Research On Image Saliency Detection Using Aggregation Degree Of Color And Texture

Visual saliency detection aims to extract the region that attracts greater attention of human vision system than other parts of an image, which is rapidly developed, widely exploited but still remains a challenge in the computer vision field. Research shows that the saliency regions are more aggregative in color and texture distribution than the background regions. Based on this character of saliency regions, we introduce and define the conception of aggregation degree, prove the property of aggregation degree of data sets, and propose a novel bottom-up saliency detection framework using the aggregation degree of color and texture. The algorithm model of this paper contains several steps as follows.Firstly, the color saliency areas are extracted from five color channels L, a, b, H, S of Lab and HSV color space, which are obtained via color space conversion from the original image and divided into several parts by FCM. And the significance of each saliency area is evaluated by the aggregation degree and their divergence.For the images, the color of the target regions and background regions of which are significantly similar, the structure of these images should be taken into consider. Thus the texture saliency map is extracted with the local phase of hypercomplex, which is constructed by the detail signal of stationary wavelet transform(SWT) and its Hilbert transform.Finally, we design an effective mechanism for credit feature maps selection depending on the character of aggregation degree proved in this paper, quantize the position relationship of candidate features,and weight the credit feature maps based on their significance. The final saliency map is reconstructed by fusing the credit feature maps.Extensive experiments are performed on ASD database. And the results of saliency maps and performance comparison with seven state-of-the-art algorithms demonstrate that our algorithm achieves better performance.