Improved EdgeCraft-Based Few-Shot Cross-Domain Detection for Remote Sensing Images

摘要

Place an optical satellite image of a harbor beside a SAR image of the same location. The ships that read as neat elongated rectangles in the optical view become irregular clusters of bright pixels against a grainy, noise-like background in the SAR view. Boundaries fray; portions of the hull dissolve into speckle; nearby sea clutter occasionally produces phantom reflections that could be mistaken for small vessels. A detection model trained on thousands of optical ship examples, when fed a SAR image, may fail to recognize these objects—not because the ships are absent, but because the visual signatures are unrecognizable. The problem tightens when SAR labels are scarce. In a typical few-shot setup where K in {3, 5, 10, 30} labeled SAR bounding boxes are available, a model sees almost nothing that resembles a SAR ship during its brief exposure to the target domain. Small vessels with weak backscatter are hit hardest. Their radar returns barely clear the speckle floor, and with so few labeled examples the network has little signal to latch onto. A Multi-Scale Attention Feature Enhancement (MSA) mechanism addresses the structural challenge. Multiplicative speckle fractures target boundary continuity, but gradient magnitudes suffer far less distortion than raw pixel intensities. Edges extracted from shallow gradients therefore preserve contour information that noise erases, and the MSA module fuses these cues with deep semantic features to strengthen the representation of small, weak SAR targets. A Sparsely Supervised Cross-Domain Feature Alignment (SCA) module addresses the distribution challenge, using MMD and CORAL losses to narrow the gap between domains with the few labeled SAR boxes serving as anchor points. The few labeled SAR anchors pull features of the same class together while preserving the structural separability between foreground and background. Experiments on the HRRSD-to-SSDD transfer task show that our framework achieves better performance than mainstream few-shot detectors and cross-domain adaptation methods.