Visual saliency detection via invariant feature constrained stacked denoising autoencoder

Visual saliency detection is usually regarded as an image pre-processing method to predict and locate the position and shape of saliency regions. However, many existing saliency detection methods can only obtain the local or even incorrect position and shape of saliency regions, resulting in incomplete detection and segmentation of the salient target region. In order to solve this problem, a visual saliency detection method based on scale invariant feature and stacked denoising autoencoder is proposed. Firstly, the deep belief network would be pretrained to initialize the parameters of stacked denoising autoencoder network. Secondly, different from traditional features, scale invariant feature is not limited to the size, resolution, and content of original images. At the same time, it can help the network to restore important features of original images more accurately in multi-scale space. So, scale invariant feature is adopted to design the loss function of the network to complete self-training and update the parameters. Finally, the difference between the final reconstructed image obtained by stacked denoising autoencoder and the original is regarded as the final saliency map. In the experiment, we test the performance of the proposed method in both saliency prediction and saliency object segmentation. The experimental results show that the proposed method has good ability in saliency prediction and has the best performance in saliency object segmentation than other comparison saliency prediction methods and saliency object detection methods.

Automated summary

A visual saliency detection method based on scale invariant feature and stacked denoising autoencoder is proposed. The method utilizes a deep belief network to initialize the parameters of the autoencoder network and uses scale invariant feature to design the loss function for self-training and parameter update. Experimental results demonstrate the method's effectiveness in saliency prediction and object segmentation, outperforming other comparison methods.