Real-time segmentation system and method based on hybrid expansion network

Abstract

The invention relates to a real-time segmentation system and method based on a hybrid expansion network, and belongs to the field of computer vision. The system comprises a backbone network MobileNetv2 and a lightweight hybrid hole convolution module. The lightweight hybrid hole convolution module realizes comprehensive balance in the aspects of accuracy and high efficiency through multi-scale information and an effective attention mechanism; the lightweight hybrid hole convolution module mainly comprises a depth separable attention module and a hybrid multi-scale module; the depth separableattention module is of a single-layer hybrid convolution design. In one aspect, the system may enhance the expression of information by increasing the depth of the network. On the other hand, the depth separable convolution performs convolution integral deconvolution on each channel, thereby reducing the parameter size and the calculation cost.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a real-time segmentation system and method based on a hybrid expansion network. Background technique [0002] At present, various sensors apply image segmentation technology, which has the ability to perceive the environment, and can segment and recognize various information in the picture. The sensor determines the follow-up work through the perception of the surrounding scene. The image segmentation model applied by it is usually divided into two types: , one is an end-to-end semantic segmentation model, and the other is a real-time segmentation model. [0003] In order to improve the accuracy of segmented images, previous end-to-end methods often design more complex semantic segmentation structures, such as multi-scale, dense connection strategies, etc. Chen et al. proposed the DeepLab series, mainly proposing a pyramid pooling module (ASPP) that uses expanded convolutions with...

AI Technical Summary

Problems solved by technology

However, these real-time models have problems such as insufficient information extraction ability and loss of detailed information, resulting in a decrease in accuracy performance, and it is difficult to meet the segmentation requirements for complex road conditions in practical applications.

Images