Face Sketch Synthesis Style Similarity:
A New Structure Co-occurrence Texture Measure
Existing face sketch synthesis (FSS) similarity measures are sensitive to slight image degradation (e.g., noise, blur). However, human perception of the similarity of two sketches will consider both structure and texture as essential factors and is not sensitive to slight (“pixel-level”) mismatches. Consequently, the use of existing similarity measures can lead to better algorithms receiving a lower score than worse algorithms. This unreliable evaluation has significantly hindered the development of the FSS field. To solve this problem, we propose a novel and robust style similarity measure called Scoot-measure (Structure CO-Occurrence Texture Measure), which simultaneously evaluates “block-level” spatial structure and co-occurrence texture statistics. In addition, we further propose 4 new meta-measures and create 2 new datasets to perform a comprehensive evaluation of several widely-used FSS measures on two large databases. Experimental results demonstrate that our measure not only provides a reliable evaluation but also achieves significantly improved performance. Specifically, the study indicated a higher degree (78.8%) of correlation between our measure and human judgment than the best prior measure (58.6%). Our code will be made available.
Please take a look at Fig. 1 in which you can see several synthesized face sketches (GAN (Isola et al., 2017), FCN (Zhang et al., 2015b), MRF (Wang and Tang, 2009)). Which one do you think is closer to the ground-truth (GT) sketch drawn by the artist? While this comparison task seems trivial for humans, to date the most common measures (e.g., FSIM (Zhang et al., 2011b), SSIM (Wang et al., 2004), VIF (Sheikh and Bovik, 2006), and GMSD (Xue et al., 2014)) cannot reliably rank the algorithms according to their similarity to humans. This is the problem that we address in this paper, and to solve it we propose a novel measure that does much better than existing measures in terms of ranking algorithm syntheses compared to the GT sketch.
Face sketch synthesis (FSS) has been gaining popularity in the design community (e.g., face artistic styles synthesis) (Berger et al., 2013; Wang et al., 2014) and is being used for digital entertainment (Gao et al., 2008; Yu et al., 2012; Zhang et al., 2017b), and multimedia surveillance law enforcement (e.g., sketch based face recognition) (Tang and Wang, 2004; Best-Rowden et al., 2014). In such applications, the comparison of a synthesized sketch against a GT sketch is crucial in evaluating the quality of a face sketch synthesis algorithm.
To date, the most widely-used (Wang et al., 2018; Wang et al., 2017c; Peng et al., 2016; Zhang et al., 2017a, 2015a, 2016a; Wang et al., 2017b; Zhang et al., 2016b; Wang et al., 2014) evaluation measures are that of Structural SIMilarity (SSIM) (Wang et al., 2004), Visual Information Fidelity (VIF) (Sheikh and Bovik, 2006), Feature SIMilarity (FSIM) (Zhang et al., 2011b), and Gradient Magnitude Similarity Deviation (GMSD) (Xue et al., 2014). These measures were originally designed for pixel-level Image Quality Assessment (IQA) which aims to detect types of image degradation such as Gaussian blur, jpeg, and jpeg 2000 compression. Therefore these measures should be sensitive to slight (e.g., pixel-level) change in image. Psychophysics (Zucker et al., 1989) and prior works (e.g., style for lines of drawing) (Grabli et al., 2004; Zucker, 2005; Freeman et al., 2003) indicate that style is an essential factor in sketches. Note that human perception of the style similarity of two sketches will consider both of structure and texture (Wang et al., 2016) as the essential factor rather than being sensitive to “pixel-level” mismatches. As a consequence, current FSS measures often provide unreliable evaluation due to the different nature of their task. Consequently, a better algorithm may receive a lower score than a worse algorithm (see Fig. 1). This unreliable evaluation has significantly hindered the development of the FSS field. However, designing a reliable style similarity measure is difficult. Ideally, such a measure should:
hold good holistic-content capture capability, and should for example score a complete state-of-the-art (SOTA) sketch higher than the results of only preserving light strokes (Sec. 4.3);
obtain a high correlation to human perception so that the sketch can be directly used in various subjective applications, especially for law enforcement and entertainment (Sec. 4.4).
As far as we know, no previous measures can meet all these goals simultaneously and provide reliable evaluation results.
To this end, we propose a new and robust style similarity measure called Scoot-measure (Structure CO-Occurrence Texture Measure), which simultaneously evaluates “block-level” spatial structure and co-occurrence texture statistics. We experimentally demonstrate that our measure offers a reliable evaluation and achieves significantly improved performance. Specially, the experiment indicated a higher degree (78.8%) of correlation between our measure and the human judgment than the best prior measure (58.6%). Our main contributions are:
1) We propose a simple and robust style similarity measure for face sketch synthesis that provides a unified evaluation capturing both structure and texture. We experimentally show that our measure is significantly better than most of the current widely used measures and some alternatives including GLRLM (Galloway, 1974), Gabor (Gabor, 1946), Canny (Canny, 1987), and Sobel (Sobel, 1990).
2) To assess the quality of style similarity measures, we further propose 4 new meta-measures based on 3 hypotheses and build two new human ranked face sketches datasets. These two datasets contain 676, and 1888 synthesized face sketches, respectively. We use the two datasets to examine the ranking consistency between current measures and human judgment.
2. Related Works
As discussed in the introduction, a reliable similarity evaluation measure in FSS does not exist (Wang et al., 2016). Previous researchers just used the standard measures in IQA to evaluate the similarity of FSS.
Wang et al. (Wang et al., 2004) proposed the SSIM index. Since the structural information reflects the object structure features in the scenes, SSIM computes the structure similarity, the luminance and contrast comparison using a sliding window on the local patch.
Sheikh and Bovik (Sheikh and Bovik, 2006) proposed the VIF measure which evaluates the image quality by quantifying two kinds of information. One is obtained via the human visual system (HVS) channel, with the input ground truth and the output reference image information. The other is obtained via the distortion channel, called distortion information. And the result is the ratio of these two kinds of information.
Physiological and psychological studies found that the perceived features of human vision show great consistency with the phase consistency of Fourier series at different frequencies. Therefore, Zhang et al. (Zhang et al., 2011b) choose the phase congruency as the primary feature. Since phase congruency is not affected by the contrast, which is a key point of image quality, the image gradient magnitude is chosen as the second feature. Then they proposed a low-level feature similarity measure called FSIM.
Recently, Xue et al. (Xue et al., 2014) devised a simple measure named gradient magnitude similarity deviation (GMSD), where the pixel-wise gradient magnitude similarity is utilized to obtain image local quality, and the standard deviation of the overall GMS map is calculated as the final image quality index. Their measure achieves the SOTA performance compared with the other measures in IQA.
However, all the above-mentioned measures were originally designed for IQA which focuses on pixel-level image degradation, and so they are sensitive to slight resizing (see Fig. 7) of the image. Although some of them consider structure similarity, the structure is based on pixel-wise rather than pair-wise measurements. By observing sketches we found that the pair-wise co-occurrence texture is more suitable to capture the style similarity (also see our discussion in Sec. 5) than the pixel-wise texture.
3. Scoot measure
In this section, we introduce our novel measure to evaluate the style similarity for FSS. The important property of our measure is that it captures the pair-wise co-occurrence texture statistics in the “block-level” spatial structure. As a result, our Scoot-measure works better than the current widely-used measures. The framework is shown in Fig. 2.
3.1. Co-occurrence Texture
To uncover the secrets of face sketch and explore quantifiable factors in style, we observed the basic principles of the sketch and noted that the “graphite pencil grades” and the “pencil’s strokes” are the two fundamental elements in the sketch.
Graphite Pencil Grades. In the European system, “H” & “B” stand for “hard” & “soft” pencil, respectively. Fig. 4 illustrates the grade of graphite pencil from 9H-9B. Sketch images are expressed through a limited medium (graphite pencil) which provides no color. Illustrator Sylwia Bomba (2015) said that “if you put your hand closer to the end of the pencil, you have darker markings. Gripping further up the pencil will result in lighter markings.” In addition, after a long period of practice, artists will form their own fixed pressure style. In other words, the marking of the stroke can be varied by varying the pressure on the tip (Fig. 3 (c - d)). It is important to note that different pressures on the tip will result in various types of marking which is one of the quantifiable factors called gray tone.
Pencil Strokes. Because all of the sketches are generated by moving a tip on the paper. Hence, different paths of the tip along the paper will create various stroke shapes. One example is shown in Fig. 5; different spatial distributions of the stroke have produced various textures (e.g., sparse or dense). Thus, the stroke tone is another quantifiable factor.
In this section, we analyse the two quantifiable factors: gray tone and stroke tone.
Gray tone. To reduce the effect of slight noise and over-sensitivity to imperceptible gray tone gradient changes in sketches, intensity quantization can be introduced during the evaluation of gray tone similarity. Inspired by previous works (Clausi, 2002), we can quantize the input sketch to different grades to reduce the number of intensities to be considered: .
A typical example of such quantization is shown in Fig. 6. Human beings will consistently rank (b) higher than (c) before and after quantizing the input sketches when evaluating the style similarity. Although, the quantization technology may introduces artifacts. However, our experiments (Sec. 5 of the quantized Scoot measure) show that this process can reduce sensitivity to minor intensity variations, thus achieving the best overall performance, and also reducing the computational complexity.
Stroke tone. Although we can efficiently capture the gray tone by quantizing the input sketch to grades we notice that the stroke tone (e.g., dense & sparse) is also important to formulate the sketch style. Stroke tone and gray tone are not independent concepts; rather, they bear an inextricable relationship to one another. Both gray tone and stroke tone are innate properties of the sketches. The gray tone is based on the varying strokes of gray-scale in a sketch image, while the stroke tone is concerned with the spatial (statistical) distribution of gray tones.
This is demonstrated in Fig. 6, which presents the ground-truth sketch (a) and two other synthesis results (b) & (c). These two images are the results of SOTA face synthesis algorithms. Intuitively, result (b) is better than (c), since (b) preserves the texture of the hair and details in the face while (c) presents a smooth result and has lost much of the sketch style. In other words, (b) and (c) have different stroke patterns. The co-occurrence stroke pattern of (b) has a higher similarity than the sketch in (c).
After confirming the two quantifiable factors, we start to describe our implementation details. To simultaneously extract statistics about the “stroke tone” and their relationship to the surrounding “gray tone”, we need to characterize their spatial interrelationships. Theoretically, the sketches are quite close to textures. As a consequence, our goal is to capture the co-occurrence texture in the sketch. Such co-occurrence properties help to capture the patterns of the face sketch (especially in the hair region). Similar to the general technology (Haralick et al., 1973), we built the co-occurrence gray tones in a matrix at distance to encode the dependency of the gray tones. Specifically, this matrix is defined as:
where and denote the gray value; and are the spatial positions in the given quantized sketch ; denotes the gray value of at position ; and are the width and height of the sketch , respectively.
To extract the style features we tested the three most widely-used (Haralick, 1979) statistics: Homogeneity, Contrast, and Energy.
Homogeneity reflects how much the texture changes in local regions, it will be high if the gray tone of each pixel pair is similar. The homogeneity is defined as:
Contrast represents the difference between a pixel in and its neighbor summed over the whole sketch. This reflects that a low-contrast sketch is not characterized by low gray tones but rather by low spatial frequencies. The contrast is highly correlated with spatial frequencies. The contrast equals 0 for a constant tone sketch.
Energy measures textural uniformity. When only similar gray tones of pixels occur in an sketch () patch, a few elements in will be close to 1, while others will be close to 0. Energy will reach the maximum if there is only one gray tone in a sketch () patch. Thus, high energy corresponds to the sketch’s gray tone distribution having either a periodic or constant form.
CE (Contrast and Energy) Feature. Our experiments (see Tab. 1) show that the combination of two statistics Contrast and Energy achieves the best average performance. In this work, we concatenate the two statistics as the CE feature to represent the style of the input sketch .
3.2. Spatial Structure
Note that the matrix (Eq. 1) we built is based at the “image-level”. However, the weakness of this approach is that it only captures the global statistics, and the structure of the sketch is ignored. Therefore, for instance, it would be difficult to distinguish men and women with the same hairstyle.
To holistically represent the spatial structure, we follow the spatial envelope method (Oliva and Torralba, 2001) to capture the statistics from the “block-level” spatial structure in the sketch. Firstly, we divide the whole sketch image into a grid of blocks111 We set to achieve the best performance in our all experiments. before extracting the CE feature. Our experiments demonstrate that the process can help to derive content information. Second, we compute the co-occurrence matrix for all blocks and normalize each matrix such that the sum of its components is 1. Finally, we concatenate the statistics of all the blocks into a dimension vector
Note that each of the statistics is only based on a single direction (e.g., , that is d = (0, 1)), since the direction of the spatial distribution is also very important to capture the style such as “hair direction”, “the direction of shadowing strokes”. To exploit this observation for efficiently extracting the stroke direction style, we compute the average feature of four orientations222 Due to the symmetry of the co-occurrence matrix , the statistical features in 4 orientations are actually equivalent to 8 neighbors direction at 1 distance. Empirically, we set 4 orientations to achieve the best performance. vectors to capture more directional information:
where denotes the th direction and the average CE feature .
3.3. Scoot measure
Having obtained the style feature vectors of the GT sketch and synthesis sketch , we use the Euclidean distance between their feature vectors to evaluate their style similarity. Thus, our style Scoot-measure333 In all of our experiment, we set grades to achieve the best performance. can be defined as:
where denotes -norms. denote the quantized , respectively. corresponds to identical style. Using this measure to evaluate the three synthesized face sketches in Fig. 1, we can correctly rank the sketches consistent with the human ranking.
In this section, we compare our Scoot-measure with 8 measures including 4 popular measures (FSIM, SSIM, VIF, GMSD) and 4 alternative baseline measures (GLRLM (Galloway, 1974), Gabor (Gabor, 1946), Canny (Canny, 1987), Sobel (Sobel, 1990)) which related with texture or edge. For the four baseline measures, we only replace our CE feature (describe in Sec. 3.1.3) with these features.
Meta-measure. To test the quality of our measure, we adopt the meta-measure methodology. The meta-measures consist of assuming some plausible hypotheses about the results and assessing how well each measure reflects these hypotheses (Pont-Tuset and Marques, 2013; Fan et al., 2017). We design 4 novel meta-measures based on our plausible hypotheses (Sec. 1) and the experiment results are summarized in Tab. 2.
Dataset. All of the meta-measures were performed on 2 widely used databases: CUFS (Wang and Tang, 2009) and CUFSF (Zhang et al., 2011a). The CUFS database includes 606 face photo-sketch pairs consisting of 3 sub-sections: CUHK Student (188 pairs), XM2VTS (295 pairs) and Purdue AR (123 pairs). The CUFSF database contains 1194 pairs of persons from the FERET database (Phillips et al., 2000). There is one face photo and one GT sketch drawn by the artist for each person in both CUFS and CUFSF databases. Following (Wang et al., 2018), we use 338 pairs (CUFS) and 944 pairs (CUFSF) test set images to conduct our experiments.
Sketch Synthesis Results. Sketch synthesis results were generated for each test image using 10 SOTA models (FCN (Zhang et al., 2015b), GAN (Isola et al., 2017), LLE (Liu et al., 2005), LR (Wang et al., 2017c), MRF (Wang and Tang, 2009), MWF (Zhou et al., 2012), RSLCR (Wang et al., 2018), SSD (Song et al., 2014), DGFL (Zhu et al., 2017), BFSS (Wang et al., 2017a)).
4.1. Meta-measure 1: Stability to Slight Resizing
The first meta-measure specifies that the rankings of synthetic results should not change much with slight changes in the GT sketch. Therefore, we perform a slight 5 pixels downsizing of the GT by using nearest-neighbor interpolation.
Fig. 7 gives an example. The hair of the GT sketch in (b) drawn by the artist has a slight size discrepancy compared to the photo (a). We observe that about 5 pixels deviation (Fig. 7 c) in the boundary is common. While the two sketches (e) & (f) are almost identical, commonly-used measures including SSIM, VIF, and GMSD switched the ranking of the two synthesized results (g, h) when using (e) or (f). However, our measure consistently ranked (g) higher than (h).
Here, we applied the (Best and Roberts, 1975) measure to test the measure ranking stability before and after the GT downsizing was performed. Fig. 8 and Tab. 2 show the results: the lower the result is, the more stable an evaluation measure is to slightly downsizing. We can see a significant ( 88% and 92%) improvement over the existing SSIM, FSIM, GMSD, and VIF measures in both the CUFS and CUFSF databases. These improvements are mainly because our evaluation measure considers “block-level” features rather than “pixel-level”.
4.2. Meta-measure 2: Rotation Sensitivity
In real-wold situations, sketches drawn by artists may also have slight rotations compared to the original photographs. Thus, our second meta-measure verifies the sensitivity of GT rotation for the evaluation measure. We did a slight counter-clockwise rotation () for each GT sketch. Fig. 9 shows an example. When the GT (a) is switched to the slightly rotated GT (b), the ranking results should not change much.
We got the ranking results for each measure by using GT sketches and slightly rotated GT sketches separately and applied the same measure () as meta-measure 1 mentioned to evaluate the rotation sensitivity. The sensitivity results are shown in Fig. 10, Tab. 2. Thanks to our use of “block-level” statistics, our measure again significantly outperforms the current measures over the CUFS and CUFSF databases.
4.3. Meta-measure 3: Content Capture Capability
The third meta-measure describes that a good evaluation measure should not only evaluate the style similarity but also can capture the content similarity. Fig. 11 presents an example. We expect that an evaluation measure should prefer the SOTA synthesized result over the light strokes444 To test the third meta-measure, we use a simple threshold of gray scale (e.g., 170) to separate the sketch (Fig. 11 GT) into darker strokes & lighter strokes. The lighter strokes image loses the main texture features of the face (e.g., hair, eye, beard), resulting in an incomplete sketch. results. To our surprise, only our measure gives the correct order.
We use the mean score of 10 SOTA synthesis algorithms (FCN, GAN, LLE, LR, MRF, MWF, RSLCR, SSD, DGFL, BFSS). The mean score is robust to situations in which a certain model generates a poor result. We recorded the number of times the mean score of SOTA synthesized algorithms is higher than a light stroke’s score. The results are shown in Fig. 12 & Tab. 2. In the CUFS database we can see a great improvement over the most widely-used measures. A slight improvement is also achieved for the CUFSF database.
The reason why other measures fail in this meta-measure is shown in Fig. 11. In terms of pixel-level matching, it is obvious that the regions where dark strokes are removed are totally different from the corresponding parts in (a). But at other positions the pixels are identical to the GT. Previous measures only consider “pixel-level” matching, and will rank the light strokes sketch higher. However, the synthesized sketch (b) is better than the light one (c) in terms of both style and content.
4.4. Meta-measure 4: Human Judgment
The fourth meta-measure specifies that the ranking result according to an evaluation measure should agree with the human ranking. As far as we know, there is no such human ranking sketch database publicly available.
Source images. As mentioned in Sec. 4, our source sketch synthesized results are from the two widely-used datasets: CUFS and CUFSF and 10 SOTA sketch synthesis algorithms. Thus, we have 3380 (10 338) and 9440 (10 944) synthesized results for CUFS and CUFSF, respectively.
Human judgment. For each photo of the CUFS and CUFSF databases, we have 10 sketches synthesized by 10 SOTA algorithms. Due to the similarity of the synthesized results of some algorithms, it is difficult for viewers to judge the ranking. Following (Wang et al., 2016), we trained the viewers to assess synthesized sketch quality based on two criterion: texture similarity and content similarity. To minimize the ambiguity of human ranking, we asked 5 viewers to select 2 synthesized sketches for each photo through the following 3 stage processes:
i) We let the first group of viewers select 4 out of 10 sketches for each photo. The 4 sketches should consist of two good and two bad ones. Thus we are left with 1352 (4 338) and 3776 (4 944) sketches for CUFS and CUFSF, respectively.
ii) For the 4 sketches in each photo, the second group of viewers is further asked to select 3 sketches for which they can rank them easily. Based on the voting results of viewers, we pick out the 3 most selected sketches.
iii) The last group of viewers are asked to pick out a pair of sketches that are most obvious to rank. Note that we have 5 volunteers involved in the whole process for cross-checking the ranking. Finally, we create two new human ranked datasets: CUFS-Ranked and CUFSF-Ranked (see the supplementary for all images in these two datasets) 555They include 776 & 1888 human ranked images, respectively. The two datasets will be made publicly available.. Fig. 13 shows the human ranked examples.
Here, we examined the the consistency between human ranking and evaluation measure ranking. Results are given in Fig. 14, which shows a significant (about 26.3%) improvement over the best prior measure in CUFS. This improvement is due to our consideration of style similarity which human perception considers as an essential factor when evaluating sketches.
In this section, we will discuss the combination of statistics in Sec. 5.1 and other alternative features in Sec. 5.2. These discussions indicate that the pair-wise “co-occurrence” properties are quite important for the style similarity evaluation for FSS.
5.1. Statistical Feature Selection
According to Sec. 3, we considered three widely-used features of co-occurrence matrices: homogeneity (), contrast () and energy (). In order to achieve the best performance, we need to explore the best combination of these statistical features. We have applied our four meta-measures to test the performance of our measure using each single feature, each feature pair and the combination of all three features.
The results are shown in Tab. 1. All possibilities () perform well in MM4 (human judgment). and are insensitive to resizing (MM1) and rotation (MM2), while they are not good at content capture (MM3). is the opposite compared to and . Thus, using a single feature is not good. The results of combining two features show that if is combined with , the sensitivity to resizing and rotating will still be high, while partially overcoming the weakness of . The performance of shows no improvement compared to the combination of . Previous work in (Baraldi and Parmiggiani, 1995) also found energy and contrast to be the most efficient features for discriminating textural patterns. Therefore, we choose contrast and energy as our final combination “CE” feature to extract the style features.
|CUFS (3380 synthesized images)||CUFSF (9440 synthesized images)|
|SSIM (Wang et al., 2004)||0.162||0.086||81.4%||37.3%||0.073||0.074||97.4%||36.8%|
|FSIM (Zhang et al., 2011b)||0.268||0.123||14.2%||50.0%||0.151||0.058||32.4%||37.5%|
|VIF (Sheikh and Bovik, 2006)||0.322||0.236||43.5%||44.1%||0.111||0.150||22.2%||52.8%|
|GMSD (Xue et al., 2014)||0.417||0.210||21.9%||42.6%||0.259||0.132||63.6%||58.6%|
5.2. Alternative Features
As described in Sec. 3, sketches are quite close to textures. There are many other texture & edge-based features (e.g., GLRLM (Galloway, 1974), Gabor (Gabor, 1946), Canny (Canny, 1987), Sobel (Sobel, 1990)). Here, we select the most wide-used features as candidate alternatives to replace our “CE” feature. For GLRLM, we select all five statistics mentioned in the original version. Results are shown in Tab. 3. Gabor and GLRLM are texture features, while the other two are edge-based. With all the texture features our measure provides a high consistency with human ranking (MM4). However, GLRLM does perform well according to MM1 & 2 & 3. Gabor is reasonable in terms of MM1 & 2, but fails in MM3. For edge-based features, Canny fails according to all meta-measures. Sobel is very stable to slight resizing (MM1) or rotating (MM2), but fails to capture content (MM3) and is not consistent with human judgment (MM4).
New insight. Note that the alternative features do not consider the pair-wise “co-occurrence” texture properties (see Fig. 15), only our measure considers this cue and provides the best performance compared to the other features. Tab. 3 indicates that the pair-wise co-occurrence texture is suitable for FSS evaluation.
In this paper, we summarized the widely-used image quality measures for face sketch synthesis evaluation based on “pixel-level” degradation, and enumerated their limitations. Then we focused on the basic principles of sketching and proposed the use of structure co-occurrence texture as an alternative motivating principle for the design of face sketch style similarity measures. To amend the flaws of existing measures, we proposed the novel Scoot-measure which simultaneously captures both “block-level” spatial structure and co-occurrence texture. Also, we concluded that the pair-wise co-occurrence properties are more important than pixel-level properties for style similarity evaluation of face sketch synthesis images.
In addition, extensive experiments with 4 new meta-measures show that the proposed measure provides a reliable evaluation and achieves the best performance. Finally, we created two new human ranked face sketch datasets (containing 776 images & 1888 images) to examine the correlation between evaluation measures and human judgment. The study indicated a higher degree (78.8%) of correlation between our measure and human judgment. The proposed Scoot-measure is motivated from substantially different design principles, and we see it as complementary to the existing approaches.
All in all, our measure provides new insights into face sketch synthesis evaluation where current measures fail to truly examine the pros and cons of synthesis algorithms. We encourage the community to consider this measure in future model evaluations and comparisons.
Future works. Though our measure is only tested on face sketch datasets, it is not limited to evaluating face sketch synthesis. We will also investigate the potential to use the measure to describe styles in sketches and investigate new face sketch synthesis models. To promote the development of this field, our code and dataset will be made publicly available.
- Baraldi and Parmiggiani (1995) Andrea Baraldi and Flavio Parmiggiani. 1995. An investigation of the textural characteristics associated with gray level cooccurrence matrix statistical parameters. IEEE T Geosci. Remote. 33, 2 (1995), 293–304.
- Berger et al. (2013) Itamar Berger, Ariel Shamir, Moshe Mahler, Elizabeth Carter, and Jessica Hodgins. 2013. Style and abstraction in portrait sketching. ACM TOG 32, 4 (2013), 55.
- Best and Roberts (1975) DJ Best and DE Roberts. 1975. Algorithm AS 89: the upper tail probabilities of Spearman’s rho. J R STAT SOC C-APPL 24, 3 (1975), 377–379.
- Best-Rowden et al. (2014) Lacey Best-Rowden, Hu Han, Charles Otto, Brendan F Klare, and Anil K Jain. 2014. Unconstrained face recognition: Identifying a person of interest from a media collection. IEEE Transactions on Information Forensics and Security 9, 12 (2014), 2144–2157.
- Canny (1987) John Canny. 1987. A computational approach to edge detection. In Readings in Computer Vision. Elsevier, 184–203.
- Clausi (2002) David A Clausi. 2002. An analysis of co-occurrence texture statistics as a function of grey level quantization. CAN J Remote. Sens. 28, 1 (2002), 45–62.
- Fan et al. (2017) Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. 2017. Structure-measure: A New Way to Evaluate Foreground Maps. ICCV (2017).
- Freeman et al. (2003) William T Freeman, Joshua B Tenenbaum, and Egon C Pasztor. 2003. Learning style translation for the lines of a drawing. ACM TOG 22, 1 (2003), 33–46.
- Gabor (1946) Dennis Gabor. 1946. Theory of communication. Part 1: The analysis of information. Journal of the Institution of Electrical Engineers-Part III: Radio and Communication Engineering 93, 26 (1946), 429–441.
- Galloway (1974) Mary M Galloway. 1974. Texture analysis using grey level run lengths. NASA STI/Recon Technical Report N 75 (1974).
- Gao et al. (2008) Xinbo Gao, Juanjuan Zhong, Dacheng Tao, and Xuelong Li. 2008. Local face sketch synthesis learning. Neurocomputing 71, 10-12 (2008), 1921–1930.
- Grabli et al. (2004) Stéphane Grabli, Emmanuel Turquin, Frédo Durand, and François X Sillion. 2004. Programmable style for NPR line drawing. Rendering Techniques (Eurographics Symposium on Rendering) (2004).
- Haralick (1979) Robert M Haralick. 1979. Statistical and structural approaches to texture. Proc. IEEE 67, 5 (1979), 786–804.
- Haralick et al. (1973) Robert M Haralick, Karthikeyan Shanmugam, et al. 1973. Textural features for image classification. IEEE Transactions on systems, man, and cybernetics 6 (1973), 610–621.
- Isola et al. (2017) Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. 2017. Image-To-Image Translation With Conditional Adversarial Networks. In CVPR. 1125–1134.
- Liu et al. (2005) Qingshan Liu, Xiaoou Tang, Hongliang Jin, Hanqing Lu, and Songde Ma. 2005. A nonlinear approach for face sketch synthesis and recognition. In CVPR, Vol. 1. IEEE, 1005–1010.
- Oliva and Torralba (2001) Aude Oliva and Antonio Torralba. 2001. Modeling the shape of the scene: A holistic representation of the spatial envelope. International journal of computer vision 42, 3 (2001), 145–175.
- Peng et al. (2016) Chunlei Peng, Xinbo Gao, Nannan Wang, Dacheng Tao, Xuelong Li, and Jie Li. 2016. Multiple representations-based face sketch–photo synthesis. IEEE TNNLS 27, 11 (2016), 2201–2215.
- Phillips et al. (2000) P Jonathon Phillips, Hyeonjoon Moon, Syed A Rizvi, and Patrick J Rauss. 2000. The FERET evaluation methodology for face-recognition algorithms. IEEE TPAMI 22, 10 (2000), 1090–1104.
- Pont-Tuset and Marques (2013) Jordi Pont-Tuset and Ferran Marques. 2013. Measures and meta-measures for the supervised evaluation of image segmentation. In CVPR. 2131–2138.
- Sheikh and Bovik (2006) Hamid R Sheikh and Alan C Bovik. 2006. Image information and visual quality. IEEE TIP 15, 2 (2006), 430–444.
- Sobel (1990) Irvin Sobel. 1990. An isotropic 3 3 image gradient operator. Machine vision for three-dimensional scenes (1990), 376–379.
- Song et al. (2014) Yibing Song, Linchao Bao, Qingxiong Yang, and Ming-Hsuan Yang. 2014. Real-time exemplar-based face sketch synthesis. In ECCV. Springer, 800–813.
- Sylwia et al. (2015) Bomba Sylwia, Cai Rovina, Croes Brun, Gerard Justin, and Lewis Marisa. 2015. Beginner’s Guide to Sketching. 3dtotal Publishing.
- Tang and Wang (2004) Xiaoou Tang and Xiaogang Wang. 2004. Face sketch recognition. IEEE Transactions on Circuits and Systems for Video Technology 14, 1 (2004), 50–57.
- Wang et al. (2018) Nannan Wang, Xinbo Gao, and Jie Li. 2018. Random sampling for fast face sketch synthesis. Pattern Recognition 76 (2018), 215–227.
- Wang et al. (2016) Nannan Wang, Xinbo Gao, Jie Li, Bin Song, and Zan Li. 2016. Evaluation on synthesized face sketches. Neurocomputing 214 (2016), 991–1000.
- Wang et al. (2017a) Nannan Wang, Xinbo Gao, Leiyu Sun, and Jie Li. 2017a. Bayesian face sketch synthesis. IEEE Transactions on Image Processing 26, 3 (2017), 1264–1274.
- Wang et al. (2014) Nannan Wang, Dacheng Tao, Xinbo Gao, Xuelong Li, and Jie Li. 2014. A comprehensive survey to face hallucination. IJCV 106, 1 (2014), 9–30.
- Wang et al. (2017b) Nannan Wang, Shengchuan Zhang, Xinbo Gao, Jie Li, Bin Song, and Zan Li. 2017b. Unified framework for face sketch synthesis. Signal Processing 130 (2017), 1–11.
- Wang et al. (2017c) Nannan Wang, Mingrui Zhu, Jie Li, Bin Song, and Zan Li. 2017c. Data-driven vs. model-driven: Fast face sketch synthesis. Neurocomputing (2017).
- Wang and Tang (2009) Xiaogang Wang and Xiaoou Tang. 2009. Face photo-sketch synthesis and recognition. IEEE TPAMI 31, 11 (2009), 1955–1967.
- Wang et al. (2004) Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity. IEEE TIP 13, 4 (2004), 600–612.
- Winkenbach and Salesin (1994) Georges Winkenbach and David H Salesin. 1994. Computer-generated pen-and-ink illustration. In ACM SIGGRAPH. 91–100.
- Xue et al. (2014) Wufeng Xue, Lei Zhang, Xuanqin Mou, and Alan C Bovik. 2014. Gradient magnitude similarity deviation: A highly efficient perceptual image quality index. IEEE Transactions on Image Processing 23, 2 (2014), 684–695.
- Yu et al. (2012) Jun Yu, Meng Wang, and Dacheng Tao. 2012. Semisupervised multiview distance metric learning for cartoon synthesis. IEEE Transactions on Image Processing 21, 11 (2012), 4636–4648.
- Zhang et al. (2015b) Liliang Zhang, Liang Lin, Xian Wu, Shengyong Ding, and Lei Zhang. 2015b. End-to-end photo-sketch generation via fully convolutional representation learning. In ICMR. ACM, 627–634.
- Zhang et al. (2011b) Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. 2011b. FSIM: A feature similarity index for image quality assessment. IEEE TIP 20, 8 (2011), 2378–2386.
- Zhang et al. (2017b) Mingjin Zhang, Jie Li, Nannan Wang, and Xinbo Gao. 2017b. Compositional Model-Based Sketch Generator in Facial Entertainment. IEEE Transactions on Cybernetics (2017).
- Zhang et al. (2016a) Shengchuan Zhang, Xinbo Gao, Nannan Wang, and Jie Li. 2016a. Robust face sketch style synthesis. IEEE Transactions on Image Processing 25, 1 (2016), 220–232.
- Zhang et al. (2017a) Shengchuan Zhang, Xinbo Gao, Nannan Wang, and Jie Li. 2017a. Face sketch synthesis from a single photo–sketch pair. IEEE Transactions on Circuits and Systems for Video Technology 27, 2 (2017), 275–287.
- Zhang et al. (2015a) Shengchuan Zhang, Xinbo Gao, Nannan Wang, Jie Li, and Mingjin Zhang. 2015a. Face sketch synthesis via sparse representation-based greedy search. IEEE transactions on image processing 24, 8 (2015), 2466–2477.
- Zhang et al. (2011a) Wei Zhang, Xiaogang Wang, and Xiaoou Tang. 2011a. Coupled information-theoretic encoding for face photo-sketch recognition. In CVPR. IEEE, 513–520.
- Zhang et al. (2016b) Yuqian Zhang, Nannan Wang, Shengchuan Zhang, Jie Li, and Xinbo Gao. 2016b. Fast face sketch synthesis via kd-tree search. In European Conference on Computer Vision. Springer, 64–77.
- Zhou et al. (2012) Hao Zhou, Zhanghui Kuang, and Kwan-Yee K Wong. 2012. Markov weight fields for face sketch synthesis. In CVPR. IEEE, 1091–1097.
- Zhu et al. (2017) Mingrui Zhu, Nannan Wang, Xinbo Gao, and Jie Li. 2017. Deep graphical feature learning for face sketch synthesis. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. AAAI Press, 3574–3580.
- Zucker (2005) Steven W Zucker. 2005. Which computation runs in visual cortical columns? 23 Problems in Systems Neuroscience (2005), 215.
- Zucker et al. (1989) Steven W Zucker, Allan Dobbins, and Lee Iverson. 1989. Two stages of curve detection suggest two styles of visual computation. Neural computation 1, 1 (1989), 68–81.