Image understanding method based on cognition

Abstract

The present invention relates to an image understanding method based on cognition. The method comprises the steps of: an establishment step of a high-cognition-degree neural network training model: associating an image data set to be classified with an image label data set, performing reclassification of the internal portion of the image data set to be classified according to an association result, performing training of a deep convolutional neural network of the data set after reclassification, and obtaining a high-cognition-degree neural network training model; and an image understanding step: allowing images to be understood to be processed by the high-cognition-degree neural network training model to obtain labels corresponding to the images. Compared to the prior art, the image understanding method based on cognition is high in understanding accuracy, more accords with the actual condition, effectively shortens the image understanding time, etc.

Description

technical field [0001] The invention relates to the field of image semantic understanding, in particular to a cognition-based image understanding method. Background technique [0002] With the development of computer vision and natural language, more and more research and applications combine the two, use algorithmic models to understand the semantic content of pictures, and express the content in natural language. In the process of image understanding, deep convolutional neural networks are usually used to extract the visual features of pictures. A common method is to continue training and parameter fine-tuning on the trained classification task pre-training model. However, the category division method as the training target plays a vital role in the feature learning of the neural network. Different from traditional classification tasks, it is not necessary to accurately classify all objects in the picture in the content understanding of pictures, but it is necessary to l...

AI Technical Summary

Problems solved by technology

[0004] First, the types of objects in the data set are randomly selected, which is not suitable for feature learning of image understanding tasks

[0005] Second, the classification in the data set is too fine, and some category labels are beyond the cognitive ability of ordinary humans

[0006] Third, the training samples of some classes in the dataset are not representative

[0007] Fourth, although for pure visual tasks, the deep convolutional neural network pre-training model can improve the training performance to a certain extent, but it does not consider the impact of human cognitive ability on visual understanding

Images