Seismic data processing is crucial to identifying Earth’s subsurface layers. However, during seismic data acquisition, seismic data are often corrupted by random and erratic noise, which may degrade seismic data quality and lead to misinterpretation of Earth layers. Random noise elimination has been addressed using various standard denoising techniques, and erratic noise poses a significant challenge due to its discontinuous high-amplitude peaks. Existing dictionary-based denoising methods, such as double sparsity dictionary learning (DSDL), etc., involve iterative processes in various stages that are computationally expensive. To address these limitations, we propose a convolutional dictionary learning (CDL) method using ℓ2,1 norm to provide group sparsity and alternating direction method of multipliers (ADMMs). The ℓ2,1 norm combines local sparsity, internal patch self-similarity, and their correlation to external reference estimates to effectively preserve weak signals through group sparsity. CDL method effectively eliminates both random and erratic noise while preserving critical signal features and reducing computational complexity. Performance of CDL is evaluated in terms of evaluation metrics such as mean squared error (MSE), signal-to-noise ratio (SNR), and energy values. The proposed CDL method demonstrated better denoising with signal feature preservation and reduced computational time.