A Method for Recognizing Feature Regions of Object Model Based on Instance Segmentation
Guohua Chen, Yiyou Zhang, Jian Chun Xing · 2021
The instance segmentation technology for the feature region of the object model has the problem of slow network model convergence and general recognition accuracy. It is particularly important to enhance the accuracy of the local feature recognition of the object model by improving the instance segmentation network. Carry out all-round data enhancement operations for the data set model, and propose a new instance segmentation network model based on the maskrcnn architecture improvement, which includes deepening the feature extraction architecture, using the nonlinear activation function mish, and the loss function regression in the target detection The use of ciou in, and the use of the training optimizer ranger, improve the accuracy of the recognition of the local features of the two-dimensional object model, and perform the instance segmentation of the feature region. The data set verification confirms the convergence performance and model accuracy of the new instance segmentation architecture.