The deep stacked autoencoder, as a prominent deep learning architecture, has been widely applied in fields such as data science and pattern recognition. However, existing deep stacked autoencoders focus on transforming the features of individual samples, often overlooking the inter-sample correlation, which can lead to suboptimal feature quality. To address this limitation, this paper introduces a novel deep stacked autoencoder architecture called the dual-stage joint projection envelope embedded stacked autoencoder. Unlike traditional deep stacked autoencoders, the proposed model transforms deep features based on the correlation information between samples, rather than focusing solely on the samples themselves. The model is composed of two primary components: the dual-stage joint projection envelope and the embedded stacked autoencoder. The dual-stage joint projection envelope utilizes a manifold sample-pair envelope module to extract local correlation information from the original samples and reconstruct the first layer of enveloped samples. A descending clustering module is then employed to capture global correlations and reconstruct the second layer of enveloped samples. Additionally, the dual-stage inter-consistency maintenance module enhances the representational power of the second-layer enveloped samples. Subsequently, two sets of deep features are extracted by training two embedded stacked autoencoders on these two layers of enveloped samples. The paper concludes with four sets of experiments: ablation studies, algorithm comparisons, parameter sensitivity analysis, and complexity analysis. Experimental results demonstrate that the deep features extracted by the proposed dual-stage joint projection envelope embedded stacked autoencoder exhibit both high quality and stability.