基于AFFormer的小目标增强识别算法改进

打开文本图片集

中图分类号：TP242

文献标识码：A

Abstract： Lightweight semantic segmentation networks limited by model capacity often hampers the extraction of fine-grained features，especially for small-scale objects，leading to a significant drop in segmentation accuracy. To address this issue，we propose a novel small-object-enhanced segmentation framework based on an Adaptive Frequency Transformer （AFFormer）. By introducing only a modest number of additional parameters，our approach significantly boosts the model's ability to recognize small objects. Specifically，three innovative modules are designed： a Multi-Scale Feature Fusion Module，which enhances the representation of small objects through cross-scale feature interaction；an Efficient Visual Mamba Module （Efficient ViM），which captures local contextual information with low computational overhead，thereby reinforcing feature representation while preserving generalization ability；and a Frequency-Aware Feature Fusion Module，which leverages frequency-domain decomposition to process low and high-frequency components separately， effectively mitigating intra-class inconsistency and boundary ambiguity. Experiments conducted on the high-resolution Cityscapes dataset demonstrate that our proposed model achieves a mean Intersection over Union（mIoU） of 78.72% ，surpassing the baseline by 0.95% ，thereby validating the effectiveness of the proposed method.

Keywords： AFFormer; semantic segmentation; small target recognition; lightweight

图像语义分割是将图像中的待定对象或区域单独定位、划分、并定义为像素级分类任务[1-2]，便于应用到自动驾驶、工业生产等各类下游任务，因此受到广泛关注。（剩余11503字）