Segmentation model training method for image with vertex and segmentation method for image with vertex

Abstract

The invention relates to an image segmentation technology, and discloses a segmentation model training method for an image with a vertex and a segmentation method for the image with the vertex. The segmentation model training method comprises the steps of obtaining a training sample set, wherein the training sample set is provided with a true value mask; inputting the training sample set into a coarse segmentation model to obtain a coarse segmentation result mask of each image, wherein the coarse segmentation result mask comprises a corresponding coarse segmentation result; extracting vertexes of the image on a truth value mask of the training sample set, distributing a neighborhood range for each vertex, mapping each vertex on the truth value mask and the corresponding neighborhood range to a corresponding coarse segmentation result mask, calculating the minimum distance between a coarse segmentation result in the mapped neighborhood range and the mapped vertex, accumulating and summing the minimum distances of all neighborhood ranges of each image to serve as a vertex distance loss parameter; and iteratively optimizing the coarse segmentation model by taking the vertex distance loss parameter as a constraint condition to obtain a fine segmentation model.

Description

technical field [0001] The present application relates to image segmentation technology, in particular to segmentation model training of images with vertices and segmentation technology of images with vertices. Background technique [0002] The production of high-precision maps based on raster maps generally requires steps such as map collection, vectorization of map elements, and map data production. Among them, the vectorization step of map elements is very important. [0003] The current image segmentation methods mainly include artificial vectorization methods, traditional image processing methods for extracting target boundaries, and existing methods based on deep learning image segmentation technology to extract target boundaries. First of all, for the method using artificial vector method, the efficiency is low and error-prone, which greatly increases the labor cost. Secondly, for the method of using traditional image processing to extract the target boundary, becau...

AI Technical Summary

Problems solved by technology

First of all, for the method using artificial vector method, the efficiency is low and error-prone, which greatly increases the labor cost

Secondly, for the method of using traditional image processing to extract the target boundary, because this method has high requirements on imaging quality and poor adaptability to complex environments, and in the process of high-precision map collection, the quality of acquired images (raster images) is generally low , most of the actual scenes are relatively complex, which makes the effect of extracting the target boundary by this method unstable, and the accuracy of the boundary extraction results is generally low

Thirdly, for the existing method of extracting target boundaries based on deep learning image segmentation technology, although the performance is relatively stable, the target boundaries in the extraction results are generally rough and have a large difference from the real boundaries of the target.

[0004] However, for map elements with vertices such as direction arrows, diamond-shaped deceleration signs, and triangular deceleration signs, due to their large number, variety, vertices, and curved edges in some arrows, none of the above methods can satisfy their high-efficiency fine segmentation. Require

Images