<p>Video-based action recognition remains a challenging task due to the difficulty in accurately modeling spatio-temporal dynamics and distinguishing foreground motion from static background clutter. Existing methods often struggle with capturing long-range temporal dependencies and tend to overfit to irrelevant background features, leading to reduced recognition performance in complex scenes. To address these limitations, we propose a biologically inspired two-branch convolutional network, termed Key-information Spatio-temporal Correlation Network (KSC-Net). The architecture integrates two novel modules. First, a Dynamic Feature Filter (DF) is introduced to enhance sensitivity to salient motion by suppressing redundant visual signals through second-order temporal difference and Laplacian-based spatial filtering. This module mimics the edge-enhancing and motion-focusing mechanisms of human vision. Second, a Spatio-temporal Self-similarity Gated (SG) module captures long-range correlations by computing feature similarity across frames and adaptively regulating memory propagation using a bi-directional gated structure with temporal offset pooling. Extensive experiments on public benchmarks including Kinetics-400, UCF-101, and HMDB-51 demonstrate that our proposed model achieves superior Top-1 recognition accuracy compared to state-of-the-art methods, validating the effectiveness of the proposed biologically inspired spatio-temporal modeling framework.</p>