Abstract: In order to solve the problem that the saliency detection target boundary is vague, the paper proposes a saliency feature extraction method based on multi-scale region covariance. This method firstly extracts the multi-scale features of an image, then combines region covariance to extract the image bottom features, calculates the image’s multi-scale uncertainty weights, and the weights are optimized for the final saliency features, which are obtained by fusion. In the paper, our proposed model compares the experimental results with the commonly used feature extraction algorithms. The experimental results show that the proposed algorithm is closer to the actual boundary of the object, the different weight is given to the different multi-scale in the process of multi-scale regional covariance saliency feature extraction, which could enhance the part that is focused by human eyes on the image, and the effect of saliency feature extraction can be improved.

Key words: saliency feature, region covariance, multi-scale transform, multi-feature fusion