Real-time semantic segmentation method based on context attention mechanism and information fusion

Abstract

The invention discloses a real-time semantic segmentation method based on a context attention mechanism and information fusion, and the method comprises the steps that: firstly a real-time semantic segmentation network is constructed, the preprocessing of an image in a data set obtained in advance and a corresponding label is carried out, and the preprocessed image and the corresponding label areinput into the constructed network for training; in the network, the input image passes through three groups of independent downsampling modules, so that the resolution of the input image is distinguished to be 1 / 2, 1 / 4 and 1 / 8 of the original resolution, and three feature maps with different sizes are respectively subjected to feature fusion with subsequent features of different stages of a backbone network; after the features of each stage are fused, an attention module is accessed; and the fused features are subjected to classified convolution operation to output a final prediction result,the final prediction result is compared with a corresponding semantic annotation image, and a cross entropy loss function is calculated as a target function, so that a trained network model is obtained. According to the method, the high precision of semantic segmentation is guaranteed, and the efficient reasoning speed and the memory capacity suitable for boundary equipment are also guaranteed.

Description

technical field [0001] The invention belongs to the fields of computer vision and pattern recognition, and in particular relates to a real-time semantic segmentation method based on a context attention mechanism and information fusion. Background technique [0002] Semantic segmentation is one of the key problems in the field of computer vision today. It is classified at the image pixel level, and pixels belonging to the same class are classified into one class. Therefore, contextual semantic information is very important for semantic segmentation. In practice, virtual reality, human-computer interaction, and autonomous driving will all use semantic segmentation technology, and an accurate understanding of the surrounding scene has an important impact on the decision-making of practical applications. [0003] The current best image semantic segmentation methods are all implemented based on deep convolutional neural network methods, and they are all based on encoding and de...

AI Technical Summary

Problems solved by technology

However, this also leads to another important problem. Many existing lightweight networks will pursue faster speeds based on blindly sacrificing segmentation accuracy.

Images