Neural network migration learning method based on virtual image dataset

Abstract

The invention belongs to the field of machine learning technology and discloses a neural network migration learning method based on a virtual image dataset. The method comprises the steps that virtual image data is acquired through graphic simulation software, classified marking is performed on a virtual dataset, and an image dataset is obtained; a neural network is utilized to train a target recognition model through preprocessing of the virtual dataset, and a parameter model trained through the virtual dataset is saved; and migration learning is performed, and a new target dataset is utilized to train the saved pre-training model again. According to the method, a needed scene or target is collected through a virtual engine to construct a dataset based on the virtual image dataset, the pre-training model is obtained through training of the virtual dataset, and the migration learning method is utilized to achieve the practical application effect on a small quantity of datasets.

Description

technical field [0001] The invention belongs to the technical field of machine deep learning, and in particular relates to a neural network migration learning method based on a virtual image data set. Background technique [0002] With the development of machine learning, recognition technology in various complex scenes has become more and more important, but it is difficult to collect target data in some extreme environments and data sets from special perspectives (such as the drone's overlooking angle). Some public datasets cannot meet the needs of pattern recognition training datasets in specific scenarios. However, training a complex neural network requires a large amount of labeled data, which not only consumes a lot of manpower and material resources, but even with a large amount of training data, it takes days or even weeks to train a complex convolutional neural network. The problem of difficulty in collecting data sets in the environment and special perspectives, a...

AI Technical Summary

Problems solved by technology

[0003] To sum up, the problems existing in the existing technology are: the existing transfer learning technology cannot solve the difficulty of collecting and training training data sets in extreme environments and special perspectives, resulting in small training databases, cumbersome manual labeling data sets, and many The training time under the model is too long

Images