Abstract: To address the scarcity of radar jamming samples and severe noise interference in complex electromagnetic environments, this paper proposes a few-shot recognition network with cross-scale nonlinear metric, termed Task-driven Scale-Calibrated Relation Network (TSCR-RN). This network achieves task-adaptive feature reconstruction and denoising through a Task-driven Cross-Scale Channel Rectification (TSCR) mechanism. Furthermore, by integrating a Multi-Scale Relation Network (MSRN) based on high-order statistical characteristics, it effectively overcomes the performance degradation associated with traditional linear distance metrics under strong noise. Simulation experiments demonstrate that the proposed model achieves an average recognition accuracy of 84.23% using only 1% of the training samples. Even under an extremely low jamming-to-noise ratio (JNR) of -10 dB, it maintains excellent noise resistance and robustness.