Training method of image semantic segmentation model and server

Abstract

The embodiment of the invention discloses a training method of an image semantic segmentation model and a server, which are used for positioning all object regions from an original image and improvingthe segmentation quality of image semantic segmentation. The embodiment of the invention provides a training method of an image semantic segmentation model. The training method comprises the following steps: acquiring an original image for model training; performing full-image classification annotation on the original image by using a multi-magnification hole convolutional neural network model toobtain a global object positioning image in the original image under different dispersion degrees, the dispersion degrees being used for indicating distribution of an object region positioned by themulti-magnification hole convolutional neural network model on a target object; and using the global object positioning graph as supervision information of an image semantic segmentation network model, and training the image semantic segmentation network model through the supervision information.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a training method and server for an image semantic segmentation model. Background technique [0002] Image semantic segmentation is the basis of image understanding, and image semantic segmentation is very important in autonomous driving, drone applications, and wearable device applications. An image is composed of many pixels, and semantic segmentation is to group pixels according to the different semantic meanings expressed in the image. [0003] In the prior art, a conventional deep convolutional neural network is usually trained as an image semantic segmentation network. First, the input image is classified into the whole image, and then the object area corresponding to the entire image classification label in the image is located according to the network, and then these areas are used as Supervised information for image semantic segmentation, through which the original im...

AI Technical Summary

Problems solved by technology

[0004] In the prior art, the conventional convolution operation is used to locate the object area corresponding to the classification and labeling of the whole image. Usually, only one or more of the most discriminative parts of the entire object can be located, and it is difficult to locate all object areas. Therefore, the existing There is a problem in the image semantic segmentation in the technology that it cannot locate all object areas

Images