An Instance Segmentation Method Fused with Atrous Convolution and Edge Information

Abstract

The invention discloses an instance segmentation method for fusing atrous convolution and edge information, comprising: establishing an instance segmentation model including a feature pyramid extraction network, an area of ​​interest extraction network, a preliminary segmentation network, and an edge detection network; The network includes a first feature extraction network, a second feature extraction network, and a mixed dilated convolutional layer; the mixed dilated convolutional layer is used to mix the top-level feature maps of the feature pyramids output by the first feature extraction network and the second feature extraction network Hole convolution; preliminary segmentation network, used to classify, position regression and segment the region of interest output by the region of interest extraction network; edge detection network, used to perform edge detection on the segmentation result to obtain the final image segmentation result; The trained instance segmentation model performs instance segmentation; the method of the invention can avoid the loss of feature information, improve the image edge fitting effect, and improve the segmentation accuracy.

Description

Technical field [0001] The present invention belongs to the field of image processing and machine vision, and more specifically, the present invention relates to an example segmentation method of fusion cavity and edge information. Background technique [0002] The instance division is one of the image segmentation, which divides the scene image into a plurality of regions, each area corresponding to an object, and indicates a category tag of the area. Compared to other segmentation methods, instance segmentation is closer to our human beings to the world, and allows for subsequent processing of the scene constituent elements, such as moving the pedestrians. However, for human perception, the image is divided into multiple subsets, more dependent on subjective judgment, and there is no objective standard or specified guidance instance segmentation, so the result of instance segmentation has no standard answer; in addition, the image is included The information is complex, and the...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the defects of the prior art, the purpose of the present invention is to provide an instance segmentation method that integrates atrous convolution and edge information, aiming to solve the problem that the existing instance segmentation method has feature information loss, and the segmentation result has edge errors, which leads to the segmentation accuracy low problem

Images