边缘检测算子在道路图像语义分割中的应用
首发时间:2021-04-29
摘要:道路场景解析是常见的图像语义分割任务,讨论如何解决在道路场景里的图片分割模糊、边缘不连续、小物体难以识别等问题对于实际应用有着重要的意义。针对上述问题,我们在常见的编解码结构的语义分割模型中加入了一个边缘特征金字塔模块。该模块通过利用特征金字塔思想将不同尺度的边缘特征进行融合,可以有效地提高对小物体的识别能力。而边缘特征又可以弥补分割特征的不足,有效解决边缘不连续的问题。同时该模块在对编解码器各自进行特征提取后,利用欧式距离有效地对二者的特征表达能力进行相似性比较,可以提高解码器的解码能力,最终达到提升精度的目的。最后我们利用公开数据集Cityscapes来进行实验验证,本文方法相较于SegNet提升了7.5%,较ENet提升了6.2%,并通过消融实验证明了边缘检测的有效性。
关键词: 模式识别与智能系统; 特征金字塔; 边缘检测; 语义分割; 道路解析
For information in English, please click here
Edge Detection Operators Used in Road Image Semantic Segmentation
Abstract:Road scene parsing is a common task in semantic segmentation. Discussing how to deal with the problem of vague segmentation, margin discontinuity and poor recognition of small target has important significance in practical applications. Considering the above issues, we extend the segmentation network which has an encoder-decoder structure by an edge feature pyramid module. This module can produce a multi-scale feature representation by feature pyramid thought, and it can improve the ability to recognize small targets. Edge features can make up for CNN features\' shortcomings. What\'s more, this module can help network produce smooth segmentations. Meanwhile, after extracting features of the encoder and decoder, the module uses Euclidean distance to compare the similarity between encoder and decoder, which can increase the decoder\'s ability to restore from the encoder. Finally, the experiment on Cityscapes datasets demonstrates that the accuracy of our method has a 7.5% improvement over SegNet and a 6.2% improvement over ENet. And the ablation experiment validates the effectiveness of Sobel operators.
Keywords: pattern recognition and intelligence system Feature pyramid edge detection semantic segmentation road scene parsing
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
边缘检测算子在道路图像语义分割中的应用
评论
全部评论0/1000