Single-cell RNA-sequencing (scRNA-seq) has emerged as a powerful high throughput technique that enables the characterization of transcriptomic profiles at single-cell resolution. However, scRNA-seq data has an excess number of zeros in expressed genes due to a low amount of sequenced mRNA in each cell. This missing detection in a portion of mRNA molecules (dropout) presents a fundamental challenge for various types of data analyses. Here we introduce scIDS, a novel imputation method that is a combination of deep autoencoder neural networks and subspace regression to reliably recover the missing values in scRNA-seq data. We compare scIDS with two widely used methods using eight datasets. Extensive experiments demonstrate that scIDS outperforms existing approaches in improving the identification of cell populations while preserving the biological landscape.