Training method for context insensitivity of semantic segmentation model

Abstract

The invention discloses a training method for context insensitivity of a semantic segmentation model. The training method is used for greatly accelerating a semantic segmentation algorithm of a video.The method specifically comprises the following steps: 1) obtaining a plurality of groups of image data sets for training semantic segmentation, and defining an algorithm target; 2) learning on the data set by using a model based on a full convolutional network structure; 3) generating a new training sample by using a class erasure sample generator; and 4) optimizing the original data set and thenew sample generated in the step 3) in combination with consistency constraints by using the network parameters obtained in the step 2) to obtain a model insensitive to the context. According to themethod, the scene understanding ability of semantic segmentation is mined, and the trained model has better generalization ability under the conditions of data erasing, data interference, style migration and the like.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to a context-insensitive training method for a semantic segmentation model. Background technique [0002] Semantic segmentation is a computer vision task that assigns each pixel in an image to a semantic label. The current industry-leading semantic segmentation technologies are all based on variants of the fully convolutional neural network (FCN), and most of these technical methods use contextual information to obtain better segmentation results. For example, PSPNet adds global pyramid pooling technology on the basis of FCN to increase context information. The DeepLab series of algorithms tried a variety of different atrous convolution architectures to obtain multi-scale context information. As a result, current semantic segmentation techniques are sensitive to contextual information. [0003] However, this context-sensitive strategy can lead to poor generalization abi...

AI Technical Summary

Problems solved by technology

[0003] However, this context-sensitive strategy will lead to weak model generalization ability, and cannot really make the model understand the scene with human-like ability

The context-sensitive model actually learns a joint probability distribution of semantic tags on the training data set. Once it encounters a scene that the model is not familiar with, the performance of the model will drop significantly.

Images