基于Transformer的边缘检测网络

打开文本图片集

关键词：边缘检测;Transformer;注意力机制；多级聚合特征金字塔；多尺度注意力增强中图分类号：TP394.1;TH691.9 文献标识码：Adoi：10.37188/OPE.20253322.3564 CSTR：32169.14.OPE.20253322.3564

Abstract：The currnt mainstream edge detection method based on convolutional neural network has limitations in receptive field range and fine-grained edge perception.With the development of Vision Transformer，its global modeling ability and flexible information interaction mechanism bring new possibilities foredge detection tasks. To solve this issue，this paper proposed an encoder-decoder model named TFEdge，which combined Transformer，Multi-Level Aggregation Feature Pyramid （MLAFP），and Multi-Scale Attention Aggregation （MSAA） modules for high-precision edge detection. The model introduced the Dilated Neighborhood Attention Transformer as the backbone network and extracted the global context information and local edge clues of the image through a multi-stage cascade design. Simultaneously，the Multi-Level Aggregation Feature Pyramid was designed to aggregate the deep and shallow features of each stage，endowing the shallow features with more abundant semantic features to suppress image noise and improve the detection ability of indistinct boundaries.Finally，the Multi-Scale Attention Aggregation module， based on an attention mechanism，was proposed to further enhance feature representation by aggregating the cross-scale spatial and channel attention information of feature maps.The experiment is evaluated on the BSDS5OO and NYUDv2 datasets. The ODS and OIS F-scores of TFEdge on the BSDS500 are 0.857 and 0.874，respectively，while on the NYUDv2 they are 0.788 and O.801，respectively. Compared with many existing methods，TFEdge shows superior edge detection performance in both quantitative and qualitative results.

Key words： edge detection; transformer; attention mechanism；multi-level aggregation feature pyramid ; multi-scale attention aggregation

1引言

边缘检测是数字图像处理和计算机视觉中的基础任务之一，旨在从图像中识别并提取出物体与背景之间的边界，为后续图像分割1、目标检测2和目标追踪[3等各种高级视觉任务提供关键信息，具有广泛的应用场景。（剩余18274字）