Image semantic segmentation model training method and image semantic segmentation method

Abstract

The invention discloses a training method for an image semantic segmentation model and an image semantic segmentation method. The training method includes: respectively inputting training images with pre-marked semantic segmentation information into at least two basic semantic segmentation sub-models to obtain at least two corresponding A feature map containing semantic information; at least two feature maps and their pre-labeled semantic segmentation information are input into the attention model at the same time to calculate the weight of each feature map; the fusion unit is used to combine at least two of the features The graph is fused according to the corresponding weight to obtain the predicted semantic segmentation result of the training image; according to the predicted semantic segmentation result of the training image and the pre-labeled semantic segmentation information, at least two basic semantic segmentation sub-models and the attention model Parameters are corrected; using several training images to iteratively execute the above training steps until the training results of at least two basic semantic segmentation sub-models and the attention model meet the preset convergence conditions.

Description

technical field [0001] The invention relates to the fields of computer vision and image processing, in particular to an image semantic segmentation method based on attention model fusion and a model training method. Background technique [0002] Image semantic segmentation is an important research content in the field of computer vision. Its goal is to segment the image into regions with different semantic information, and label the corresponding semantic labels of each region. For example, after image semantic segmentation of an image Semantic tags (such as tables, walls, sky, people, dogs, etc.) can be added to objects in the image, which can be applied to fields such as driverless driving. [0003] Currently, the most mainstream solutions for image semantic segmentation are mainly based on Convolutional Neural Networks (CNN for short), which learn the semantic feature representation of images. For example, a fully convolutional network (Fully Convolutional Networks, refe...

AI Technical Summary

Problems solved by technology

[0007] Aiming at the problem that the improvement of feature fusion effect is not obvious, the present invention proposes an image semantic segmentation method based on attention model fusion, by using multiple basic semantic segmentation sub-models to extract features with semantic information from the input image, and Reasonably calculate the weight of each feature fusion through the attention model, and then fuse multiple feature maps according to the corresponding weights, and finally use the fused features to predict the segmentation results, so as to improve the accuracy of image semantic segmentation prediction.

Images