［Purpose /significance］Image resource retrieval for digital library needs an accurate and fast response. However, traditional image retrieval methods based on deep learning have the problems of long time for training model, unrefined feature and slow retrieval speed. We proposed Multi-Group-Two-Layer Sparse Auto-Encoder (MGTLSAE) to learn refine features for fast image resource retrieval.［Method /process］Firstly, extracted deep learning features of the image resources with the pre-train model GoogLeNet as initial features which were clustered with the density peak(DP) clustering algorithm. Secondly, designed the network of MGTL-SAE using the number of classes of clusters, initialized the mean and variance parameters of MGTL-SAE with the image blocks selecting from the images according to the image resource sub-sets partitioned by the clusters, estimate the parameters of MGTL-SAE to learn features by the gradient descent algorithm until the cost function is converged, iterate the DP clustering algorithm and estimating the parameters of MGTL-SAE to compute the refined feature until the feature bases are converged. Lastly, retrieved the query sample image fast within the sub-set of image resources partitioned by the DP clusters of refined features.[Result /conclusion] Experiment results show that our method has higher precision, recall and mAP(mean Average Precision) and less retrieval time than the methods based on CaffeNet and GoogLeNet . It also provides a new perspective for the research of image resource retrieval in the digital library.