Image segmentation method combining thermodynamic diagram channel and improved U-Net

Abstract

The invention discloses an image segmentation method combining thermodynamic diagram channel and improved U-Net. The method comprises the following steps: 1) obtaining an original image; 2) preprocessing the original image to obtain a thermodynamic diagram; 3) constructing a multi-scale convolution module to improve a U-Net network structure; 4) entering a thermal map prepared in step 2) into theimproved U-Net network structure for image segmentation to obtain the segmented image. The multi-scale convolution module and a thermodynamic diagram channel are added to the U-Net network structure.Compared with a traditional U-Net model, the multi-scale convolution module and the thermodynamic diagram channel are added, so that not only can complete feature information be obtained, but also theprecision of a target edge can be effectively improved, and the edge is smoother.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image segmentation method combined with a thermal map channel and improved U-Net. Background technique [0002] The images required for multi-view 3D reconstruction often have the characteristics of rich colors and fine textures. Manual segmentation is extremely time-consuming, resulting in a small amount of manually labeled data. Thanks to its excellent structural design, U-Net has the characteristics of small data scale required for training and well preserved feature location information, so the present invention chooses U-Net as the basic structure. [0003] U-Net consists of a contraction path (i.e. encoder) and an expansion path (i.e. decoder). The shrinking path consists of a typical convolutional neural network, using two 3x3 convolutions in each step of the shrinking path, and using a linear activation unit (ReLU) after each convolution operation, and using ...

AI Technical Summary

Problems solved by technology

When the target is large, the edge is often discontinuous and the target area is incomplete

On the contrary, if the convolution kernel is set too large, the field of view will be too large, so that the extracted features cannot obtain effective local information, resulting in inaccurate segmentation results.

Images