Abstract:To address the high cost of full-scale fine-tuning for large-scale visual models and the limitations of existing parameter-efficient fine-tuning (PEFT) methods that neglect visual signal characteristics, this study proposes an adapter fine-tuning method SP-Adapter based on spatial perception and sample-adaptive modulation. This approach reconstructs one-dimensional image sequence features into feature maps with spatial topology through spatial reconstruction. It explicitly injects spatial inductive biases using convolutional operators to enhance the model’s ability to model spatial visual features. Building upon this foundation, SP-Adapter introduces a channel attention mechanism adaptively generated from sample context. It dynamically recalibrates feature response weights in the high-dimensional latent space based on semantic differences in input samples, thereby enhancing task-relevant discriminative visual representations and improving the model’s adaptability in downstream tasks. Experimental results on the VTAB-1K and FGVC benchmark datasets demonstrate that SP-Adapter exhibits strong generalization capabilities with minimal parameter tuning, outperforming other classical efficient parameter fine-tuning methods.