As current saliency detection algorithms based on deep convolutional network have problems of incomplete target and noisy background detected from complex scene images, a new algorithm of deep feature-oriented saliency detection composed with basic feature extraction and high-level feature which guided cross-level aggregating delivery was proposed. It was based on the improvement of an extant Encoded Low level distance map with Deep features (ELD) model. Firstly, according to the characteristics of convolutional features at different levels, a cross-level feature fusion network model of high-level feature guidance was established. Then, saliency clustering propagation by using high-level feature guidance on initial saliency map that generated by improved neural network was implemented. Finally, final saliency map with more details and less noise was generated by using fully-connected conditional random field after saliency propagation. The experimental results on ECSSD and DUT-ORMON data sets show that, the Precision-Recall (PR) performance of the proposed algorithm is better than ELD algorithms, and F-measure(F) is increased by 7.5% and 11%, respectively, while its Mean Average Errors (MAE) are decreased by 16% and 15%, respectively,which also can obtain more robust results in complex image scene fields of target recognition, pattern recognition, image indexing, and so on.