Adaptive multi-scale feature fusion with spatial translation for semantic segmentation

In image segmentation tasks, contextual information is crucial as it provides essential semantic details. Multi-scale feature extraction methods help models capture this contextual information comprehensively, but they can introduce redundancy and insufficient receptive fields in some areas, particularly with large objects or complex scenes. To address these issues, we propose the Adaptive Feature Perception Module (AFPM). Inspired by the visual system, we combine the pyramid model with dilated convolutions and incorporate a spatial shift mechanism for extensive information capture.This module adaptively adjusts its focus and perception range to maximize target feature capture.Meanwhile, we introduce the Channel and Spectral Attention Module(CSAM) to model dependencies between channels and spectral domains,enabling the network to learn more discriminative features and improve segmentation accuracy. Based on these enhancements,we propose a new network model called AMFFNet. We validated its effectiveness by comparing it with several state-of-the-art methods on the PASCAL VOC 2012, Cityscapes and ADE20K datasets. The results demonstrate that AMFFNet offers superior performance.