Adaptive adversarial learning-based urban traffic scene semantic segmentation method and system

Abstract

The invention discloses an adaptive adversarial learning-based urban traffic scene semantic segmentation method and system. The method comprises steps of obtaining training data of the semantic segmentation model, and preprocessing the game synthesis data set GTA5 to obtain a new synthesis data set SG-GTA5 which is close to urban scene real data set Cityscapes in distribution; constructing a generative adversarial network model for semantic segmentation; based on the training data set, performing self-adaptive confrontation learning on the generative adversarial network model, using self-adaptive learning rates in confrontation learning of different feature layers, adjusting loss values of the feature layers through the learning rates, and then dynamically updating network parameters to obtain an optimized generative adversarial network model; and carrying out verification on the city scene real data set CityScapes. According to the method, the semantic segmentation precision of a complex urban traffic scene which lacks annotation information and has more scale targets can be improved, and the generalization ability of a semantic segmentation model is enhanced.

Description

technical field [0001] The invention relates to the field of semantic segmentation of images based on weak / semi-supervision, in particular to a method and system for semantic segmentation of urban traffic scenes based on adaptive confrontation learning. Background technique [0002] Semantic segmentation refers to dividing an image into several groups of pixel regions with characteristic semantics, identifying the category of each region, and finally obtaining an image with pixel semantic annotations. Through image semantic segmentation, complex images are easier to understand and analyze. Semantic segmentation is the basis of visual analysis and understanding for autonomous driving, visual navigation, image retrieval, object recognition, and detection and tracking. [0003] Because deep learning can learn high-level semantic features and has a strong ability to fit complex scenes, deep learning has made breakthroughs in the field of computer vision research. In the networ...

AI Technical Summary

Problems solved by technology

[0008] (1) The image semantic segmentation method based on full supervision requires a large number of images and corresponding labeled samples, and needs to be relabeled for different scenes or the same scene under different lighting and climate conditions, which is time-consuming and laborious; non-supervised Image semantic segmentation does not require additional labeling, but the segmentation accuracy is often very low;

[0009] (2) Urban traffic scenes are complex and easily affected by lighting and climate conditions, so the semantic segmentation of such scenes usually has class drift and class infection, which affects the accuracy of model segmentation and model mobility;

[0010] (3) The current popular semantic segmentation methods based on domain-adaptive adversarial generator networks usually directly conduct adversarial training on the source domain (synthetic dataset) and the target domain (real dataset), however, when training directly on the synthetic dataset , the model cannot be well generalized to the real data set, the main reason is that the color, texture and other feature distributions of the source domain and target domain images are quite different, the adversarial loss value in the training process is large, and the model in the backpropagation process Gradient explosion is prone to occur in the middle, thus affecting the segmentation accuracy

[0011] (4) The current popular semantic segmentation based on domain adaptive confrontation generator network usually adopts the method of fusion of multiple feature layers, but fixed penalty coefficients are added to the cross entropy loss value and confrontation loss value of different feature layers, and through multiple It is obtained by manually adjusting the second experiment, and the robustness to semantic segmentation of complex traffic scenes is poor

Images