Research Article Saliency Detection via Fusing Color Contrast and Hash Fingerprint

. Saliency detection is a technique for automatically extracting regions of interest from the background and has been widely used in the computer vision field. This study proposes a simple and effective saliency detection method combining color contrast and hash fingerprint. In our solution, the input image is segmented into nonoverlapping superpixels, so as to perform the saliency detection at the region level to reduce computational complexity. A background optimization selection is used to construct an accurate background template. Based on this, a saliency map that highlights the whole salient region is obtained by estimating color contrast. Besides, another saliency map that enhances the salient region while restraining the background is also generated through hash fingerprint matching. Ultimately, the final saliency map can be obtained by fusing the two saliency maps. Comparing the performance with other methods, the proposed algorithm works better even in the presence of complex background or very large salient regions.


Introduction
Saliency detection can quickly screen important information so as to locate objects of interest in an image. It has been widely used in many applications, such as target detection [1], target recognition [2], image compression [3], and image retrieval [4]. Existing saliency detection methods can be roughly divided into top-down models and bottom-up models [5,6]. e top-down models generally utilize highlevel features or supervised learning methods to measure the saliency. For example, Liu et al. [7] proposed a binary saliency estimation model by training a conditional random eld to combine a set of novel features. Wang et al. [8] used a trained classi er called the auto-context model for salient region detection. Lu et al. [9] generated a set of salient seeds through learning and used the seeds to identify the salient regions. Dai et al. [10] extracted high-level semantic features through a region-based fully convolutional network so as to obtain the saliency map. Jiang et al. [11] used the supervised learning approach to map the regional feature vector to a saliency score and fused the saliency scores across multiple levels to yield the saliency map. Although these kinds of models achieve good detection performance, the calculations are comparatively expensive and time consuming.
Comparatively, the bottom-up models primarily exploit low-level cues such as texture, intensity, color, and edge to measure saliency.
us, this kind of method provides simpler and faster solutions and is more widely used in saliency detection. Itti et al. [12] presented a saliency map by combining three feature maps including color, intensity, and orientation at di erent scales. Achanta et al. [13] proposed a frequency-tuned approach to estimate the contrast di erence between the mean image feature vector and the corresponding image pixel vector value in the Gaussian blurred version. Zhai and Shah [14] calculated the saliency using the sum of the color distance between the pixel and all the other pixels in the image. Achanta and Süsstrunk [15] estimated saliency using the contrast between the color and luminance features of each pixel and its maximum symmetry surrounding area. ese methods have simple calculations but su er great di culty in handling images with complex backgrounds. To solve this problem, some algorithms introduced background prior and treated the image boundary region as the background for saliency detection. Wei et al. [16] proposed a saliency measure called geodesic saliency that uses boundary and connectivity priors to provide more clues for the saliency calculation. Yang et al. [17] determined the saliency by the similarity of the image elements (pixel or region) with foreground cues or background. However, these approaches naively select the image border as the background so that they may fail when salient regions locate along the image boundary. erefore, some research made further optimization in order to improve the detection performance. For instance, Zhu et al. [18] proposed a robust background measure based on boundary connectivity to constrain background regions and obtained salient regions using foreground cues. Besides, in our earlier work [19], an accurate background template was first constructed, and based on it the saliency map was obtained through image sparse representation and color features combination.
Aiming at the abovementioned problems, we propose a novel and effective salient detection algorithm to highlight salient regions while restraining the image background. First, simple linear iterative clustering (SLIC) [20] is utilized to segment the original image into superpixels, and the subsequent operations are carried out based on each superpixel, which helps to reduce computation complexity. en, the accurate background template is generated through background optimization, which can improve the accuracy of subsequent saliency detection. e first saliency map corresponding to the original image is obtained by calculating the color contrast between superpixels based on the background template. On the other hand, the original image is divided into blocks of equal size, and the hash fingerprint of each block is calculated based on discrete cosine transformation (DCT). us, the second saliency map is obtained by matching hash fingerprints of image blocks. Finally, the two saliency maps are fused to generate the final saliency map. e framework of the proposed algorithm is shown in Figure 1.
e main contributions of this study are summarized as follows. First, we design a background optimization scheme to ensure that salient regions can be completely detected even when they are situated at the boundary of the image. Second, we use hash fingerprint matching to rapidly extract salient regions and well suppress the background. ird, we propose a fusion strategy to produce the final saliency map. Comprehensive experiments demonstrate that the proposed method has desirable detection accuracy.

Background Optimization Selection.
To retain more structural information about the original image, we first utilize the SLIC algorithm to segment the image into nonoverlapping superpixels. Besides, the number of superpixels is much smaller than that of the original pixels, which helps increase the calculation speed. Most of the existing methods based on SLIC usually choose the edge superpixels as the background for saliency detection. However, in some cases, the salient regions usually extend to the edge of the image. If the edge superpixels of the image are directly labeled as the background, it might cause the subsequent saliency detection to be degraded.
To address this issue, an optimization scheme is designed to generate a refined background template. After the image is divided into superpixels, all outermost superpixels are chosen to build the candidate background template B, consisting of n superpixels. If a certain superpixel in B belongs to the salient region, there will be a big difference in color features between it and other superpixels. Moreover, the larger the color difference is, the more likely it belongs to the salient region. While considering the effect of the spatial distance between superpixels, for each superpixel in B, we calculate the spatially weighted sum of color distances between it and all other superpixels and then judge whether it belongs to the background. Specifically, for superpixel i, the spatially weighted sum is denoted by where d c (i, j) and d s (i, j) are, respectively, the color distance and the spatial distance between two superpixels i and j, σ is empirically set to 0.25. en the superpixels whose spatially weighted sum is greater than a specific threshold α will be eliminated from B to obtain an accurate background template. e threshold should be determined adaptively according to the input image, and it is defined as α � S Bmax −λ·D, where λ is a constant factor and set to 8 for the best performance, S Bmax and D are the largest value, and the variance of S B (i) i � 1. . .n , respectively.
is means that only certain outermost superpixels are selected to construct the final background template B opt . More details about background optimization selection are available in our earlier work [19].

Color Contrast-Based Saliency
Map. Naturally, the more a superpixel varies in color from its surrounding regions, the larger the saliency value of the superpixel. Also, the closer the spatial distance between the superpixel and its surrounding superpixels, the more the contribution of the color contrast between them to the saliency value. us, we use the color contrast weighted by the distances between a superpixel and other superpixels to measure the saliency value of the superpixel. Assume that the rest region of the image is denoted as F and contains m superpixels after removing superpixels belonging to B opt . For the superpixel i in F, the saliency value S F (i) can be formulated in a similar way to (1), the only difference is that superpixels used for calculation belong to F.
To further highlight the salient region, the saliency of each superpixel in F is enhanced by integrating the color contrast between the superpixel and the background template. Specifically, the modified saliency value of the superpixel i is denoted by where c i is the color vector of superpixel i in CIELab space, c a is the average color vector of all superpixels in B opt , and ||·|| nrm denotes the normalized Euclidian distance, that is the color contrast is normalized to [0, 1]. Figure 2 shows example results of the background optimization and the color contrast-based saliency map. It can be seen in the Figures 2(c) and 2(d), the background optimization selection can refine the background template. e outermost superpixels corresponding to the salient object are eliminated from the candidate background template. Besides, it is clear that the color contrast-based saliency map illustrated in Figure 2(e) can highlight the whole salient region well, but at the same time may not effectively suppress some background regions. erefore, it is essential to perform further operations to improve the detection results.

Hash Fingerprint-Based Saliency Map.
Background usually contains more low-frequency information compared to the salient region. e hash algorithm based on DCT can effectively make use of the low-frequency information in image blocks, and the hash fingerprints can measure the similarity between image blocks. erefore, hash fingerprint matching can be used to estimate the saliency of each image block and meanwhile better suppress the background.
In order to extract the hash fingerprint, the input image is first converted to a grayscale image and then divided into b nonoverlapping blocks of size 8 × 8. Note that the image should be resized by interpolation if its width and height are not integer multiples of 8. For each block, we can obtain the 8 × 8 coefficient matrix by applying a DCT. Next, we traverse the coefficient matrix using the mean of all coefficients as the threshold, meanwhile setting the coefficients less than the mean to zero and other coefficients to 1. Finally, we convert the 8 × 8 matrices made up of zeros and ones into a 64 × 1 flattened vector using a zigzag scan so as to obtain the hash fingerprint of the image block. Based on this, the saliency is represented as the spatially weighted sum of hamming distance. us, for image block i, the saliency value is calculated by where (x, y) and (u, v) are separately the coordinates of block i and another block j, h i and h j correspond to the hash fingerprints of the two blocks, and d (h i , h j ) represents the hamming distance between h i and h j . Examples of hash fingerprint-based saliency detection are illustrated in Figure 3 can be seen from Figure 3(c) that the detection results can effectively remove most of the background. Nevertheless, there are two problems with this solution. e first is the outline of the saliency map is not too clear due to blocking artifacts, and this is because that the image is divided into small blocks and each block is characterized as a saliency value.
e second is the extracted salient object is not complete and uniform enough. e reason behind this is that the edge region of the salient object that has a relatively large saliency value is highlighted, whereas other regions are just the opposite.

Saliency Map Fusion.
As described above, color contrast-based saliency detection has predominance in highlighting the whole salient region, while hash fingerprintbased saliency detection has an advantage in background suppression. erefore, it is feasible to combine the two to generate a final saliency map. Inspired by the idea proposed in [21], the final saliency map can be obtained by retaining and balancing the respective advantages of the two saliency maps. Specifically, the final saliency map is generated by using the hash fingerprint-based saliency map to refine the color contrast-based saliency map, and the formula is described as follows: where the constant factor c is set to 5, which is used to control the influence degree of S h on the final saliency map. It is clear that, when S f remains unchanged, the final saliency value decreases with the decrease of S h . As a result, the saliency value of the background region will be small because the corresponding S h is usually very small. is means that the background can be suppressed effectively via the saliency map fusion, as shown in Figure 4(d).

Experiment Results
In this section, we evaluate the performance of the proposed method and compare it with eight state-of-the-art saliency detection algorithms, including FT [13], IT [12], GBMR [17], CA [22], LC [14], MSS [15], SF [23], and our earlier work [19] named SRFC. ese methods are selected for comparison because they belong to bottom-up models and are the most classical saliency detection algorithms. Moreover, the executable codes for the above methods are publicly available.
In the experiments, five representative public datasets, DUTS [24], ECSSD [25], MSRA10K [26], SOD [27], and  Journal of Electrical and Computer Engineering HKU_IS [28], and a self-made dataset DADA-W are adopted to compare various saliency detection methods. Note that DADA-W is an image dataset that we captured from various water environments in which the salient regions include garbage, algae, spilled oil, etc. We randomly choose 60 images from each of the six datasets for performance testing. Due to space constraints, only a few representative images are selected and the comparison results are shown in Figure 5, where Figure 5 It can be seen from Figure 5 that the saliency maps of the proposed method are more complete and have fewer error detections in most cases. Specifically, our method can not only suppress the background but also highlight the whole salient region. Especially for complex scenarios, the proposed method and SRFC obviously outperform the others because they both adopt the background optimization selection. For example, for the 11th image, all the state-of-the-art methods except for the two methods we proposed can hardly completely extract lotus leaves from the water. Besides, compare to SRFC, the proposed method is more effective and can completely extract the whole salient object from the background. For certain images, there maybe exist small holes and fragmentation in the detection results of SRFC, as seen from the results of the 3rd and 9th images. e reason is that the goal of SRFC is to distinguish the salient region from the background as far as possible, whereas the proposed method puts much emphasis on highlighting the whole salient region while suppressing the background.
ree standard metrics, namely the area under ROC curve (AUC) score, mean absolute error (MAE) score, and F β-score (β � 0.3) are utilized to quantitatively compare the   detection performance of the proposed method against the other methods. Table 1 shows a detailed comparison, in which the best three metrics are marked with red, blue, and green, respectively. From Table 1, we can see that the proposed method is always ranked in the top three over all datasets, which means that our method has certain versatility for various images. Compared to other methods, our method increases by an average of 23%, 11%, and 9% in terms of MAE scores, F β -scores, and AUC scores, respectively.
Furthermore, the precision-recall (PR) curve and the receiver operating characteristics (ROC) curve are also used for intuitive comparison. Figures 6 and 7, respectively, show the PR and ROC curves of different algorithms over six datasets. It is clear that the proposed method achieves the most desirable detection performance. Concretely, the PR and ROC curves of our method are much better than those of FT, SF, and LC, but slightly better than those of IT, CA, and MSS. e reason is that these methods do not make full use of image background information. On the other hand, the PR and ROC curves of GBMR and SRFC are close to those of our method. However, when there exist multiple salient regions in the image or high color contrast in a salient region, our method clearly outperforms both of them.

Conclusion
In this study, we propose a novel and effective saliency detection algorithm. First, we create an accurate background template using background optimization selection. Based on this, saliency detection based on color contrast can be significantly improved to better highlight the salient region. en, the saliency detection based on hash fingerprint is introduced to suppress the background. Finally, due to the two maps being complementary to each other, the final saliency map is obtained by fusing the two detection results. Experimental results show that the proposed algorithm achieves good detection results and outperforms other traditional saliency detection algorithms.

Conflicts of Interest
e authors declare no conflicts of interest.

Authors' Contributions
Yong Wang contributed to conceptualization, original draft preparation, and funding acquisition; Yin Lv and Xuanrui Zhang were involved in formal analysis and validation.