and length of the vectors from the feature points to the candidate block centers in the reference image.
3.5. Sharpness metric
where the (j, k) are the pixel coordinates in a correspondence block B i , M is the size of the correspondence block B i , b is a parameter for the reduced measurement area, and x j, k is the diagonal wavelet coefficient. The sharpness metric uses m correspondence blocks linked to the m highest-valued candidate blocks. The overall sharpness value is the average value of m correspondence blocks. In this study, m was set between 1 and 24 when studying how the value of m affects performance.
The SIFT algorithm is robust and finds correspondence blocks well, but the correspondence block edge areas can include structures from outside the original candidate-block area. The sharpness metric compensates for this inaccuracy with a reduced measurement area. If the candidate-block size used for block searching is M pixels, the measurement area within the correspondence block in the distorted images is M - b pixels. In this study, the value of M was set to 50, 75, 100, and 125 pixels when studying how the value of M affects performance. The value range of the M was based on the pretest. The parameter value of b was set to 25 pixels.
4. Experimental settings for data collection
4.1. The image contents
The proposed method was validated using two datasets (Datasets I and II). Both datasets included five views (image contents). View contents were designed based on the photo space approach [44, 45]. The photo space describes typical shooting distances and illuminance levels for specific imaging-application areas. Our application was camera-phone benchmarking. By using two different datasets that were collected at different time periods, we were able to validate the robustness of the objective method.
The cameras for the tests were selected to cover a wide quality range: the selection consisted of low-, moderate-, and high-quality mobile phone cameras and moderate-quality compact cameras. The cameras' pixel counts ranged from 3 to 12 Mpix. The views of Dataset I were captured by 13 cameras, and those of Dataset 14 cameras. In addition, all views were captured by a high-quality reference camera. The reference camera was a Canon EOS 5D with a Canon EF 24-80 mm lens. The performance (e.g., signal-to-noise ratio and detail reproduction) of the reference camera was considerably higher than that of the cameras to be tested, which was the only requirement.
Every view was captured several times by each camera. The cameras were set to their automatic mode, as the study benchmarked consumer products. The automatic mode is typically used by consumer photographers. Based on an expert evaluation, one image was selected to represent each camera using the focus on the content target as the criterion. Images with a random white balance or exposure error were discarded.
The most notable differences between Datasets I and II can be found in Contents 4 and 5. They were captured outdoors in Finland, but in different seasons and from different shooting positions. Dataset I was captured in autumn and Dataset II in winter. The differences in Contents 1 and 2 relate only to the people in the images and their clothes. Content 3 is identical for the two sets.
4.2. Subjective test settings
Because of the display size in the subjective tests, images were scaled to a size of 1600 × 1200 pixels. The interpolation method was bicubic. Black borders were also added around the images to match the image file and display resolutions (1920 × 1200). The test setup included two Eizo ColorEdge CG241W displays and a small display. The test images were shown on one display one at a time, and the reference image (Dataset I) or several reference images (Dataset II) were shown on the other. The user interface included sliders for quality attributes and was mounted on the small display.
The observers first evaluated the overall quality value of a test image and selected the attribute values, one of which was sharpness. The other attributes were lightness, saturation, and graininess. The continuous scale was from 0 to 100. All test images representing one content were shown sequentially. The order of the images and contents were randomized between the observers. Near visual acuity, near contrast vision (near F.A.C.T), and color vision were controlled before participation. The viewing distance was approximately 80 cm, and the ambient illuminance was 20 lux. The displays were calibrated using the sRGB standard. We utilized the subjective sharpness data for this article. The following subsections describe the datasets' properties in more detail.
4.2.1. Dataset I
Dataset I included 65 test images (13 cameras × 5 contents). University students were used as observers (n = 25). They were all naïve with respect to image quality. Before the test, all test images and high- and low-quality example images were shown to subjects at a rate of one second per picture. The reference image was shown on one display during the test, and the test images on the other display. The image-quality value of the reference image was set to 90 on a scale of 0-100. The reference image functioned as a high-quality anchor image. A quality value of 90 out of 100 left some latitude for the observers with high-quality test images. Reference images were tuned based on consumer preference expectations. Typical consumers prefer sharp, high contrast, and colorful images.
4.2.2. Dataset II
Dataset II included 70 test images (14 cameras × 5 contents). University students were used as the observers (n = 30). They were all naïve with respect to image quality. Before a single test image of a given content was shown, all test images of the content in question were shown to the observer as a slide show. This process was repeated before each test image was evaluated. We call this method a dynamic reference method. Before the test started, high- and low-quality example images were also shown to the observers. The main difference between Datasets I and II is that the observers saw all images of a given content before a single test image of Dataset II was assessed. Small differences between images were easier to find with this type of presentation. The differences between the Datasets were motivated by the continuous need to improve the procedures for subjective image quality testing.
4.3. Subjective sharpness data
Because the view and shooting environments were the same for both sets, Content 3 (the studio image) gives a good example of how the test method affected the data. The studio image is easy for cameras, as signal processing in modern cameras can handle simple views. This consideration can be seen in the results achieved in Dataset I. The observers had difficulty seeing differences between the images. For Dataset II, the observers saw larger differences between the images. This distinction was due to the dynamic reference, which provided observers with a clear reference for the sharpness scale within the image set under consideration. Based on these results, the dynamic reference method functions well if differences are small but exist.
The problem with the dynamic reference method relates to content-specific normalization. Observers use content-specific anchors when quality values are given. The worst image from the content set functions as a low-quality anchor (quality value = 0), and the best image functions as a high-quality anchor (quality value = 100). We can expect that the same sharpness values for images of different contents would not be identical within Dataset II because some contents are more difficult for cameras to detect than others. For example, the sharpness value of 50 for Content 3 is not the same as the sharpness value of 50 for the other contents. This effect should be considered when data are used for metric validation. Before Dataset II can be used, the objective metric should be normalized in a content-specific manner.
4.4. Reference images
We used two different reference image types: one for the subjective test (Dataset I) and one for the reference-image framework for the objective image-quality metric. The difference between the reference image types arose from the image-processing aims. For the objective image-quality metric, image processing was minimized. Only the default algorithms of white balance, demosaicing, and JPEG format compression were applied. The purpose was to characterize the views of Figure 5.
For the subjective test, the aim was to produce a preferable image. In addition to the default algorithms, image sharpness, contrast, and colorfulness were adjusted according to known consumer preferences. These are assumed to favor sharp, high-contrast, and colorful images.
4.5. Other metrics
We compared the proposed method to the state-of-the-art NR, RR, and test-target metrics. Section 5.2 presents the results. The FR metrics were omitted from the performance analyses reported in Section 5.2; they would not give meaningful results without additional pre-processing, as they require pixel-accurate alignment.
The NR sharpness metrics were from Marziliano et al. , Ferzli and Karam , and Narvekar and Karam , and the RR metric was from Wang and Simoncelli . The NR metrics are local, and previous studies have reported their performance to be high [30, 31]. The RR metric performance  was high in our pretest utilizing the test images captured by the different cameras in this study. We used the published code  for the Ferzli and Karam NR metric , the published code  for the Narveker and Karam NR metric , and the published code  for the Wang and Simoncelli RR metric . We used Marziliano et al.'s  for the MATLAB implementation of their NR metric.
The RR metric  has been developed for overall image quality. The metric compares the wavelet coefficient distributions between the reference and distorted images. Equation (2) calculates the overall image quality. Its high performance at predicting sharpness in our pretests relates to the correlation between image contrast and wavelet coefficient energy. Image contrast relates to the perceived sharpness and detail reproduction.
In addition to the above algorithmic metrics, MTF50 values were calculated. The Mica test-target  images were captured under laboratory conditions. Low-illuminance lighting (100 lux) was used to simulate the conditions of Contents 1 and 2, and high-illuminance lighting (1000 lux) was used for the other contents. The Mica target included low-contrast edges for the frequency-response calculations. A low contrast compensates for signal sharpening effects in cameras. The IE Analyzer v4.0.5 software was used for the calculations, and the reported values are for an average of ten test-target images.
5. Performance results
This section presents the performance of the proposed method. The proposed and other algorithmic methods were applied to the same interpolated image files used in the subjective tests. An interpolation algorithm can filter the images' structure and noise energy. Using the same interpolated image files for the objective metrics, we assessed the same images that the observers saw in the subjective study.
First, we analyzed the character of the proposed metric with different candidate-block sizes and different numbers of the correspondence blocks. Section 5.1 presents these results. Based on these measurements, we selected the optimal settings for the performance comparison between the proposed and the state-of-the-art metrics. Section 5.2 gives these results. Section 5.3 compares the performance of the proposed method for the Gaussian-blurred and JPEG2000-compressed images from the well-known LIVE database. The performance metrics were the Pearson linear correlation coefficient (LCC), Spearman rank-ordered correlation coefficient (ROCC), and root-mean square error (RMSE).
5.1. The influences of block size and number of blocks
The effects of the measurement-region size on the LCC values between the objective and subjective data were studied using measurement area sizes of 25, 50, 75, and 100 pixels. The candidate-block sizes were thus 50, 75, 100, and 125 pixels when located in the reference images. The reference camera's candidate blocks were sorted in ascending order based on the objective function value. When m = 1, the proposed sharpness metric used only the correspondence block of the candidate block with the highest sharpness (i = 1). When m = 2, the proposed sharpness metric used the correspondence blocks of the two highest-sharpness candidate blocks (i = 1,2), and a similar pattern was used for other values of m.
Figure 8b shows the LCC plots as a function of correspondence-block number, i. The proposed sharpness metric was calculated using a single correspondence block. The candidate-block size was 100 pixels. Based on Figure 8b, the selected correspondence block affects performance. The first clear decrease comes when i = 6. The second decrease comes when i = 21. The performance of five first blocks sorted by their Equation (4) values is thus high compared to the other blocks.
Based on these results, we concluded that the optimal candidate block size is 100 pixels and the optimal number of blocks is 5. The candidate blocks were located using a block size of 100 pixels, and the measurement area inside the correspondence blocks was 75 pixels.
5.2. Comparison between the metrics
where p1, p2, p3 and p4 are fitting parameters of the model, Spred is the predicted sharpness, and v i is the metric value for image i. The fitting parameters, p i , were obtained by calculating the minimum least-squares, nonlinear regression using the fminsearch function in MATLAB.
The LCC values of the benchmark metrics and the proposed metric with subjective sharpness.
Based on Table 1 the performance of the proposed metric was higher than that of the previously published metrics. When the data were fitted over all contents, the LCC of the proposed metric was highest. The content-specific performance of the proposed metric was highest, except for Contents 3 and 4 in Dataset I and Content 4 in Dataset II. In these cases, the performance of the Wang and Simoncelli RR metric (Content 3 in Dataset I and Content 4 in Dataset II) or the test-target MTF50 (Content 4 in Dataset I) was the highest.
Dataset II was more difficult for the proposed metric than Dataset I. Overall, the LCC for Dataset I was higher than for Dataset II; even the LCCs for Contents 3 and 4 in Dataset I were low compared to the other contents. Figure 9 shows the higher data dispersion of Dataset II compared to Dataset I.
The mean validation performance values for the metric  and the proposed metric.
The content-specific validation performance values for Dataset I. Boldface indicates the best performer.
The content-specific validation performance values for Dataset II.
Tables 2, 3, and 4 show that the proposed metric performance was also high for the validation data. The mean performance was clearly higher than the Wang and Simoncelli RR metric performance. According to Tables 3 and 4, the content-specific LCC values for the proposed metric were higher than the LCC values of the Wang and Simoncelli metric, except for Content 3 in Datasets I and II.
The RMSE is the most interesting feature of the data. It indicates that the objective metrics outperform the subjective data metrics in predicting sharpness. This result mirrors nature and real problems relating to subjective data. In any case, it should be remembered that goodness-of-fit type metrics, such as the RMSE, are feasible data-quality indicators only for objective data. The dispersion of subjective data may depend on the study material and instructions given. Subjective data should be handled as a probability distribution. It should be characterized and expressed using at least the mean and standard deviation.
5.3. Performance using the LIVE image database
The evaluation of the proposed metric for images from the LIVE database
Based on the results, the performance for the Gaussian-blurred images is high and that for the JPEG2000-compressed images is lower. A reason for low performance of the JPEG2000 images can be the ringing distortion because of heavy compression. The proposed method has not been developed for handling distortions of this type.
The analogy between the proposed and test-target methods arises from using known image properties. Test-target metrics determine the properties of test-target patches, and the proposed method can identify the properties of selected regions in natural images. The advantage of the test-target method is that the test images always contain areas with desired properties for the attribute of interest. The proposed method tries to find appropriate regions in the image. In some cases, performance can be lower if appropriate regions cannot be found in the image.
Based on the evaluation results, the proposed method is promising. The proposed method solves problems related to the NR, FR, and test-target methods. It considers image properties in a more advanced manner than the NR method. It can select the best areas of a scene based on quality attributes. For example, it does not try to interpret noise energy as image sharpness energy. Compared to FR methods, the proposed method does not require a pixel-wise reference image and is thus applicable to performance studies of digital cameras, as in benchmarking camera phones. Compared to the test-target methods, the proposed method does not require the distinctly tedious process of preparing objective test images. With the proposed method, the same natural images captured by cameras can be used for both subjective and objective measurements.
The proposed method predicts the perceived sharpness. The image-quality space of imaging devices is multi-dimensional. Sharpness is one low-level quality attribute. Other low-level attributes, including noise and color reproduction accuracy, are fairly simple to implement in the proposed framework. Both can be measured using the same procedure, which locates the appropriate regions for measurement in the reference image using block data. Higher-level attributes, including naturalness and clarity, set higher requirements for the method. We can expect that new components need to be added in the framework. The higher-level attributes relate strongly to image content and semantics. The framework will require advanced computational methods for content understanding before higher-level attributes or overall image quality can be calculated.
This study was partially supported by the Nokia Mobile Solutions/Symbian Smartphones. The authors thank Emeritus Professor Hannu Saarelma for reviewing the manuscript and Fredrik Hollsten and Jussi Tarvainen for the test images.
- Sheikh HR, Sabir M, Bovik AC: A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Trans Image Process 2006, 15(11):3441-3452.View ArticleGoogle Scholar
- Lubin J: A human vision system model for objective picture quality measurements. In Proceedings of International Broadcasting Conference. Amsterdam, Netherlands; 1997.Google Scholar
- Wang Z, Bovik AC, Sheikh HR, Simoncelli EP: Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process 2004, 13(4):600-612. 10.1109/TIP.2003.819861View ArticleGoogle Scholar
- Sheikh HR, Bovik AC, Veciana G: An information fidelity criterion for image quality assessment using natural scene statistics. IEEE Trans Image Process 2005, 14(12):2117-2128.View ArticleGoogle Scholar
- Yang C-L, Gao W-R, Po L-M: Discrete wavelet transform-based structural similarity for image quality assessment. In Proceedings of IEEE International Conference on Image Processing. San Diego, CA, USA, 12-15 October 2008; 2008:377-380.Google Scholar
- Yang WY, Wu L, Fan Y, Wang Z: A method of image quality assessment based on region of interest. In Proceedings of World Congress on Intelligent Control and Automation. Chongqing, China, 25-27 June 2008; 2008:6840-6843.Google Scholar
- Geary B, Grecos C: Image quality assessment using a rotated gaussian discrimination function. In Proceedings of Computer Vision and Pattern Recognition Workshop. San Francisco, CA, USA, 13-18 June 2010; 2010:47-52.Google Scholar
- Li C, Bovik AC: Content-partitioned structural similarity index for image quality assessment. Signal Process: Image Commun 2010, 25: 517-526. 10.1016/j.image.2010.03.004Google Scholar
- Tong Y, Konik H, Cheikh FA, Tremeau A: Full reference image quality assessment based on saliency map analysis. J Imag Sci Technol 2010, 54(3):030503. 10.2352/J.ImagingSci.Technol.2010.54.3.030503View ArticleGoogle Scholar
- Moorthy AK, Bovik AC: Visual importance pooling for image quality assessment. IEEE J Sel Topics Signal Process 2009, 3(2):193-201.View ArticleGoogle Scholar
- Engelke U, Zepernick H-J: Framework for optimal region of interest-based quality assessment in wireless imaging. J Electron Imag 2010, 19(1):011005. 10.1117/1.3267097View ArticleGoogle Scholar
- Meur OL, Ninassi A, Callet PL, Barba D: Over visual attention for free-viewing and quality assessment tasks. Impact of the regions of interest on a video quality metric. Signal Process: Image Commun 2010, 25: 547-558. 10.1016/j.image.2010.05.006Google Scholar
- Moorthy AK, Bovik AC: A two-step framework for constructing blind image quality indices. IEEE Signal Process Lett 2010, 17(5):513-516.View ArticleGoogle Scholar
- Eerola T: Computational visual quality of digitally printed images. Dissertation, Lappeenranta University of Technology; 2010.Google Scholar
- Cui L, Allen AR: An image quality metric based on a colour appearance model. In Advanced Concepts for Intelligent Vision Systems, ACIVS 2008. Volume 5259. Edited by: J Blanc-Talon, S Bourennane, W Philips, D Popescu, P Schenders. Springer, Heidelberg; 2008:696. France, 20-24 October 2008. Lecture Notes in Computer Science 10.1007/978-3-540-88458-3_63View ArticleGoogle Scholar
- Radun J, Leisti T, Virtanen T, Häkkinen J, Vuori T, Nyman G: Evaluating the multivariate visual quality performance of image-processing components. ACM Trans Appl Percept 2010, 7(3):Article 16.View ArticleGoogle Scholar
- Leisti T, Radun J, Virtanen T, Halonen R, Nyman G: Subjective experience of image quality: attributes, definitions and decision making of subjective image quality. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2009. San Jose, CA, USA, 18-22 January 2009; 2009. 7242, 72420DGoogle Scholar
- Wang Z, Simoncelli EP: Reduced-reference image quality assessment using a wavelet-domain natural image statistic model. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2005. Volume 5666. San Jose, CA, USA, 17 January 2005; 2005:149-159.Google Scholar
- Li Q, Wang Z: Reduced-reference image quality assessment using divisive normalization-based image representation. IEEE J Sel Topics Signal Process 2009, 3(2):202-211.View ArticleGoogle Scholar
- ISO 12233:2000(E). Photography-Electronic still-picture cameras-Resolution measurements 2000.Google Scholar
- ISO 15739:2003(E). Photography-Electronic still-picture cameras-Noise measurements 2003.Google Scholar
- ISO 17321-1:2006(E). Graphic technology and p-tography-Colour characterization of digital still cameras (DSCs)-Part 1: Stimuli, metrology and test procedures 2006.Google Scholar
- ISO 14524:1999(E). Photography-Electronic still-picture cameras-Methods for measuring opto-electronic conversion functions (OECFs) 1999.Google Scholar
- Koren N: The Imatest program: comparing cameras with different amount of sharpening. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2006. Volume 6069. San Jose, CA, USA, 23 January 2006; 2006:195-203.Google Scholar
- Okano Y: MTF Analysis and its measurements for digital still camera. In Proceedings of IS&T 50th Annual Conference. Volume 50. Cambridge, MA, May 1997; 1997:383-387.Google Scholar
- Loebich CM, Wueller D, Klingen B, Jaeger A: Digital camera resolution measurement using sinusoidal siemens stars. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2007. Volume 6502. San Jose, CA, USA, 28 January 2007; 2007:65020N.Google Scholar
- Artmann U, Wueller D: Interaction of image noise, spatial resolution, and low contrast fine detail preservation in digital image processing. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2009. Volume 7250. San Jose, CA, USA, 18-22 January 2009; 2009:72500I.Google Scholar
- Cao F, Guichard F, Hornung H: Measuring texture sharpness of a digital camera. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2009. Volume 7250. San Jose, CA, USA, 18-22 January 2009; 2009:27500H.Google Scholar
- Marziliano P, Dufaux F, Winkler S, Ebrahimi T: Perceptual blur and ringing metrics: application to JPEG2000. Signal Process: Image Commun 2004, 19(2):163-172. 10.1016/j.image.2003.08.003Google Scholar
- Ferzli R, Karam LJ: A no-reference objective image sharpness metric based on the notion of just noticeable blur (JNB). IEEE Trans Image Process 2009, 18(4):717-728.MathSciNetView ArticleGoogle Scholar
- Narvekar ND, Karam LJ: A no-reference image blur metric based on the cumulative probability of blur detection (CPBD). IEEE Trans Image Process 2011, 20(9):2678-2683.MathSciNetView ArticleGoogle Scholar
- Liang L, Chen J, Ma S, Zhao D, Gao W: A no-reference perceptual blur metric using histogram of gradient profile sharpness. In Proceedings of IEEE International Conference on Image Processing. Cairo, Egypt, 7-10 November 2009; 2009:4369-4372.Google Scholar
- Caviedes J, Gurbuz S: No-reference sharpness metric based on local edge Kurtosis. In Proceedings of IEEE International Conference on Image Processing. Rochester, New York, USA, 22-25 September 2002; 2002:53-56.Google Scholar
- Zhu X, Milanfar P: A no-reference sharpness metric sensitive to blur and noise. In Proceedings of First International Workshop on Quality of Multimedia Experience. San Diego, CA, USA, 29-31 July 2009; 2009:64-69.Google Scholar
- Chen M-J, Bovik AC: No-reference image blur assessment using multiscale gradient. In Proceedings of First International Workshop on Quality of Multimedia Experience. San Diego, CA, USA, 29-31 July 2009; 2009:70-74.Google Scholar
- Wee C-Y, Paramesram R: Image sharpness measure using eigenvalues. In Proceedings of Signal Processing, ICSP 2008. Beijing, China, 26-29 October 2008; 2008:840-843.Google Scholar
- Sheikh HR, Bovik AC, Cormack L: No-reference quality assessment using natural scene statistics: JPEG2000. IEEE Trans Image Process 2005, 14(11):1918-1927.View ArticleGoogle Scholar
- Zhang F-Y, Sun T, Tu YF, Qin Q-Q: Reduced reference image quality assessment based on wavelet domain singular value decomposition. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2009. Volume 7498. San Jose, CA, USA, 18-22 January 2009; 2009:74984H.Google Scholar
- Xue W, Mou X: Reduced reference image quality assessment based on weibull statistics. In Proceedings of Second International Workshop on Quality of Multimedia Experience. Trondheim, Norway, 21-23 June 2010; 2010:1-6.Google Scholar
- Cheng G, Cheng L: Reduced reference image quality assessment based on dual derivative priors. IEEE Electron Lett 2009, 45(18):937-939. 10.1049/el.2009.1210View ArticleGoogle Scholar
- Nuutinen M, Orenius O, Säämänen T, Oittinen P: Reference image method for measuring quality of photographs produced by digital cameras. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2009. Volume 7867. San Jose, CA, USA, 17-21 January 2010; 2010:78670M.Google Scholar
- Lowe DG: Distinctive image features from scale-invariant keypoints. Int J Comput Vis 2004, 60(2):91-110.View ArticleGoogle Scholar
- Vedaldi A, Fulkerson B: VLFeat: an open and portable library of computer vision algorithms.[http://www.vlfeat.org]
- CPIQ Phase 1 v.1.10. Camera Phone Image Quality-Phase 1-Fundamentals and review of considered test methods International Imaging Industry Association (I3A); 2007.Google Scholar
- Segur RK: Using photographic space to improve the evaluation of consumer cameras. In Proceedings of IS&T Image Processing, Image Quality, Image Capture, Systems Conference, PICS. Volume 3. Portland, USA, March 2000; 2000:221-224.Google Scholar
- Ferzli R, Karam LJ: JNB Sharpness Metric Software.[http://ivulab.asu.edu/Quality/JNBM]
- Narvekar ND, Karam LJ: CPBD Sharpness Metric Software.[http://ivulab.asu.edu/Quality/CPBD]
- Reduced-reference image quality assessment[http://www.cns.nyu.edu/~zwang/files/research/rriqa/index.html]
- Tervonen A, Nivala I, Ryytty P, Saari H, Ojanen H, Viinikanoja J: Integrated measurement system for miniature camera modules. In Proceedings of IS&T/SPIE International Symposium on Electronic Imaging 2009. Volume 6196. San Jose, CA, USA, 15-19 January 2006; 2006:61960L.Google Scholar
- VQEG, Final Report from the Video Quality Expert Group on the Validation of Objective Models of Video Quality Assessment 2000.Google Scholar
- Sheikh HR, Wang Z, Cormack L, Bovik AC: LIVE image quality assessment database release 2.[http://live.ece.utexas.edu/research/quality]
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.