AMSD-Net: Adaptive Multiscale Dynamic Network With Dynamic Feature Modulation for Small-Target Detection in Remote Sensing

Abstract

Small-target detection in remote-sensing images is crucial in fields like urban monitoring, maritime surveillance, and disaster assessment. Nevertheless, it confronts challenges such as significant multiscale variation, intense background interference, and dense target distribution. These issues can readily result in insufficient feature representation, scale discontinuity, and problems of false and missed detections. To tackle these problems, this letter presents the adaptive multiscale dynamic network (AMSD-Net). The hierarchical scale interaction fusion (HSIF) structure is utilized to boost the complementarity between high- and low-dimensional features, mitigate scale discontinuity, and reduce missed detections of small targets. The dynamic feature modulation (DFM) module enhances the capture of small-target details and suppresses noise. The shared and spatial-channel synergistic head (SCS-Head) decreases redundancy and minimizes false detections. Finally, the method was verified on two public datasets, proving its effectiveness in small-target detection.