DeepLabv3plus-IRCNet image semantic segmentation algorithm based on coding and decoding structure

Abstract

The invention discloses a DeepLabv3plus-IRCNet image semantic segmentation algorithm based on a coding and decoding structure, and relates to the technical field of image semantic segmentation. The method is based on DeepLabv3plus, and comprises the following steps: extracting picture features by a deep convolutional neural network formed by introducing depth separable convolution inverted residual modules in series, and introducing a feature map segmentation module under 1 / 16 resolution; amplifying each segmentation feature map to the size before segmentation; extracting features in a parameter sharing mode; splicing the corresponding positions of each output feature map and then performing fusion with the feature map amplified to the same size in the decoding stage, so that the extraction capability of the model for features of a small target object is improved, and the problem of small target loss caused by gradual reduction of the resolution of the feature map in the down-samplingprocess of the image is effectively solved.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a DeepLabv3plus-IRCNet image semantic segmentation algorithm based on a codec structure. Background technique [0002] Image semantic segmentation is one of the fundamental research topics in the field of computer vision, and its goal is to assign a semantic label to each pixel in an image. The current development of the semantic segmentation model based on deep convolutional neural networks (DNCCs) is mainly due to the proposal of the full convolutional neural network (FCN), which replaces all the fully connected layers behind the convolutional neural network with convolutional layers. , upsampled using a bilinear interpolation algorithm, and outputs a split map with the input size. However, there were two problems with FCN at that time: 1) The resolution of the feature map was constantly shrinking, resulting in the loss of some pixels of small objects; 2) The image context infor...

AI Technical Summary

Problems solved by technology

However, there were two problems with FCN at that time: 1) The resolution of the feature map was constantly shrinking, resulting in the loss of some pixels of small objects; 2) The image context information (the relationship between pixels and pixels) was not fully considered, and the rich information could not be fully utilized. Spatial location information

[0003] These methods are all for the purpose of enabling the model to better extract feature information at different resolutions of the middle layer. They are all directly operated on the entire feature map, but in a certain local area of ​​​​an image there are often For extremely small targets, ordinary convolution operations cannot extract the features of small target objects well, and it is difficult to obtain better semantic segmentation results

Images