Super-large image classification method based on graph neural network

Abstract

The invention belongs to the field of image classification, and relates to a super-large image classification method based on a graph neural network. Different sub-image screening methods can be adopted for different super-large images, and the feature extraction network is further adjusted, so that feature extraction of the sub-images is more accurate; a super-large image is constructed into image data; a differentiable pooling operation is introduced into a traditional graph convolutional neural network; and global information of the super-large image can be mined, feature information of a hidden layer can be mined in the training process through micro-pooling operation, the relevance of all the sub-images in the feature space is fully analyzed, and the super-large image can be classified more accurately.

Description

technical field [0001] The invention belongs to the field of image classification, and relates to a super-large image classification method based on a graph neural network. Background technique [0002] With the rapid development of science and technology and digital imaging technology, high-resolution imaging equipment is more and more widely used, the types of image data that can be obtained are constantly enriched, and the amount of data contained in a single image has also increased from several megabytes in the past to the present. Hundreds of megabytes, several gigabytes or even dozens of gigabytes, this kind of images that contain a huge amount of data and have ultra-high resolution cannot be read into the computer memory for direct processing at one time. The most representative super-large images include electronic scanning pictures of microscope imaging, pathological slices in medicine, and satellite remote sensing images, etc. To analyze and process these super-l...

AI Technical Summary

Problems solved by technology

However, convolutional neural network technology cannot be directly applied to super-large image classification tasks for the following reasons: (1) The size of the pictures is different: the size of natural pictures is usually on the MB level, which can be directly read into the computer memory and sent to the neural network. The network performs feature extraction and classification, but the average size of a super-large image is on the order of GB. Direct reading and display will cause computer memory overflow, which cannot be directly processed by the neural network. Usually, the entire image needs to be cropped to obtain multiple sub-images. image, feature extraction and analysis on sub-images; (2) The situation of image labels is different: it is relatively easy to obtain natural images and their accurate labels. However, in the classification task of very large images, due to the large Labeling at the sub-image level takes a lot of time and effort, and is difficult to implement. Therefore, the label of the sub-image is not necessarily available. Even if it is available, it is likely to be different from the label of the original image, so the label data is more complicated; (3) The nature of the picture itself is different: for natural pictures, the whole image is directly processed to complete the classification task, but for very large images, sometimes the task-related area, that is, the ROI (Region of Interest) only accounts for the whole picture. A small part, most of them are background areas or areas that have nothing to do with the task, and the classification task is more complicated; (4) The quality of super large images is uneven: due to the different specifications of scanning machines, there are large differences between different images of super large images, so it is necessary to Perform more complex data preprocessing

Images