Multi-scale serial convolutional deep learning microscopic image segmentation method

Abstract

The invention discloses a multi-scale serial convolution deep learning microscopic image segmentation method, and the method comprises the following steps of S1, obtaining an environmental microorganism microscopic image, and carrying out gray processing on the image; S2, inputting the image after gray processing into a pre-trained MIaMIA-Net model, and outputting an image segmentation result, wherein the MIaMIA-Net model is a neural network model which is based on an Inception structure and a U-Net structure and is used for obtaining a multi-scale convolution kernel combined action effect through the serial convolution operation. According to the method, the parameter quantity and memory occupation of the model are reduced while the adaptability to a multi-scale target is improved, and the segmentation precision of an environmental microorganism microscopic image is high.

Description

technical field [0001] The invention relates to the technical field of microscopic image processing, in particular to a multi-scale serial convolution deep learning microscopic image segmentation method. Background technique [0002] In the prior art, the U-Net convolutional neural network deep learning algorithm is often used to perform image segmentation processing on microscopic images of environmental microorganisms. The U-Net structure includes the U-Net downsampling structure of the contraction path and the U-Net upsampling structure of the expansion path. [0003] Currently, the U-Net downsampling structure includes alternate downsampling convolution operations and 2×2 maximum pooling operations. The U-Net downsampling structure uses environmental microscopic images as input, and the last downsampling convolution operation The obtained feature map is used as the output; the downsampling convolution operation is two 3×3 convolutions, the length of the feature map afte...

AI Technical Summary

Problems solved by technology

[0005] Due to the difference in the shape of different types of microorganisms and the different shooting conditions of microbial micrographs, there is a problem of different segmentation target sizes in microbial images; each layer of U-Net has two 3×3 convolution kernels for convolution operations , therefore, U-Net has poor adaptability to targets of different scales

When the background of the microscopic image of environmental microorganisms is complex, the U-Net model has limited ability to eliminate noise, and the results obtained by segmentation perform poorly

The U-Net model is very easy to fall into local extreme points during the training process, resulting in loss that cannot be reduced, training fails, and the training model cannot be used

And the U-Net model has a large amount of parameters and takes up a lot of memory

Images