Capsule network image positioning improved algorithm based on CNN

Abstract

The invention discloses an improved convolutional neural network-based pneumonia image focus positioning algorithm, which is characterized in that on the basis of a CNN simplified model, a batch normalization layer, a down-sampling layer and a Dropout layer are added to prevent over-fitting feature extraction, a ReLu activation function is added behind the Dropout layer, dimension reduction is further performed on a feature map, the expression ability of the CNN model is improved, and the accuracy of positioning the focus of a pneumonia image is improved. And transmitting the features extracted by the CNN model into a capsule network, updating the positioning information of the clustering center through a membership matrix, and outputting the positioning coordinates of the focus. The features can be automatically extracted through the neural network, and the focus area of the lung cancer image is positioned. According to the invention, the improved CNN simplified model is combined with the capsule network, the overfitting phenomenon is effectively prevented, and the image positioning accuracy is improved under the condition that a small number of training sets are used.

Description

technical field [0001] The invention belongs to the field of deep learning and image processing, and relates to the application of image positioning under the improved deep neural network technology. Background technique [0002] At present, the development of computer hardware performance has enabled deep learning technology to fit medical image positioning well. Traditional image processing techniques are mostly used in support vector machines. Such techniques require artificial feature extraction, which requires a lot of professional background of doctors. Deep learning technology does not require artificial feature extraction, and is an end-to-end learning method. Among them, xception is another improvement to Inception V3 proposed by Google after Inception. Its advantage is that the accuracy rate is improved, but the number of training sets required is also large. Through subsequent experiments, it was found that in order to obtain better training results, either incr...

AI Technical Summary

Problems solved by technology

But in this way, it will cause another problem, that is, an overly deep network can easily lead to overfitting

[0003] In order to solve the problem of convolutional neural network overfitting and dependence on the number of samples, the present invention proposes an algorithm based on the improved CNN simplified model, so that the improved model can complete the image of pneumonia lesions without a large number of training sets. positioning work

Images