Semantic segmentation method based on dense scale dynamic network

Abstract

The semantic segmentation method based on the dense scale dynamic network solves the difficult problem of segmentation caused by the existing DCNN in semantic segmentation and the large-scale changes of similar objects, and belongs to the technical field of semantic segmentation. The present invention provides a dense scale module DSM, which includes n layers of parallel units connected in sequence, and each layer of parallel units includes m depth dynamic local hole convolution DDLACs and a 1×1 convolution; The atrous convolution is adjusted to be generated by dynamic local and conditional parameterized convolution, including the conditional filter generation network CFGN and the corresponding dynamic local filtering operation unit DLFO; the feature map output by the backbone network is input into each CFGN to Generate a set of filters. The feature map output by the backbone network is input into the first DLFO of the DSM, each DLFO is segmented on the basis of the corresponding filter set, and a fine segmentation map is generated through n layers of parallel units.

Description

technical field [0001] The invention relates to a semantic segmentation method based on a dense scale dynamic network, and belongs to the technical field of semantic segmentation. Background technique [0002] Semantic segmentation plays an important role in many applications such as autonomous driving, medical imaging, robotic vision, etc. Although deep convolutional network (DCNN)-based semantic segmentation algorithms outperform traditional algorithms, they still face two major challenges: 1) Most DCNNs have spatial invariance, which reduces the localization accuracy during segmentation; 2) In image data, the large-scale variation of homogeneous objects causes segmentation difficulties. [0003] The current mainstream DCNN consists of filters (weights) shared in the spatial domain. They are spatially invariant to local image transformations. This invariance means that objects in the image can still be accurately recognized by DCNN after being rotated or translated. Th...

Images