Robust Target-Sensitive Region Segmentation via Adaptive Frequency Fusion and Dynamic Optimization

Target-sensitive region segmentation plays a crucial role in environmental understanding and perceptual decisionmaking. Its accuracy and robustness directly affect the reliability of smart devices in consumer electronics, such as autonomous vehicles and augmented reality glasses, as well as in medical consumer electronics, like intelligent endoscopic lesion segmentation systems. However, in complex scenarios, common cross-domain challenges, such as boundary ambiguity, semantic confusion, and high-frequency interference, often arise. Current mainstream methods overly rely on spatial-domain feature modeling, which limits their ability to explore implicit cross-scale semantic correlations in the frequency domain, thereby reducing segmentation robustness under heterogeneous interference. To address these challenges, we propose the Adaptive Frequency Fusion Dynamic Optimization Network (AFFDO-Net), which integrates three key modules: 1) The Frequency-Adaptive Enhancement Fusion Module (FAEFM), which dynamically fuses high- and low-frequency image information through a frequency-domain decoupling strategy, improving both global localization and edge-sharpening capabilities; 2) The Dynamic Response Optimization Module (DROM), which utilizes a Progressive Channel Self-Attention Mechanism (PCSM) and a Positional Information Mapping Module (PIMM) to dynamically optimize the fused features, alleviating intra-class inconsistency and boundary offset caused by the fusion of high- and low-frequency information; 3) The class-large kernel attention wavelet convolution module (SWC-LKA) embedded in the decoder, which mitigates noise interference, such as reflection artifacts, through its frequency-space dual-domain filtering characteristics, significantly enhancing the detail retention capability of the segmented regions. We conducted experiments on five publicly available datasets and one private dataset (Colps), and the results demonstrate that our network outperforms existing methods.