Xinghui Dong

The Importance of Long-Range Interactions to Texture Similarity

We have tested 51 sets of texture features for estimating the perceptual similarity between textures. Our results show that these computational features only agree with human judgments at an average rate of 57.76%. In a second experiment we show that the agreement rates, between humans and computational features, increase when humans are not allowed to use long-range interactions beyond 19×19 pixels. We believe that this experiment provides evidence that humans exploit long-range interactions which are not normally available to computational features.

How Well Do Computational Features Perceptually Rank Textures? A Comparative Evaluation

Inspired by studies [4, 23, 40] which compared rankings obtained by search engines and human observers, in this paper we compare texture rankings derived by 51 sets of computational features against perceptual texture rankings obtained from a free-grouping experiment with 30 human observers, using a unify evaluation framework. Experimental results show that the MRSAR [37], VZNEIGHBORHOOD [62], LBPHF [2] and LBPBASIC [3] feature sets perform better than their counterparts. However, none of those feature sets are ideal. The best average G and M measures (measures of ranking accuracy from 0 to 1) [15, 5] obtained are 0.36 and 0.25 respectively. We suggest that this poor performance may be due to the small local neighborhood used to calculate higher-order features which cannot capture the long-range interactions that humans have been shown to exploit [14, 16, 49, 56].

Texture Similarity Estimation Using Contours

In a study of 51 computational features sets Dong et al. [1] showed that none of these managed to estimate texture similarity well and, coincidently, none of these computed higher order statistics (HOS) over large regions (that is larger than 19×19 pixels). Yet it is well-known that the human visual system is extremely adept at extracting long-range aperiodic (and periodic) “contour” characteristics from images [5, 6]. It is our hypothesis that HOS computed over larger spatial extent in the form of contour data are important for estimating perceptual texture similarity. However, to the authors’ knowledge the use of contour data (rather than edge data) has not been proposed before as the basis for a set of feature vectors. We provide results of an experiment with 334 textures that shows that contour data is more important than local image patches, or 2nd-order global data, to human observers. We also propose a contour-based feature set that exploits the long-range HOS encoded in the spatial distribution and orientation of contour segments. We compare it against the 51 feature sets tested by Dong et al. [1, 2] and another contour model derived from shape recognition. The results show that the proposed method outperforms all the other feature sets in a pairs-of-pairs task and all but two feature sets in a ranking task. We attribute this promising performance to the fact that this new feature set encodes long-range HOS.

Segmentation of Chinese Postal Envelope Images for Address Block Location

In this paper, we propose a simple segmentation approach for camera-captured Chinese envelope images. We first apply a moving-window thresholding algorithm, which is less curvature-biased and less sensitive to noise than other local thresholding methods, to generate binary images. Then the skew images are corrected by using a skew detection and correction algorithm. In the following stage rectangular frames on the envelopes containing postcode are removed by using opening operators in mathematical morphology. Finally, a post-processing procedure is used to remove remaining thin lines. In this stage, connected components are labeled. We test 800 camera-captured envelope images in our experiments, including handwritten and machine-printed envelopes. For almost all of these images, the proposed approach can accurately separate the address block, stamp and postmark from the background.

Texture synthesis based on multiple seed-blocks and support vector machines

We introduce a new method for texture synthesis based on multiple seed-blocks and support vector machines (SVM). First the sample texture is used to train the SVM model with class labels assigned to gray levels. During the synthesis process, each time we generate one patch in the left-to-right order in the result texture. The size of each patch is smaller than that of the sample, and we search a seed-block in the already generated patches to ensure the synthesized patch has similar texture characteristics as the sample. Support vector machines are used to generate pixel values within each patch. The advantage of using SVM is that the sample is not required during the synthesis stage since it has been modeled by a linear model. Unlike previous work in, which can only synthesize highly structured texture, the proposed method can successfully synthesize both random and structured textures. It is also extended to synthesize 3D surface texture or Bidirectional Texture Functions (BTF).

Relations between Surface Gradient Maps in Frequency Domain and Application in Diffuse Component Detection

This paper examines the relationships between surface gradient maps in frequency domain and applies the relationships to detect the diffuse component using a set of images. In our method, we first assume the diffuse component can be expressed by the Lambertian model, and then we obtain a set of surface gradient pairs by selecting different pixel values produced by different illumination directions. We use the relation between surface gradient maps as a constraint to select the optimized surface gradient maps pair, which can be further used to obtain diffuse parameters. Experimental results based on real surface textures are presented.

Texture Segmentation Based on Probabilistic Index Maps

The Probabilistic Index Map (PIM) model was originally proposed for video processing to extract background of video frames. In this paper, we introduce the PIM model for texture segmentation. We first extract texture features based on Laws and Gabor filters respectively. Then we present a Fuzzy K-Means method to generate the index map and palette, and use the PIM model to improve the segmentation accuracy. Based on the comparison of experimental results produced using different features and different resolutions, we show the proposed method is effective for texture segmentation.

Enteromorpha detection in aerial images using support vector machines

In this paper, we introduce a simple approach for detecting enteromorpha based on statistical learning of image features using support vector machines (SVM). The approach first classifies an enteromorpha image into two classes: enteromorpha and background. Then it extracts features from those two classes and uses them for training the SVM model. Finally, the predicting process is carried out in a pixel by pixel manner using the learned model. The model uses saturation in NTSC color space or filtered images by Gabor filter as the input features while the output class label is treated as 1 or 2 (enteromorpha or background), which is assigned to the location that is being predicted. In fact, this application is only a two-class pattern classification problem. Experimental results show that the method can be effectively applied to detecting enteromorpha in aerial images.

Perceptually Motivated Image Features Using Contours

Dong et al. examined the ability of 51 computational feature sets to estimate human perceptual texture similarity; however, none performed well for this task. While it is well-known that the human visual system is extremely adept at exploiting longer-range aperiodic (and periodic) “contour” characteristics in images, none of the investigated feature sets exploit higher order statistics (HOS) over larger image regions (>19×19 pixels). We, therefore, hypothesise that long-range HOS, in the form of contour data, are useful for perceptual texture similarity estimation. We present the results of a psychophysical experiment that shows that contour data are more important, than local image patches, or global second-order data, to human observers for this task. Inspired by this finding, we propose a set of perceptually motivated image features (PMIF) that encode the long-range HOS computed from spatial and angular distributions of contour segments. We use two perceptual texture similarity estimation tasks to compare PMIF against the 51 feature sets referred to above and four commonly used contour representations. This new feature set is also examined in the context of two additional tasks: sketch-based image retrieval and natural scene recognition. The results show that the proposed feature set performs better, or at least comparably to, all the other feature sets. We attribute this promising performance to the fact that the proposed feature set exploits both short-range and long-range HOS.

Perceptual texture retrieval using spatial distributions of textons (SDoT)

It has been shown that the spatial information of local image characteristics is important to human perception and computational features. Inspired by these studies, we propose a set of new computational texture features based on the spatial distributions of textons (SDoT). First, gradient magnitude and gradient direction spectra are computed from a texture image. Second, the multiple gradient spectra simultaneous autoregressive (MGSSAR) models are estimated for each image. Both model coefficients and the variance of the model estimation error jointly construct a local feature space. Third, &-means is used to learn textons from the local features. All textons learned from a texture database are combined into a dictionary. Fourth, vector quantization is utilized to map a texture from the local feature space into the texton space. Finally, an aura matrix is computed from the texton map of each texture in order to encode the spatial distributions of the textons. The results of a perceptual texture retrieval experiment show that the proposed feature set performs more consistently with human observers than 56 existing feature sets. We attribute this to the fact that the proposed feature set encodes the spatial information of textons.