Deeplabv3+ semantic segmentation model based on feature cross attention mechanism

Abstract

Aiming at the problem that the deeplabv3+ model is not accurate in segmentation of the image target edge, the image feature fitting is slow, and the attention information cannot be effectively used. It is proposed to add a feature cross attention module (FCA) to the model. The cross-attention network is composed of two branches and a feature cross attention module. Among them, the shallow branch is used to extract low-level spatial information, and the deep branch is used to extract high-level context features to make important feature extraction more refined. This paper designs and realizes the connection between Feature Cross Attention module and Deeplabv3+ coding module, input the output features of the Deeplabv3+ encoding module into the feature cross attention module for convolution operation to realize the recalibration of the original features. The decoding module of Deeplabv3+ obtains spatial features and channel features from two branches respectively, and then merges the obtained features to obtain more important features. The improved model was validated by the Pascal Voc2012 data set, and the results showed that the ratio of average intersection and the average pixel accuracy were increased by 1.96% and 2.84%, respectively. The model added with FCA can effectively improve the shortcomings of the original model, can segment the target more finely, and better solve the problem of rough segmentation boundary.