Semantic segmentation method based on dense scale dynamic network

Abstract

The invention discloses a semantic segmentation method based on a dense scale dynamic network, solves the problem that segmentation is difficult due to the positioning precision and large-scale change of similar objects during semantic segmentation of an existing DCNN, and belongs to the technical field of semantic segmentation. The invention provides a dense scale module (DSM), which comprises n layers of parallel units connected in sequence, and each layer of parallel unit comprises m depth dynamic local cavity convolution DDLAC and a 1*1 convolution; each DDLAC is generated by adjusting deep cavity convolution into dynamic local and conditional parametric convolution, and comprises a conditional filter generation network (CFGN) and a corresponding dynamic local filtering operation unit (DLFO); and inputting the feature map output by the backbone network into each CFGN to generate a filter set. A feature map output by a backbone network is input into a first DLFO of a DSM, each DLFO is segmented on the basis of a corresponding filter set, and a fine segmented 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, robot vision, etc. Although the semantic segmentation algorithm based on deep convolutional network (DCNN) outperforms traditional algorithms, it still faces two major challenges: 1) most DCNNs are spatially invariant, which reduces the localization accuracy during segmentation; 2) In image data, the large-scale variation of the same object makes segmentation difficult. [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 after the object in the image is rotated or translated, it can still be accurately recognized by DCNN. T...

Images