Nonlinear Membrane-System-Based Multiscale Feature Purification Learning Framework

Deep neural networks have achieved remarkable performance in visual representation learning, but their hierarchical feature aggregation often introduces redundant and task-irrelevant responses, limiting representation quality in fine-grained visual analysis tasks. To address this issue, this paper proposes FPNet-NSNP, an NSNP-inspired multiscale feature purification framework for image forgery localization. The proposed framework integrates three complementary components: a Cross-layer Feature Fusion (CFF) module for hierarchical feature interaction, a Visual Perception (VP) module for multiscale contextual representation, and a Feature Purification (FP) module for adaptive feature refinement. In particular, the proposed ConvSNP unit incorporates nonlinear response regulation inspired by Nonlinear Spiking Neural P Systems (NSNP), enabling more effective feature transformation and redundancy reduction during propagation. Extensive experiments on benchmark image forgery localization datasets demonstrate that FPNet-NSNP achieves competitive performance compared with existing methods. Moreover, robustness evaluations under noise perturbation and JPEG compression, together with cross-domain experiments on medical image manipulation data, further verify the stability and generalization capability of the proposed framework. The results indicate that NSNP-inspired nonlinear regulation provides an effective mechanism for improving multiscale visual representation learning.