Model training and instance segmentation method, device and system and storage medium

Abstract

The embodiment of the invention provides a model training method, device and system and a storage medium. The network training method comprises the steps that a training image and a training image mask are acquired; the training image is input into a front-end model structure in an instance segmentation model to obtain a target region alignment feature map, wherein the target region alignment feature map is a feature map which corresponds to a first target region used for indicating the position of a target object in the training image and is subjected to alignment operation; the target regionalignment feature map is input into a mask generator in the instance segmentation model to obtain a generated segmentation mask corresponding to the target region alignment feature map; a real segmentation mask corresponding to the target region alignment feature map is determined based on the training image mask; and the target region alignment feature map, the generated segmentation mask and the real segmentation mask are utilized to perform adversarial training on the mask generator and a discriminator. Through the method, the problem that the mask generator is influenced by image multimodality can be effectively solved.

Description

technical field [0001] The present invention relates to the field of image recognition, and more specifically relates to a model training method, device and system, an instance segmentation method, device and system, and a storage medium. Background technique [0002] Instance segmentation has always been a very important and highly concerned issue and technical method in the field of computer vision. Its goal is to give an image and give the segmentation mask (Segmentation Mask) and category information of all target objects in the image through an algorithm. [0003] The current mainstream framework for instance segmentation is the Mask Region Convolutional Neural Network (Mask-RCNN) framework. In Mask-RCNN, a number of target regions (RoI for short, also known as regions of interest) are given to each image through the region proposal network (RPN), and then the RoI is aligned to obtain the RoI alignment feature map, and finally The RoI-aligned feature map can be fed int...

AI Technical Summary

Problems solved by technology

The current mask generator training method cannot handle the multi-peak problem of the image well, that is, when

Images