Vison Old, Vision New.
After reading this paper , you will get new insights of computer vision.
For more detail information, please contact with me via gmail [168fangjunwen@gmail.com].
- computer vision
- image saliency
- image segmentation image superpixel segmentation demo
A novel object level saliency model is proposed in this letter via iterating saliency classification difference on dynamic image region partitioning.
First, the model proposed solved three problems which is caused by static background based methods. Then dynamic background is represented and computed on the input image via dynamic image partitioning. Unlike existing static background based methods, we calculate saliency difference based on dynamic background rather than static image region. This strategy makes the saliency result more precisely to the location of image objects. We apply two saliency classification difference on the dynamic background.
Second, saliency classification difference is iterated on saliency maps generating by dynamic image region partitioning. This makes the saliency results more robust. To get a more robust result, the dynamic image partitioning is operated on an image in four directions (i.e., left to right, right to left, top to bottom, bottom to top).
Third, the final saliency map is generated by combining four saliency maps based on four direction scanning. The four direction combination enables the proposed method to uniformly highlight the salient object and simultaneously suppress the background effectively. Extensive experiments on two large dataset demonstrate that the proposed method performs favorably against the classic methods in terms of accuracy and efficiency.
Index Terms—Dynamic image region partitioning, iterating, dynamic background, four direction scanning, saliency map.
In this letter, we present a bottom-up object level saliency detection model by exploiting dynamic image region partitioning, iterating saliency difference and four direction scanning.
Firstly, we get more precise results than the background based methods with dynamic image region partitioning.
Secondly, iterating the saliency classification difference automatically separates the salient object and the background.
Thirdly, the four direction scanning makes our algorithm more robust.
Saliency maps on a large public dataset demonstrate that the proposed method can highlight the whole object region uniformly and suppress the background region effectively.
In addition, the proposed method performs favorably against the classic methods in accuracy, which shows that the proposed dynamic image region partitioning, iterating saliency classification results and four direction scanning are useful for saliency detection. In the future work, we will investigate a more complicated classification strategy to boost classification results at different scales and explore more applications of dynamic image region partitioning to other saliency algorithms.