Image multi-scale semantic segmentation method and device based on fusion network

Abstract

The invention relates to an image multi-scale semantic segmentation method and device based on a fusion network. The image multi-scale semantic segmentation method comprises the following steps: constructing a fusion network, wherein the fusion network comprises an fcn-based network, a deeplab-based network, a feature fusion module and an optimization segmentation module, and the feature fusion module is connected with the fcn-based network, the deeplab-based network and the optimization segmentation module; and performing semantic segmentation on an input image through the fusion network. Compared with the prior art, the image multi-scale semantic segmentation method has the advantages of high segmentation precision and the like.

Description

technical field [0001] The invention relates to the technical field of visual image processing, in particular to a fusion network-based image multi-scale semantic segmentation method and device. Background technique [0002] Semantic segmentation is the process of marking each pixel in an image to its category, and is the basis of visual analysis technologies such as autonomous driving, medical image processing, image retrieval, and object classification. [0003] Before the emergence of fully convolutional networks, although convolutional neural networks have achieved great success in the field of object recognition, due to the limitations of fully connected layers and pooling layers, the mainstream method in the field of semantic segmentation is still TextonForest. ) or traditional methods such as Random Forest. [0004] In 2014, Long et al. at UC Berkeley proposed that figure 1 The FCN (Fully Convolutional Network) model shown replaces the fully connected layer at the e...

AI Technical Summary

Problems solved by technology

[0007] Although DeepLab v2 achieves high segmentation accuracy, there is a "girdding issue" in the atrous convolution, that is, the atrous convolution inserts a 0 value between the two sampling pixels of the convolution kernel. If the expansion rate is too large, the convolution will Too sparse, poor ability to capture information; and not conducive to model learning - because some local information is lost, and some information on long distances may not be relevant

[0008] To sum up, the FCN network has a good ability to extract local features, but it directly enlarges the feature map to the size of the original image by upsampling when restoring the size of the feature map, which can only produce smooth feature maps, and lacks the ability to distinguish between different scales. Feature Extraction Ability

Although DeepLab extracts more global features through atrous convolution, it loses some local information.

Images