Adversarial-network-based semi-supervised semantic segmentation method

Abstract

The invention proposes an adversarial-network-based semi-supervised semantic segmentation method comprising two parts of network construction and a training process. To be specific, a ResNet-101-model-contained DeepLab-v2 frame pretrained in an ImageNet database is used as a segmentation network; a last classification network layer is removed and the space between last two convolutional layers ischanged from 2 to 1; an extended convolutional network is employed to increase a receiving domain and spatial Pyramid pooling is used after the last layer; and a full convolutional network is used asan authentication network and an up sampling network is used for adjusting an output image matching the size of an input image again. According to the invention, on the basis of the adversarial network, a semi-supervised semantic segmentation method is provided; the full convolution discriminator allows the system to carry out semi-supervised learning and an additional supervision signal is provided, so that the performance of image semantic segmentation is improved.

Description

technical field [0001] The present invention relates to the field of semantic segmentation, in particular to a semi-supervised semantic segmentation method based on confrontation network. Background technique [0002] Semantic segmentation technology, in simple terms, is given a picture and classifies each pixel in the picture. Image semantic segmentation is an important branch in the field of artificial intelligence and an important part of image understanding in machine vision technology. Semantic segmentation technology plays a very important role in practical applications. For example, in an automatic driving system, semantic segmentation technology can recognize and understand street view images well, output a more realistic scene map, and enable the automatic driving system to make safer and more reliable driving operations; In human-machine applications, semantic segmentation technology is conducive to more accurate positioning of the landing point of drones; in wea...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the problems such as poor image semantic segmentation effect in existing methods, the purpose of the present invention is to provide a semi-supervised semantic segmentation method based on confrontation network. Firstly, DeepLab-v2 with ResNet-101 model and pre-trained in ImageNet database frame as a segmentation network; then remove the last classification network layer, and change the spacing between the last two convolutional layers from 2 to 1; then use an extended convolutional network to increase the receptive field, using space after the last layer Pyramid pooling; the final discriminator network uses a fully convolutional network with an upsampling network layer to rescale the output image to match the size of the input image

Images