Image segmentation method and device, computer equipment and storage medium

Abstract

The invention relates to an image segmentation method and device, computer equipment and a storage medium. A neural network model in the image segmentation method does not contain a target detection network in a traditional image segmentation model. The method comprises the steps of establishing a corresponding relation between different objects in a pre-segmented image and channels of a neural network model output feature map; based on the corresponding relation, directly obtaining a feature map containing a preset number of channels after an image is input, wherein each channel of the feature map corresponds to different target objects in the input image; and on the basis, obtaining a segmentation result based on each channel segmentation image of the feature map. Therefore, according tothe method, the network structure of the instance segmentation model is simplified, and the instance segmentation efficiency is high.

Description

technical field [0001] The present application relates to the technical field of computer vision, in particular to an image segmentation method, device, computer equipment and storage medium. Background technique [0002] With the development of deep learning and the popularization of intelligent terminals, the application of image segmentation is becoming more and more extensive. For example, fast and accurate multi-person portrait segmentation is one of the technologies in demand at this stage. Multi-person portrait segmentation is a kind of instance segmentation. Most of the existing instance segmentation methods are oriented to multi-category segmentation problems, and their processes include multiple operations including detection, which is relatively complicated. [0003] Taking the most widely used Mask R-CNN algorithm in recent years as an example, its model includes convolutional network, Regionproposal network, ROI Align, and network layers such as detection class...

AI Technical Summary

Problems solved by technology

Its defects are as follows: the model network is complex and uses a lot of resources; it also uses a variety of custom operations in addition to convolution, which is difficult to implement and optimize in a variety of different platforms; sampling the features of the detection frame area to the same smaller The corresponding segmentation result is obtained after the size, and the accuracy of the segmentation edge is low

[0004] In addition, although a "bottom-up" segmentation process is proposed in some methods for portrait segmentation, such as Richeimer, Bounding Box Embedding for Single Shot Person InstanceSegmentation, arXiv preprint arXiv:1807.07674 (2018), however, such image segmentation methods It still includes operations such as human detection or key point detection, and its implementation has high model complexity, which is difficult to apply to actual needs

Images