Blood cell subtype image classification method based on multi-scale fusion

Abstract

A blood cell subtype image classification method based on multi-scale fusion is provided. The method comprises the following steps: (1) wherein a training set contains four erythrocyte subtype images;(2) based on an Xception Entry flow module, constructing a shallow feature extraction network, and outputting four feature maps of different scales; (3) respectively connecting four intermediate feature extraction networks formed by cascading Xception Middle flow modules for the four outputs in step (2); (4) respectively connecting four modified Xception Exit flow modules for the four outputs instep (3), extracting deep feature information, and outputting four high-dimensional feature vectors; (5) fusing the information of 4 high-dimensional feature vectors output in (4), and performing inference and prediction on image categories; (6) on the classification network, using the data-amplified blood cell subtype image training set for training; and (7) using the network trained in step (6)to predict on a blood cell subtype image test set, and outputting categories to which the blood cell subtype images belong. According to the method disclosed by the present invention, the performanceof a blood cell subtype image classifier can be enhanced.

Description

technical field [0001] The invention belongs to the fields of image processing, computer vision, deep learning, and image classification, and specifically relates to a blood cell subtype image classification method based on multi-scale fusion. Background technique [0002] At present, the most powerful image processing technology in the industry is deep learning. In the field of image classification, the accuracy of ImageNet competition as a benchmark for computer vision classification algorithms is already deeply rooted in the hearts of the people. Since 2012, Convolutional Neural Networks and Deep Learning have dominated the competition's leaderboards. In 2015, Kaiming He et al. proposed the ResidualNeural Network structure. By using the residual module and replacing the fully connected layer with Global Average Pooling, they successfully trained a 152-layer deep network and won the ImageNet championship of the year, with top-5 accuracy. Reached 93.3%. In 2017, Francois...

AI Technical Summary

Problems solved by technology

This type of network often improves the receptive field of the convolution kernel and reduces the amount of computation through multiple pooling operations, but this operation also often causes a sharp drop in image resolution, causing some details of the image to be lost. This may be unacceptable for medical imaging tasks of fine tissue

Images