Adversarial-based lightweight network semantic segmentation method

Abstract

The invention relates to an adversarial-based lightweight network semantic segmentation method, which is used for solving the problems of low prediction accuracy, low network processing speed and difficulty in meeting the requirement of real-time prediction. The invention provides a lightweight semantic segmentation method based on adversarial from the perspective of improving semantic segmentation speed and precision. The method comprises the following steps: firstly, improving the network information acquisition capability by reducing the number of channels, reducing the parameter quantity in jump connection by utilizing asymmetric convolution, increasing the receptive field of a feature map by utilizing cavity convolution and disturbing the operation of the channels, and constructing alightweight asymmetric encoding and decoding semantic segmentation network; using confrontation ideas, and judging the segmented image and the calibrated semantic label by using a judgment network, designing a judgment loss function and a segmentation loss function, and alternately updating the segmentation network and the judgment network by using a back propagation method until the judgment network cannot distinguish the label and the real label generated by the segmentation network, thereby realizing semantic segmentation of the image. According to the method, the lightweight model and theadversarial idea are utilized, so that the segmentation precision is relatively high while the real-time performance of the segmentation network is ensured.

Description

technical field [0001] The invention belongs to the field of semantic segmentation of computer vision images, and in particular relates to a confrontation-based lightweight network semantic segmentation method. Background technique [0002] With the development of science and technology, more and more new technologies, such as service robot food delivery and unmanned driving, are applied to people's production and life. In this context, researchers have conducted a lot of research in related fields. Robotic food delivery and unmanned driving both need to establish a perception of the surrounding environment first, and image semantic segmentation can help machines deeply understand the scene they are in. Image segmentation is a basic task in computer vision, and it is also the basis for realizing machine perception of the environment and even interaction with humans. The effectiveness of segmentation directly affects the level of machine intelligence. Image semantic segment...

AI Technical Summary

Problems solved by technology

Although ENet has made great progress in semantic segmentation speed, it can process 70 images per second, but its accuracy rate has also dropped significantly, with an accuracy rate of only 58%.

In 2017, Zhan et al. [4] An image cascading network (ICNet) based on a spatial pyramid pooling network is proposed. Under the guidance of appropriate correct labels, different resolution feature map branches in the network are fused to achieve a real-time semantic segmentation accuracy of 67%, but each Only 30 images can be processed per second

In 2019, Wang et al. [5] A lightweight network (LEDNet) is proposed, which disrupts the feature map channels of the residual network, which can reduce the amount of calculation and collect more useful information. The calculation speed and accuracy have been improved, but the accuracy is also low. Only 70%, there is still a lot of room for improvement in the accuracy of real-time semantic segmentation

[0005] In summary, although the existing real-time semantic segmentation methods have improved the speed of semantic segmentation, they have more or less sacrificed accuracy.

Images