Plane segmentation method based on superpixels and graph convolutional network

Abstract

The invention discloses a plane segmentation method based on superpixels and a graph convolutional network. The method comprises the following steps: preprocessing an input color image to obtain a preprocessed image; then, according to the image resolution, segmenting the preprocessed image into a proper number of superpixels, and converting the superpixels into an undirected graph structure; constructing a graph convolutional network, and training the graph convolutional network by using a data set; and finally, predicting a graph formed by the superpixels by using the trained graph convolutional network, and carrying out plane classification on each superpixel so as to complete plane segmentation. According to the method, the image is segmented into the superpixels, so that the edge information in the original image can be well reserved, the learning burden of a subsequent image neural network is reduced, and the problem that the segmented plane edge is too large in gap with the actual situation is prevented; and a specific label is extracted through a specific algorithm by utilizing an existing data set and serves as a training set for subsequent supervised neural network learning, so that the problem that the data set is not segmented for a plane is solved, and huge cost of manual marking is avoided.

Description

technical field [0001] The invention belongs to the field of plane segmentation calculations, in particular to a plane segmentation method based on superpixels and graph convolutional networks. [0002] technical background [0003] Computer vision is a highly interdisciplinary and complex subject with rapid development and is widely used in security, transportation, medical and other fields, and is closely related to people's lives. The goal of computer vision research is to let computers replace humans to complete visual tasks such as target detection, image description, and face recognition. On the scale of human perception, the plane is one of the most common structures in the environment. It has strong constraints, constraining a large number of points / lines and the information they carry, and various surfaces can be approximated by planes. Choose the number of planes to fit based on your accuracy requirements. In practical applications, many computer vision tasks requ...

AI Technical Summary

Problems solved by technology

However, no matter what method is used, the segmented plane edge is somewhat different from the actual situation, and it cannot be well segmented to obtain the real plane edge.

For the neural network, the problem is that its segmentation plane is essentially clustering pixels, and the network cannot learn edge information well during the clustering process

The superpixels used in the geometric method can well preserve the edge information in the image, but the ordinary convolutional neural network can only be used to process the input of regular distribution of information such as images.

Images