Scene segmentation correction method and system fusing local information

Abstract

The invention relates to a scene segmentation correction method. A local boundary correction network is adopted, a completely residual convolutional network is used as a front-end model, a confidencemap and an original image of the front-end model are spliced according to a channel as an input of the local boundary correction network so as to output local polymerization coefficients of all positions of the confidence map, the local polymerization coefficients are multiplied with the corresponding positions of the confidence map, and the products are polymerized to a center point to obtain a local boundary correction result of scene segmentation; and the local boundary correction network is trained by using a known scene segmentation data set. Meanwhile, the method further proposes serially connecting a global residual correction network and the local boundary correction network to form a cascade frame, and the cascade frame can be used for global correction and local correction on thesegmentation result of the front-end model, thus obtaining a more accurate scene segmentation correction result.

Description

technical field [0001] The method belongs to the fields of machine learning and computer vision, and in particular relates to machine learning problems for scene segmentation in computer vision. Background technique [0002] Currently popular scene segmentation methods are mainly based on Convolutional Neural Networks (CNNs). Most of these methods utilize the framework of Fully Convolutional Networks (FCNs). Many methods make further improvements on the basis of FCNs, utilizing methods such as dilated convolutions, adding multiple deconvolution layers, and capturing features in the middle layers of the network. However, these methods mainly improve the segmentation accuracy based on the idea of ​​improving the network structure. [0003] Different from the above methods, some other methods aim to improve the current segmentation results. The more famous ones include the "fully connected conditional random field" method and the "multi-scale dilated convolution" method. Th...

AI Technical Summary

Problems solved by technology

This type of method mainly has the following problems: (1) Inconsistency and discontinuity often appear in the segmentation results, (2) The segmentation boundary of the target is often inaccurate and incoherent

Images