中文 English

Image Saliency Detection With Sparse Representation of Learnt Texture Atoms

Click:

     

    Image Saliency Detection With Sparse Representation of Learnt Texture Atoms

    Lai Jiang, Mai Xu, Zhaoting Ye, Zulin Wang Beihang University, Beijing, 100191, China

     

     

    Abstract

    This  paper proposes a saliency detection method using a novel feature on  sparse representation of learnt texture atoms (SR-LTA), which are  encoded in salient and non-salient dictionaries. For salient dictionary,  a novel formulation is proposed to learn salient texture atoms from  image patches attracting extensive attention. Then, online salient  dictionary learning (OSDL) algorithm is provided to solve the proposed  formulation. Similarly, the non-salient dictionary can be learnt from  image patches without any attention. A new pixel-wise feature, namely  SR-LTA, is yielded based on the difference of sparse representation  errors regarding the learnt salient and non-salient dictionaries.  Finally, image saliency can be predicted via linear combination of the  proposed SR-LTA feature and conventional features, i.e., luminance and  contrast. For the linear combination, the weights corresponding to  different feature channels are determined by least square estimation on  the training data. The experimental results show that our method  outperforms several state-of-the-art saliency detection methods




     

    Motivations

    It  is well-known that an image patch can be seen as the composition of  texture patterns. After some experiments, we found that salient image  patches could be represented by similar patterns. The figure below shows  the possibility of learning hundreds of salient patterns (by applying  the dictionary learning algorithm) for saliency detection. In this  figure, the regions inside the red squares (enlarged in the corners) are  salient patches, in the images of the eye tracking Kienzle database  (the first row) and Doves database (the second row). Some atoms of the  dictionaries, learnt from the salient regions of other training images,  are shown in the middle of two images. In addition, the sparse  representation coefficients of  the salient patterns regarding the learnt dictionaries are also  provided. It can be seen that the salient patches across the different  images may share some similar basic patterns, and these basic patterns  may be learnt from the training data. Note that the non-salient patches  in the second row have the same characteristic.


         

     




     

    Algorithms

    This  work proposes to learn extensive positive and negative patterns from  the eye tracking data of training images, for bottom-up saliency  detection. Specifically, we first proposes a formulation with a novel  center-surround term, for learning two discriminative dictionaries.  These two dictionaries contain the atoms for basic texture patterns of  salient and non-salient regions, respectively. In light of online  dictionary learning, we develop an online salient dictionary learning  (OSDL) algorithm to solve the proposed formulation, and then the salient  and non-salient dictionaries can be learnt from the eye tracking data  of training images. Given the learnt dictionaries, a novel feature based  on sparse representation of learnt texture atoms (SR-LTA) is worked out  in our method. Such a feature is generally based on the errors of  sparse representation regarding salient and non-salient dictionaries.  Next, the saliency of an image can be predicted, via combining the  SR-LTA feature with conventional luminance and contrast features. For  the linear combination, the weights corresponding to each feature  channel are estimated via least square fitting on the training data.   Similar to other bottom-up methods, we only works on gray images with  natural scenes.


            The summary of online salient dictionary learning (OSDL) algorithm.

    undefined


    The procedure of our salieney detection method.

    undefined

     




     

    Results

    The averaged accuracy of saliency detection on test images of two databases.

    undefined

     

     

    The ROC curves of saliency detection.

    undefined

     

               

       Saliency maps of four test images from two database.

    From left to right: Input images, human fixation maps, our, BMS, Itti, Duan, GAFFE, Hou, Zhao, Judd, AWS and SALICON methods.

    undefined

     




     

    Reference

    1.    Lai  Jiang, Mai Xu, Zhaoting Ye, Zulin Wang. Image Saliency Detection with  Sparse Representation of Learnt Texture Atoms. The IEEE International  Conference on Computer Vision (ICCV) Workshops, 2015, pp. 54-62

    2.    (In  submitting) Mai Xu, Lai Jiang, Zhaoting Ye, Zulin Wang. Bottom-up  saliency detection with Sparse Representation of Learnt Texture Atoms.  Pattern Recognition 2016.



Copyright © 2018 Beihang University. All rights reserved.

Address: No. 37 Xueyuan Road, Haidian District, Beijing, P.R. China,
PV: