Abstract: To deal with the high dimensions and noise of CNN features,a method to reduce dimensions of CNN features is proposed,principal component analysis(PCA)was first used on CNN features to reduce the dimensions,the effectiveness of PCA was verified in both data aspect and human subjective perception.Then a saliency model was constructed by using the multi-level superpixels segmentation and random forest to fuse PCA-CNN features and handcrafted features.Experiments demonstrated the proposed saliency model outperformed other traditional saliency methods,the PCA-CNN features can improve performance of saliency detection.