Digital holographic microscopy (DHM) involves computationally demanding operations such as numerical wave propagation, phase retrieval, and three-dimensional (3D) reconstruction, which rely heavily on large-scale matrix computations. As spatial resolution and sampling density increase, the associated computational burden grows substantially, often necessitating optimized signal-processing strategies to enable real-time and high-fidelity reconstruction. To address this, we propose a sparse reconstruction method based on adaptive dictionary learning (DL) named as Sequential Generalized K-means (SGK) is integrated into DHM. In the proposed method, input hologram is divided into patches, and each patch is represented as linear combination of sparse dictionary atoms. Each atom in dictionary is updated sequentially by averaging the training patches associated with its fixed sparse coefficients. Unlike K-SVD, proposed method avoids the expensive singular value decomposition (SVD) for dictionary update thereby minimizes the computation time and improves the hologram reconstruction accuracy. To further enhance reconstruction quality, optimum choice of key parameters such as patch size, dictionary size, and sparsity level is essential. Experimental results demonstrate that the proposed SGK-based sparse reconstruction achieves high-fidelity reconstructions and faster convergence than conventional dictionary-learning approaches, highlighting its suitability for real-time microscopic imaging applications.