Division and identification method and device based on dense network image

Abstract

The invention provides a division and identification method and device based on a dense network image and relates to the technical field of neural networks. The division and identification method based on the dense network image comprises the following steps: firstly, capturing layering features of the dense network image by taking a convolutional layer as a feature extracting device; inputting obtained multi-scale feature patterns into an encoding module; meanwhile, connecting feature patterns with different scales with the feature patterns with the corresponding scales in the encoding moduleby utilizing jumping connection; carrying out down-sampling and semantic feature extraction on the feature patterns by applying the encoding module; carrying out up-sampling on the feature patterns by applying a decoding module and recovering detailed information to obtain the multi-scale feature patterns; then carrying out chain type residual error pooling operation on an output feature patternobtained by dense deconvolution operation, so as to improve the accuracy of boundary division and capture local contextual information; adding depth monitoring into three boundary thinning blocks in the encoding module and carrying out steps above to improve the precision of medical image division.

Description

technical field [0001] The invention relates to the technical field of neural networks, in particular to a segmentation and recognition method and device based on dense network images. Background technique [0002] Deep learning is a machine learning method based on representational learning of data. Its motivation is to establish and simulate the neural network of the human brain for analysis and learning, that is, to imitate the mechanism of the human brain to explain data, such as images, sounds and texts. In practice, deep learning represents attribute categories or features by combining low-level features to form more abstract high-level features to discover distributed features of data. [0003] In recent years, with the amazing performance of deep learning in the field of object recognition, more and more researchers have applied deep learning to the field of medical image segmentation. The traditional multi-layer nonlinear convolutional neural network has good gener...

AI Technical Summary

Problems solved by technology

With the development of technology, the classic network structure used for object segmentation is fully convolutional neural network (Fully Convolutional Networks, FCN). Since then, many networks have been improved on the basis of it. The fully convolutional neural network is divided into decoding and FCN. There are two parts of encoding, decoding is used to extract convolutional features, and encoding is used to restore the size of the original image, thus largely avoiding the resolution limitation of conventional convolutional neural networks, but its output segmentation result map is only It can be restored to 1 / 32 of the original image, but the resulting resolution is still low, and a high-resolution segmentation result map cannot be obtained

In order to solve this problem, a network called DeepLab based on dilated convolution for image segmentation has been developed. Unlike the previous two structures, it has no pooling operation, which makes the original image size not decrease, and the resolution It will not be reduced, but due to the need for convolution operations on high-resolution and high-dimensional feature maps, a large amount of computing memory is required, which requires high GPU performance

[0004] In summary, there are many limitations in the use of deep learning in the medical image segmentation process, and there is no effective solution

Images