This comes from some work I did on a contract a few years ago. While I obviously can’t publish everything related to that work, I had retained the copyright to some of the low-level implementations exactly so that I could open source them later.
(N.B.: This code is now explictly licensed, per the LICENSE file.)
What follows are some notes I had written about a broader approach that was built on this code. I expect that research in this area has since surpassed this approach.
This is an attempt to synthesize an algorithm for automatic ranking of digital photographs, based on the following basic assumptions about quality:
- good photos have the subject matter in focus;
- the focal mass will generally lie close to the points defined by the rule of thirds;
- the subject matter will also be separated from the background by brightness and saturation.
We note that the sharpness of a region is a reasonable predictor of focus, as most prominent subject matter contains detailed edge structures that are destroyed by out-of-focus/motion blur.
After the description of the algorithm and the work on which it is based, I discuss directions for improvement.
Our approach broadly builds upon the approach suggested by Lim, Yen, and Wulim-yen-wu, though their goal is simply making a decision as to whether an image is out-of-focus or not.
A related set of ideas for determining the aesthetic value of an image is given by Liu et al.liu-et-al, but their focus is on reframing an existing image to capture the most aesthetic possible cropping of an image.
Given an image, we first decompose it into square blocks of pixels. We have chosen 64x64 pixels rather than the arbitrary 100x100 suggested in lim-yen-wu to facilitate the incorporation of Discrete Cosine Transform (DCT) information in the common case that the image is JPEG encoded.
For each block, we compute the hue, saturation, and intensity of each pixel. Note that it might be possible to speed up the algorithm here by skipping this computation and using only luminance values for blurred blocks.
We wish to make a boolean decision about whether this block is likely to be in focus, based on whether the sharpness of this block exceeds a threshold. In the present implementation, we have three different sharpness metrics. We will describe each, as well as a proposed unification of the methods.
In the case that we have access to DCT information, we can perform a simple histogram algorithm due to Marichal, Ma, and Zhang (marichal-ma-zhang). For each 8x8 DCT block in this image block, we count the occurance of coefficients greater than a minimum value in a histogram. After each DCT block has been considered, we compare each value in the histogram against the histogram value for the first coefficient (0,0), summing a weighted value when the value exceeds a threshold.
This provides an extremely rapid approximation of blur extent. In the case that this value exceeds certain thresholds, it would be possible to skip further computation of more accurate sharpness metrics.
Tong et al.tong-li-zhang-zhang present an approach to blur estimation based on the Haar wavelet transform, whereby specific edge structures can be detected, and the ratio of Dirac and abrupt-step structures to roof and gradual step edge structures provides an acceptable estimation of the sharpness of the given block.
In our case, we have modified this algorithm to skip the windowing step, as we felt that adjusting the thresholds by which a point was classified was more effective than performing non-maximum suppression (NMS).
Unfortunately, even without NMS, this is the most computationally expensive of the algorithms implemented.
In his master’s thesis ramakrishnan-thesis, Harish Ramakrishnan presents some variants on this approach and evaluates their performance.
Shaked and Tastlshaked-tastl derive a model of sharpness based on the ratio between high-frequency data and low-frequency data, and present an approach to computing this metric via a pair of one-dimensional IIR filters. The advantage of this decomposable approach is that rows or columns can be skipped to speed up the process at the cost of accuracy.
Presently we use simple Butterworth high- and low-pass filters on row and column data to compute the sharpness metric. However, it seems likely that better filter designs could help improve the accuracy and performance of this model.
A Canny edge detector could be used in much the same way as the Haar transform approach given above. It is possible that this could be more efficient.
Another alternative, discussed in ramakrishnan-thesis, is the perceptual blur metric given by Marziliano et al. in marziliano-et-al. We have thus far not considered using this blur metric as it seems less efficient than the methods we already implement.
In general, we avoid using the IIR filter in preference to the wavelet transform approach or the DCT coefficient histogram. Ideally, we would use the fastest and least-accurate methods (DCT coefficients, IIR filter with a large row/column skip) to eliminate clearly blurry blocks, and then only use the more expensive wavelet transform approach on blocks whose sharpness isn’t evident.
Along with a sharpness value for each block, we also compute the means of the block’s hue, intensity, and saturation values, as well as a ratio of the number of pixels having a hue similar to that of blue sky (per lim-yen-wu).
Having computed these metrics for each block in the image, we compute several global indicators of quality. Blocks whose dominant hue is that of blue sky are ignored during this process.
We compute a composition score as a sum of the blocks which are sharp, weighted by the Manhattan distance of the block from one of the four “power points” representing intersections of the division of the image in thirds, as per the rule of thirds.
We compute brightness and saturation indices as the difference between the mean of those values for blocks considered sharp and the mean of those values for the remaining blocks.
We also compute the density of sharp blocks, as the ratio of sharp blocks to unsharp ones, and the median block sharpness.
This is where work needs to proceed: improving the quality of indices other than composition so they can be combined with appropriate weights to produce a single ranking value.
Presently we can provide a ranking based on Lim et al.’s process for deciding if an image is in focus: compute the sum of composition, brightness, and saturation weighted by density and median sharpness, as well as individual weights prioritizing composition over brightness and saturation.
The composition weighting could be improved in a number of ways. Perhaps the Manhattan distance from a power point is insufficient, and a more sophisticated model of the rule of thirds is required. Another possible direction is incorporating the idea of balanced masses as discussed in liu-et-al; note that within our presented approach, we could simplify that scheme by considering each block a unit of mass based on its sharpness value.
Lim et al.’s sky hue ratio is questionable and it remains to be seen if it is indeed an effective metric for our purposes.
lim-yen-wu Suk Hwan Lim, Jonathan Yen, Peng Wu, *Detection of Out-Of-Focus Digital Photographs*, HPL-2005-14, 2005.
tong-li-zhang-zhang Hanghang Tong, Mingjing Li, Hongjiang Zhang, Changshui Zhang, *Blur detection for digital images using wavelet transform*, Proceedings of the IEEE International Conference on Multimedia & Expo, 2004.
marichal-ma-zhang Xavier Marichal, Wei-Ying Ma, Hongjiang Zhang, *Blur determination in the compressed domain using DCT information*, Proceedings of the IEEE ICIP, pp.386-390, 1999.
liu-et-al Ligang Liu, Renjie Chen, Lior Wolf, Daniel Cohen-Or, *Optimizing Photo Composition*, Computer Graphics Forum, 29: 469-478, 2010.
ramakrishnan-thesis Harish Ramakrishnan, *Detection and Estimation of Image Blur*, Master’s Thesis, 2010.
shaked-tastl Dored Shaked, Ingeborg Tastl, *Sharpness Measure: Towards Automatic Image Enhancement*, HPL-2004-84, 2004.
marziliano-et-al Marziliano, P.; Dufaux, F.; Winkler, S.; Ebrahimi, T; *A no-reference perceptual blur metric*, Proceedings of International Conference on Image Processing, 2002.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent