A Lightweight Adapter for Efficient Fine-Tuning in Computer Vision

Abstract Pretrained vision backbones are becoming increasingly large, making full fine-tuning expensive in both training time and storage, especially when adapting a single backbone to many tasks or data domains. In the setting of Parameter-Efficient Fine-Tuning (PEFT), we propose DT1D-Adapter , a lightweight adapter that can be plugged into both convolutional networks (ConvNets) and Transformer-based models by operating on spatial features of size \((H\times W)\). DT1D-Adapter performs axial filtering via depthwise 1D convolution along the height and/or width, where dilation enlarges the receptive field with minimal parameter increase. To control the parameter budget, the filters are parameterized by symmetric, group-shared coefficients, optionally complemented with lightweight channel mixing using grouped \((1\times 1)\) projections and a small-initialized scalar residual gate to stabilize optimization under limited trainable parameters. Experiments on multiple image classification datasets show that DT1D-Adapter provides a strong accuracy--parameter trade-off, remaining competitive with common PEFT baselines such as SSF, BitFit, and VPT, and demonstrating notable efficiency compared with convolution-based adapters (e.g., Conv-Adapter) and residual-branch variants (e.g., Residual Adapters).We further report a simple video-stream inference benchmark, indicating that DT1D-Adapter remains compatible with real-time deployment on common GPU platforms.