DSG-Adapter: Parameter-Efficient Fine-Tuning for Wideband Signal Detection and Recognition

Abstract

Wideband signal detection and recognition (WSDR) is a key technology in the field of spectrum sensing. Recently, the paradigm of adapting pre-trained vision backbones to downstream WSDR tasks has become mainstream. However, conventional full fine-tuning incurs prohibitive computation and storage overheads. Parameter-efficient fine-tuning (PEFT) offers a low-cost alternative, while existing PEFT methods lack task-specific designs and systematic evaluations for the dense prediction WSDR. To address this challenge, we propose the dual-stream gated adapter (DSG-Adapter), a novel PEFT module specifically designed for WSDR. The DSG-Adapter first projects the input to a low-dimensional bottleneck space, and extracts time and frequency domain features separately using asymmetric strip depthwise separable convolutions. Subsequently, an adaptive channel-wise gated mechanism is introduced to enhance feature representations, which are then restored to the original dimension. As a plug-and-play module, the DSG-Adapter exhibits high flexibility and can be seamlessly inserted after the residual connections of attention or feed-forward network blocks. Extensive experiments demonstrate that, with only approximately 4% of the backbone network parameters fine-tuned, the DSG-Adapter significantly outperforms other PEFT methods and even surpasses full fine-tuning. Our code and dataset are now available as open-source resources at https://github.com/xk-maki/DSG-Adapter